[X86][Skylake] Fixing duplicated prefixes in the run command of Code Gen regression...

NFC.
Replaced duplicated HASWELL prefixes in run commands in the X86 Code Gen regression tests by the SKYLAKE prefix when the -mcpu is set to skylake.
The fix is needed in preparation of an upcoming patch containing the Skylake scheduling info.

Reviewers: zvi, RKSimon, aymanmus, igorb

Differential Revision: https://reviews.llvm.org/D37258

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@312103 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/test/CodeGen/X86/aes-schedule.ll b/test/CodeGen/X86/aes-schedule.ll
index ab86edac0b4cebb06e7f2c6af4a0d857dd7f94a3..56e89b62e2c41e7e16bcd43ceda3c4be96e8298f 100644 (file)
--- a/test/CodeGen/X86/aes-schedule.ll
+++ b/test/CodeGen/X86/aes-schedule.ll
@@ -5,7 +5,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -34,6 +34,12 @@ define <2 x i64> @test_aesdec(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vaesdec (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aesdec:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaesdec %xmm1, %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vaesdec (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aesdec:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaesdec %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -77,6 +83,12 @@ define <2 x i64> @test_aesdeclast(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2)
 ; HASWELL-NEXT:    vaesdeclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aesdeclast:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaesdeclast %xmm1, %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vaesdeclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aesdeclast:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaesdeclast %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -120,6 +132,12 @@ define <2 x i64> @test_aesenc(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vaesenc (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aesenc:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaesenc %xmm1, %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vaesenc (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aesenc:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaesenc %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -163,6 +181,12 @@ define <2 x i64> @test_aesenclast(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2)
 ; HASWELL-NEXT:    vaesenclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aesenclast:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaesenclast %xmm1, %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vaesenclast (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aesenclast:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaesenclast %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -210,6 +234,13 @@ define <2 x i64> @test_aesimc(<2 x i64> %a0, <2 x i64> *%a1) {
 ; HASWELL-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aesimc:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaesimc %xmm0, %xmm0 # sched: [14:2.00]
+; SKYLAKE-NEXT:    vaesimc (%rdi), %xmm1 # sched: [14:2.00]
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aesimc:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaesimc (%rdi), %xmm1 # sched: [7:1.00]
@@ -260,6 +291,13 @@ define <2 x i64> @test_aeskeygenassist(<2 x i64> %a0, <2 x i64> *%a1) {
 ; HASWELL-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_aeskeygenassist:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaeskeygenassist $7, %xmm0, %xmm0 # sched: [29:7.00]
+; SKYLAKE-NEXT:    vaeskeygenassist $7, (%rdi), %xmm1 # sched: [28:7.00]
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_aeskeygenassist:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaeskeygenassist $7, (%rdi), %xmm1 # sched: [7:1.00]

diff --git a/test/CodeGen/X86/avx-schedule.ll b/test/CodeGen/X86/avx-schedule.ll
index 17d2d9146c001564c28dd55841f3754561da454f..3f0035dd9fb86a26534aaf0ffac653cb2ab76191 100644 (file)
--- a/test/CodeGen/X86/avx-schedule.ll
+++ b/test/CodeGen/X86/avx-schedule.ll
@@ -3,7 +3,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -26,6 +26,12 @@ define <4 x double> @test_addpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vaddpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -62,6 +68,12 @@ define <8 x float> @test_addps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vaddps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -98,6 +110,12 @@ define <4 x double> @test_addsubpd(<4 x double> %a0, <4 x double> %a1, <4 x doub
 ; HASWELL-NEXT:    vaddsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addsubpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddsubpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addsubpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddsubpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -135,6 +153,12 @@ define <8 x float> @test_addsubps(<8 x float> %a0, <8 x float> %a1, <8 x float>
 ; HASWELL-NEXT:    vaddsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addsubps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddsubps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addsubps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddsubps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -175,6 +199,13 @@ define <4 x double> @test_andnotpd(<4 x double> %a0, <4 x double> %a1, <4 x doub
 ; HASWELL-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andnotpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandnpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandnpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andnotpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandnpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -223,6 +254,13 @@ define <8 x float> @test_andnotps(<8 x float> %a0, <8 x float> %a1, <8 x float>
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andnotps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandnps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandnps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andnotps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandnps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -271,6 +309,13 @@ define <4 x double> @test_andpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -317,6 +362,13 @@ define <8 x float> @test_andps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -363,6 +415,13 @@ define <4 x double> @test_blendpd(<4 x double> %a0, <4 x double> %a1, <4 x doubl
 ; HASWELL-NEXT:    vblendpd {{.*#+}} ymm0 = ymm0[0],mem[1,2],ymm0[3] sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendpd {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3] sched: [1:0.33]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vblendpd {{.*#+}} ymm0 = ymm0[0],mem[1,2],ymm0[3] sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendpd {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3] sched: [1:0.50]
@@ -402,6 +461,12 @@ define <8 x float> @test_blendps(<8 x float> %a0, <8 x float> %a1, <8 x float> *
 ; HASWELL-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2],ymm0[3],mem[4,5,6],ymm0[7] sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3,4,5,6,7] sched: [1:0.33]
+; SKYLAKE-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2],ymm0[3],mem[4,5,6],ymm0[7] sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2],ymm0[3,4,5,6,7] sched: [1:0.50]
@@ -438,6 +503,12 @@ define <4 x double> @test_blendvpd(<4 x double> %a0, <4 x double> %a1, <4 x doub
 ; HASWELL-NEXT:    vblendvpd %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendvpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendvpd %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vblendvpd %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendvpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendvpd %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:1.00]
@@ -475,6 +546,12 @@ define <8 x float> @test_blendvps(<8 x float> %a0, <8 x float> %a1, <8 x float>
 ; HASWELL-NEXT:    vblendvps %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendvps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendvps %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vblendvps %ymm2, (%rdi), %ymm0, %ymm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendvps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendvps %ymm2, %ymm1, %ymm0, %ymm0 # sched: [2:1.00]
@@ -509,6 +586,11 @@ define <8 x float> @test_broadcastf128(<4 x float> *%a0) {
 ; HASWELL-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_broadcastf128:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_broadcastf128:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] sched: [6:1.00]
@@ -539,6 +621,11 @@ define <4 x double> @test_broadcastsd_ymm(double *%a0) {
 ; HASWELL-NEXT:    vbroadcastsd (%rdi), %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_broadcastsd_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vbroadcastsd (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_broadcastsd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vbroadcastsd (%rdi), %ymm0 # sched: [6:1.00]
@@ -570,6 +657,11 @@ define <4 x float> @test_broadcastss(float *%a0) {
 ; HASWELL-NEXT:    vbroadcastss (%rdi), %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_broadcastss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vbroadcastss (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_broadcastss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vbroadcastss (%rdi), %xmm0 # sched: [5:1.00]
@@ -601,6 +693,11 @@ define <8 x float> @test_broadcastss_ymm(float *%a0) {
 ; HASWELL-NEXT:    vbroadcastss (%rdi), %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_broadcastss_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vbroadcastss (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_broadcastss_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vbroadcastss (%rdi), %ymm0 # sched: [6:1.00]
@@ -638,6 +735,13 @@ define <4 x double> @test_cmppd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vorpd %ymm0, %ymm1, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmppd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vorpd %ymm0, %ymm1, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmppd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqpd %ymm1, %ymm0, %ymm1 # sched: [3:1.00]
@@ -683,6 +787,13 @@ define <8 x float> @test_cmpps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vorps %ymm0, %ymm1, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqps %ymm1, %ymm0, %ymm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vorps %ymm0, %ymm1, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqps %ymm1, %ymm0, %ymm1 # sched: [3:1.00]
@@ -728,6 +839,13 @@ define <4 x double> @test_cvtdq2pd(<4 x i32> %a0, <4 x i32> *%a1) {
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtdq2pd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtdq2pd %xmm0, %ymm0 # sched: [6:1.00]
+; SKYLAKE-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [6:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtdq2pd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtdq2pd (%rdi), %ymm1 # sched: [8:1.00]
@@ -772,6 +890,13 @@ define <8 x float> @test_cvtdq2ps(<8 x i32> %a0, <8 x i32> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtdq2ps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtdq2ps %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtdq2ps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtdq2ps (%rdi), %ymm1 # sched: [8:1.00]
@@ -814,6 +939,13 @@ define <8 x i32> @test_cvtpd2dq(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtpd2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttpd2dq %ymm0, %xmm0 # sched: [6:1.00]
+; SKYLAKE-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtpd2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttpd2dqy (%rdi), %xmm1 # sched: [8:1.00]
@@ -856,6 +988,13 @@ define <8 x float> @test_cvtpd2ps(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtpd2ps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtpd2ps %ymm0, %xmm0 # sched: [6:1.00]
+; SKYLAKE-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtpd2ps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtpd2psy (%rdi), %xmm1 # sched: [8:1.00]
@@ -898,6 +1037,13 @@ define <8 x i32> @test_cvtps2dq(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtps2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttps2dq %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcvttps2dq (%rdi), %ymm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtps2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttps2dq (%rdi), %ymm1 # sched: [8:1.00]
@@ -937,6 +1083,12 @@ define <4 x double> @test_divpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vdivpd (%rdi), %ymm0, %ymm0 # sched: [35:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivpd %ymm1, %ymm0, %ymm0 # sched: [35:2.00]
+; SKYLAKE-NEXT:    vdivpd (%rdi), %ymm0, %ymm0 # sched: [35:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivpd %ymm1, %ymm0, %ymm0 # sched: [38:38.00]
@@ -973,6 +1125,12 @@ define <8 x float> @test_divps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vdivps (%rdi), %ymm0, %ymm0 # sched: [21:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivps %ymm1, %ymm0, %ymm0 # sched: [21:2.00]
+; SKYLAKE-NEXT:    vdivps (%rdi), %ymm0, %ymm0 # sched: [21:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivps %ymm1, %ymm0, %ymm0 # sched: [38:38.00]
@@ -1009,6 +1167,12 @@ define <8 x float> @test_dpps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2
 ; HASWELL-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [14:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_dpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [14:2.00]
+; SKYLAKE-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [14:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_dpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
@@ -1049,6 +1213,13 @@ define <4 x float> @test_extractf128(<8 x float> %a0, <8 x float> %a1, <4 x floa
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_extractf128:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vextractf128 $1, %ymm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vextractf128 $1, %ymm1, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_extractf128:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vextractf128 $1, %ymm0, %xmm0 # sched: [1:0.50]
@@ -1086,6 +1257,12 @@ define <4 x double> @test_haddpd(<4 x double> %a0, <4 x double> %a1, <4 x double
 ; HASWELL-NEXT:    vhaddpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_haddpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhaddpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_haddpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhaddpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1123,6 +1300,12 @@ define <8 x float> @test_haddps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%
 ; HASWELL-NEXT:    vhaddps (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_haddps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhaddps %ymm1, %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhaddps (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_haddps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhaddps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1160,6 +1343,12 @@ define <4 x double> @test_hsubpd(<4 x double> %a0, <4 x double> %a1, <4 x double
 ; HASWELL-NEXT:    vhsubpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_hsubpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhsubpd %ymm1, %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhsubpd (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_hsubpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhsubpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1197,6 +1386,12 @@ define <8 x float> @test_hsubps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%
 ; HASWELL-NEXT:    vhsubps (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_hsubps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhsubps %ymm1, %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhsubps (%rdi), %ymm0, %ymm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_hsubps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhsubps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1237,6 +1432,13 @@ define <8 x float> @test_insertf128(<8 x float> %a0, <4 x float> %a1, <4 x float
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_insertf128:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vinsertf128 $1, (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_insertf128:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1 # sched: [1:0.50]
@@ -1275,6 +1477,11 @@ define <32 x i8> @test_lddqu(i8* %a0) {
 ; HASWELL-NEXT:    vlddqu (%rdi), %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_lddqu:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vlddqu (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_lddqu:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vlddqu (%rdi), %ymm0 # sched: [5:1.00]
@@ -1311,6 +1518,13 @@ define <2 x double> @test_maskmovpd(i8* %a0, <2 x i64> %a1, <2 x double> %a2) {
 ; HASWELL-NEXT:    vmovapd %xmm2, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maskmovpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaskmovpd (%rdi), %xmm0, %xmm2 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vmaskmovpd %xmm1, %xmm0, (%rdi) # sched: [4:1.00]
+; SKYLAKE-NEXT:    vmovapd %xmm2, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maskmovpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaskmovpd (%rdi), %xmm0, %xmm2
@@ -1353,6 +1567,13 @@ define <4 x double> @test_maskmovpd_ymm(i8* %a0, <4 x i64> %a1, <4 x double> %a2
 ; HASWELL-NEXT:    vmovapd %ymm2, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maskmovpd_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaskmovpd (%rdi), %ymm0, %ymm2 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vmaskmovpd %ymm1, %ymm0, (%rdi) # sched: [4:1.00]
+; SKYLAKE-NEXT:    vmovapd %ymm2, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maskmovpd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaskmovpd (%rdi), %ymm0, %ymm2
@@ -1395,6 +1616,13 @@ define <4 x float> @test_maskmovps(i8* %a0, <4 x i32> %a1, <4 x float> %a2) {
 ; HASWELL-NEXT:    vmovaps %xmm2, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maskmovps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaskmovps (%rdi), %xmm0, %xmm2 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vmaskmovps %xmm1, %xmm0, (%rdi) # sched: [4:1.00]
+; SKYLAKE-NEXT:    vmovaps %xmm2, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maskmovps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaskmovps (%rdi), %xmm0, %xmm2
@@ -1437,6 +1665,13 @@ define <8 x float> @test_maskmovps_ymm(i8* %a0, <8 x i32> %a1, <8 x float> %a2)
 ; HASWELL-NEXT:    vmovaps %ymm2, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maskmovps_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaskmovps (%rdi), %ymm0, %ymm2 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vmaskmovps %ymm1, %ymm0, (%rdi) # sched: [4:1.00]
+; SKYLAKE-NEXT:    vmovaps %ymm2, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maskmovps_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaskmovps (%rdi), %ymm0, %ymm2
@@ -1476,6 +1711,12 @@ define <4 x double> @test_maxpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vmaxpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
@@ -1513,6 +1754,12 @@ define <8 x float> @test_maxps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vmaxps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
@@ -1550,6 +1797,12 @@ define <4 x double> @test_minpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vminpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
@@ -1587,6 +1840,12 @@ define <8 x float> @test_minps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vminps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
@@ -1627,6 +1886,13 @@ define <4 x double> @test_movapd(<4 x double> *%a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vmovapd %ymm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movapd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovapd (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovapd %ymm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movapd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovapd (%rdi), %ymm0 # sched: [5:1.00]
@@ -1668,6 +1934,13 @@ define <8 x float> @test_movaps(<8 x float> *%a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vmovaps %ymm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movaps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovaps (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovaps %ymm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movaps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovaps (%rdi), %ymm0 # sched: [5:1.00]
@@ -1709,6 +1982,13 @@ define <4 x double> @test_movddup(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movddup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovddup {{.*#+}} ymm1 = mem[0,0,2,2] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movddup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovddup {{.*#+}} ymm1 = mem[0,0,2,2] sched: [5:1.00]
@@ -1748,6 +2028,12 @@ define i32 @test_movmskpd(<4 x double> %a0) {
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movmskpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovmskpd %ymm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movmskpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovmskpd %ymm0, %eax # sched: [1:0.50]
@@ -1782,6 +2068,12 @@ define i32 @test_movmskps(<8 x float> %a0) {
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movmskps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovmskps %ymm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movmskps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovmskps %ymm0, %eax # sched: [1:0.50]
@@ -1816,6 +2108,12 @@ define <4 x double> @test_movntpd(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovntpd %ymm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1851,6 +2149,12 @@ define <8 x float> @test_movntps(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vmovntps %ymm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovntps %ymm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:2.00]
@@ -1889,6 +2193,13 @@ define <8 x float> @test_movshdup(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movshdup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovshdup {{.*#+}} ymm0 = ymm0[1,1,3,3,5,5,7,7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovshdup {{.*#+}} ymm1 = mem[1,1,3,3,5,5,7,7] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movshdup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovshdup {{.*#+}} ymm1 = mem[1,1,3,3,5,5,7,7] sched: [5:1.00]
@@ -1931,6 +2242,13 @@ define <8 x float> @test_movsldup(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movsldup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovsldup {{.*#+}} ymm0 = ymm0[0,0,2,2,4,4,6,6] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovsldup {{.*#+}} ymm1 = mem[0,0,2,2,4,4,6,6] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movsldup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovsldup {{.*#+}} ymm1 = mem[0,0,2,2,4,4,6,6] sched: [5:1.00]
@@ -1975,6 +2293,13 @@ define <4 x double> @test_movupd(<4 x double> *%a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vmovupd %ymm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movupd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovupd (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovupd %ymm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movupd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovupd (%rdi), %ymm0 # sched: [5:1.00]
@@ -2018,6 +2343,13 @@ define <8 x float> @test_movups(<8 x float> *%a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vmovups %ymm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movups:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovups (%rdi), %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovups %ymm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movups:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovups (%rdi), %ymm0 # sched: [5:1.00]
@@ -2056,6 +2388,12 @@ define <4 x double> @test_mulpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vmulpd (%rdi), %ymm0, %ymm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulpd %ymm1, %ymm0, %ymm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulpd (%rdi), %ymm0, %ymm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulpd %ymm1, %ymm0, %ymm0 # sched: [4:4.00]
@@ -2092,6 +2430,12 @@ define <8 x float> @test_mulps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vmulps (%rdi), %ymm0, %ymm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulps %ymm1, %ymm0, %ymm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulps (%rdi), %ymm0, %ymm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulps %ymm1, %ymm0, %ymm0 # sched: [2:2.00]
@@ -2131,6 +2475,13 @@ define <4 x double> @orpd(<4 x double> %a0, <4 x double> %a1, <4 x double> *%a2)
 ; HASWELL-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: orpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vorpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vorpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: orpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vorpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -2177,6 +2528,13 @@ define <8 x float> @test_orps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_orps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vorps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_orps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -2223,6 +2581,13 @@ define <2 x double> @test_permilpd(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilpd {{.*#+}} xmm1 = mem[1,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilpd {{.*#+}} xmm1 = mem[1,0] sched: [6:1.00]
@@ -2265,6 +2630,13 @@ define <4 x double> @test_permilpd_ymm(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilpd_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilpd {{.*#+}} ymm0 = ymm0[1,0,2,3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilpd {{.*#+}} ymm1 = mem[1,0,2,3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilpd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilpd {{.*#+}} ymm1 = mem[1,0,2,3] sched: [6:1.00]
@@ -2307,6 +2679,13 @@ define <4 x float> @test_permilps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,2,1,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilps {{.*#+}} xmm1 = mem[3,2,1,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilps {{.*#+}} xmm1 = mem[3,2,1,0] sched: [6:1.00]
@@ -2349,6 +2728,13 @@ define <8 x float> @test_permilps_ymm(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilps_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilps {{.*#+}} ymm1 = mem[3,2,1,0,7,6,5,4] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilps_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilps {{.*#+}} ymm1 = mem[3,2,1,0,7,6,5,4] sched: [6:1.00]
@@ -2388,6 +2774,12 @@ define <2 x double> @test_permilvarpd(<2 x double> %a0, <2 x i64> %a1, <2 x i64>
 ; HASWELL-NEXT:    vpermilpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilvarpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilvarpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2425,6 +2817,12 @@ define <4 x double> @test_permilvarpd_ymm(<4 x double> %a0, <4 x i64> %a1, <4 x
 ; HASWELL-NEXT:    vpermilpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilvarpd_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilvarpd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -2462,6 +2860,12 @@ define <4 x float> @test_permilvarps(<4 x float> %a0, <4 x i32> %a1, <4 x i32> *
 ; HASWELL-NEXT:    vpermilps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilvarps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilvarps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2499,6 +2903,12 @@ define <8 x float> @test_permilvarps_ymm(<8 x float> %a0, <8 x i32> %a1, <8 x i3
 ; HASWELL-NEXT:    vpermilps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_permilvarps_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpermilps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpermilps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_permilvarps_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpermilps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -2539,6 +2949,13 @@ define <8 x float> @test_rcpps(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rcpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrcpps (%rdi), %ymm1 # sched: [11:2.00]
+; SKYLAKE-NEXT:    vrcpps %ymm0, %ymm0 # sched: [11:2.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rcpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vrcpps (%rdi), %ymm1 # sched: [7:2.00]
@@ -2582,6 +2999,13 @@ define <4 x double> @test_roundpd(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundpd $7, %ymm0, %ymm0 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundpd $7, (%rdi), %ymm1 # sched: [8:1.00]
@@ -2625,6 +3049,13 @@ define <8 x float> @test_roundps(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundps $7, %ymm0, %ymm0 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundps $7, (%rdi), %ymm1 # sched: [8:1.00]
@@ -2668,6 +3099,13 @@ define <8 x float> @test_rsqrtps(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rsqrtps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrsqrtps (%rdi), %ymm1 # sched: [11:2.00]
+; SKYLAKE-NEXT:    vrsqrtps %ymm0, %ymm0 # sched: [11:2.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rsqrtps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vrsqrtps (%rdi), %ymm1 # sched: [7:2.00]
@@ -2711,6 +3149,13 @@ define <4 x double> @test_shufpd(<4 x double> %a0, <4 x double> %a1, <4 x double
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shufpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1],ymm1[0],ymm0[2],ymm1[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vshufpd {{.*#+}} ymm1 = ymm1[1],mem[0],ymm1[2],mem[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_shufpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1],ymm1[0],ymm0[2],ymm1[3] sched: [1:0.50]
@@ -2750,6 +3195,12 @@ define <8 x float> @test_shufps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%
 ; HASWELL-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,3],mem[0,0],ymm0[4,7],mem[4,4] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shufps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0],ymm1[0,0],ymm0[4,4],ymm1[4,4] sched: [1:1.00]
+; SKYLAKE-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,3],mem[0,0],ymm0[4,7],mem[4,4] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_shufps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,0],ymm1[0,0],ymm0[4,4],ymm1[4,4] sched: [1:0.50]
@@ -2789,6 +3240,13 @@ define <4 x double> @test_sqrtpd(<4 x double> %a0, <4 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtpd (%rdi), %ymm1 # sched: [35:2.00]
+; SKYLAKE-NEXT:    vsqrtpd %ymm0, %ymm0 # sched: [35:2.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsqrtpd (%rdi), %ymm1 # sched: [59:54.00]
@@ -2832,6 +3290,13 @@ define <8 x float> @test_sqrtps(<8 x float> %a0, <8 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtps (%rdi), %ymm1 # sched: [21:2.00]
+; SKYLAKE-NEXT:    vsqrtps %ymm0, %ymm0 # sched: [21:2.00]
+; SKYLAKE-NEXT:    vaddps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsqrtps (%rdi), %ymm1 # sched: [47:42.00]
@@ -2872,6 +3337,12 @@ define <4 x double> @test_subpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubpd (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubpd %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -2908,6 +3379,12 @@ define <8 x float> @test_subps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubps %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubps (%rdi), %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubps %ymm1, %ymm0, %ymm0 # sched: [3:2.00]
@@ -2953,6 +3430,15 @@ define i32 @test_testpd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2) {
 ; HASWELL-NEXT:    adcl $0, %eax # sched: [2:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_testpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    xorl %eax, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    vtestpd %xmm1, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    setb %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    vtestpd (%rdi), %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    adcl $0, %eax # sched: [2:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_testpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
@@ -3009,6 +3495,16 @@ define i32 @test_testpd_ymm(<4 x double> %a0, <4 x double> %a1, <4 x double> *%a
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_testpd_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    xorl %eax, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    vtestpd %ymm1, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    setb %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    vtestpd (%rdi), %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    adcl $0, %eax # sched: [2:0.50]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_testpd_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
@@ -3063,6 +3559,15 @@ define i32 @test_testps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {
 ; HASWELL-NEXT:    adcl $0, %eax # sched: [2:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_testps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    xorl %eax, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    vtestps %xmm1, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    setb %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    vtestps (%rdi), %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    adcl $0, %eax # sched: [2:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_testps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
@@ -3119,6 +3624,16 @@ define i32 @test_testps_ymm(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a2)
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_testps_ymm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    xorl %eax, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    vtestps %ymm1, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    setb %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    vtestps (%rdi), %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    adcl $0, %eax # sched: [2:0.50]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_testps_ymm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    xorl %eax, %eax # sched: [1:0.50]
@@ -3167,6 +3682,13 @@ define <4 x double> @test_unpckhpd(<4 x double> %a0, <4 x double> %a1, <4 x doub
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpckhpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],mem[1],ymm1[3],mem[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpckhpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3] sched: [1:0.50]
@@ -3206,6 +3728,12 @@ define <8 x float> @test_unpckhps(<8 x float> %a0, <8 x float> %a1, <8 x float>
 ; HASWELL-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpckhps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],mem[2],ymm0[3],mem[3],ymm0[6],mem[6],ymm0[7],mem[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpckhps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7] sched: [1:0.50]
@@ -3245,6 +3773,13 @@ define <4 x double> @test_unpcklpd(<4 x double> %a0, <4 x double> %a1, <4 x doub
 ; HASWELL-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpcklpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} ymm1 = ymm1[0],mem[0],ymm1[2],mem[2] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpcklpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] sched: [1:0.50]
@@ -3284,6 +3819,12 @@ define <8 x float> @test_unpcklps(<8 x float> %a0, <8 x float> %a1, <8 x float>
 ; HASWELL-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpcklps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[4],mem[4],ymm0[5],mem[5] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpcklps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5] sched: [1:0.50]
@@ -3323,6 +3864,13 @@ define <4 x double> @test_xorpd(<4 x double> %a0, <4 x double> %a1, <4 x double>
 ; HASWELL-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_xorpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vxorpd %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vxorpd (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_xorpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vxorpd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -3369,6 +3917,13 @@ define <8 x float> @test_xorps(<8 x float> %a0, <8 x float> %a1, <8 x float> *%a
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_xorps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vxorps %ymm1, %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vxorps (%rdi), %ymm0, %ymm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_xorps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vxorps %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
@@ -3409,6 +3964,11 @@ define void @test_zeroall() {
 ; HASWELL-NEXT:    vzeroall # sched: [16:16.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_zeroall:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vzeroall # sched: [16:16.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_zeroall:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vzeroall # sched: [90:?]
@@ -3439,6 +3999,11 @@ define void @test_zeroupper() {
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_zeroupper:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_zeroupper:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vzeroupper # sched: [46:?]

diff --git a/test/CodeGen/X86/avx2-schedule.ll b/test/CodeGen/X86/avx2-schedule.ll
index 7dbe2dd98e8c60049604ef15a7da61f2aa5a9908..82a73e7740d085c6962618ac8085f8bccdf0a0eb 100644 (file)
--- a/test/CodeGen/X86/avx2-schedule.ll
+++ b/test/CodeGen/X86/avx2-schedule.ll
@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=x86-64 -mattr=+avx2 | FileCheck %s --check-prefix=CHECK --check-prefix=GENERIC
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
 define <32 x i8> @test_pabsb(<32 x i8> %a0, <32 x i8> *%a1) {
@@ -19,6 +19,13 @@ define <32 x i8> @test_pabsb(<32 x i8> %a0, <32 x i8> *%a1) {
 ; HASWELL-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsb %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsb (%rdi), %ymm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pabsb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpabsb (%rdi), %ymm1 # sched: [8:0.50]
@@ -48,6 +55,13 @@ define <8 x i32> @test_pabsd(<8 x i32> %a0, <8 x i32> *%a1) {
 ; HASWELL-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsd %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsd (%rdi), %ymm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pabsd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpabsd (%rdi), %ymm1 # sched: [8:0.50]
@@ -77,6 +91,13 @@ define <16 x i16> @test_pabsw(<16 x i16> %a0, <16 x i16> *%a1) {
 ; HASWELL-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsw %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsw (%rdi), %ymm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pabsw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpabsw (%rdi), %ymm1 # sched: [8:0.50]
@@ -104,6 +125,12 @@ define <32 x i8> @test_paddb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) {
 ; HASWELL-NEXT:    vpaddb (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddb %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddb (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_paddb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpaddb %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -128,6 +155,12 @@ define <8 x i32> @test_paddd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) {
 ; HASWELL-NEXT:    vpaddd (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddd (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_paddd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpaddd %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -152,6 +185,12 @@ define <4 x i64> @test_paddq(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_paddq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -176,6 +215,12 @@ define <16 x i16> @test_paddw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) {
 ; HASWELL-NEXT:    vpaddw (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddw %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddw (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_paddw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpaddw %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -202,6 +247,13 @@ define <4 x i64> @test_pand(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pand:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpand %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpand (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pand:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpand %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -230,6 +282,13 @@ define <4 x i64> @test_pandn(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pandn:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpandn %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpandn (%rdi), %ymm0, %ymm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pandn:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpandn %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -258,6 +317,12 @@ define <8 x i32> @test_pmulld(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmulld (%rdi), %ymm0, %ymm0 # sched: [10:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmulld:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmulld %ymm1, %ymm0, %ymm0 # sched: [10:2.00]
+; SKYLAKE-NEXT:    vpmulld (%rdi), %ymm0, %ymm0 # sched: [10:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pmulld:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpmulld %ymm1, %ymm0, %ymm0 # sched: [4:1.00]
@@ -282,6 +347,12 @@ define <16 x i16> @test_pmullw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2)
 ; HASWELL-NEXT:    vpmullw (%rdi), %ymm0, %ymm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmullw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmullw %ymm1, %ymm0, %ymm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmullw (%rdi), %ymm0, %ymm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pmullw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpmullw %ymm1, %ymm0, %ymm0 # sched: [4:1.00]
@@ -308,6 +379,13 @@ define <4 x i64> @test_por(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_por:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpor (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_por:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpor %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -334,6 +412,12 @@ define <32 x i8> @test_psubb(<32 x i8> %a0, <32 x i8> %a1, <32 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsubb (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubb %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubb (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_psubb:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpsubb %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -358,6 +442,12 @@ define <8 x i32> @test_psubd(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> *%a2) {
 ; HASWELL-NEXT:    vpsubd (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubd %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubd (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_psubd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpsubd %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -382,6 +472,12 @@ define <4 x i64> @test_psubq(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpsubq (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubq (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_psubq:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpsubq %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -406,6 +502,12 @@ define <16 x i16> @test_psubw(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsubw (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubw %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubw (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_psubw:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpsubw %ymm1, %ymm0, %ymm0 # sched: [1:0.25]
@@ -432,6 +534,13 @@ define <4 x i64> @test_pxor(<4 x i64> %a0, <4 x i64> %a1, <4 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pxor:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpxor %ymm1, %ymm0, %ymm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpxor (%rdi), %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %ymm1, %ymm0, %ymm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pxor:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vpxor %ymm1, %ymm0, %ymm0 # sched: [1:0.25]

diff --git a/test/CodeGen/X86/f16c-schedule.ll b/test/CodeGen/X86/f16c-schedule.ll
index 24b7da2a22918a67fca88bfbcff0e9808a25dbb5..a439ec3bfe7c5529709dcc6bc4861a51a2182cee 100644 (file)
--- a/test/CodeGen/X86/f16c-schedule.ll
+++ b/test/CodeGen/X86/f16c-schedule.ll
@@ -2,7 +2,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=x86-64 -mattr=+f16c | FileCheck %s --check-prefix=CHECK --check-prefix=GENERIC
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=IVY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -28,6 +28,13 @@ define <4 x float> @test_vcvtph2ps_128(<8 x i16> %a0, <8 x i16> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_vcvtph2ps_128:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtph2ps (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vcvtph2ps %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_vcvtph2ps_128:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtph2ps (%rdi), %xmm1 # sched: [8:1.00]
@@ -71,6 +78,13 @@ define <8 x float> @test_vcvtph2ps_256(<8 x i16> %a0, <8 x i16> *%a1) {
 ; HASWELL-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_vcvtph2ps_256:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtph2ps (%rdi), %ymm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vcvtph2ps %xmm0, %ymm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_vcvtph2ps_256:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtph2ps (%rdi), %ymm1 # sched: [8:1.00]
@@ -111,6 +125,12 @@ define <8 x i16> @test_vcvtps2ph_128(<4 x float> %a0, <4 x float> %a1, <4 x i16>
 ; HASWELL-NEXT:    vcvtps2ph $0, %xmm1, (%rdi) # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_vcvtps2ph_128:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtps2ph $0, %xmm1, (%rdi) # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_vcvtps2ph_128:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # sched: [3:1.00]
@@ -152,6 +172,13 @@ define <8 x i16> @test_vcvtps2ph_256(<8 x float> %a0, <8 x float> %a1, <8 x i16>
 ; HASWELL-NEXT:    vzeroupper # sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_vcvtps2ph_256:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # sched: [6:1.00]
+; SKYLAKE-NEXT:    vcvtps2ph $0, %ymm1, (%rdi) # sched: [6:1.00]
+; SKYLAKE-NEXT:    vzeroupper # sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_vcvtps2ph_256:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # sched: [3:1.00]

diff --git a/test/CodeGen/X86/sse-schedule.ll b/test/CodeGen/X86/sse-schedule.ll
index 8eb7b3f35546d0f3b7d160c74ac84bf08efc00d3..e982cc9aa490927fee35a939a2818e3bf5349f4c 100644 (file)
--- a/test/CodeGen/X86/sse-schedule.ll
+++ b/test/CodeGen/X86/sse-schedule.ll
@@ -5,7 +5,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -40,6 +40,12 @@ define <4 x float> @test_addps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -88,6 +94,12 @@ define float @test_addss(float %a0, float %a1, float *%a2) {
 ; HASWELL-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -140,6 +152,12 @@ define <4 x float> @test_andps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -196,6 +214,12 @@ define <4 x float> @test_andnotps(<4 x float> %a0, <4 x float> %a1, <4 x float>
 ; HASWELL-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andnotps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandnps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andnotps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandnps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -255,6 +279,13 @@ define <4 x float> @test_cmpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vorps %xmm0, %xmm1, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqps %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
@@ -308,6 +339,12 @@ define float @test_cmpss(float %a0, float %a1, float *%a2) {
 ; HASWELL-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmpss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqss (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmpss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -401,6 +438,20 @@ define i32 @test_comiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {
 ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_comiss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %cl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %cl # sched: [1:0.25]
+; SKYLAKE-NEXT:    vcomiss (%rdi), %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %dl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    orb %cl, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_comiss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcomiss %xmm1, %xmm0 # sched: [3:1.00]
@@ -472,6 +523,13 @@ define float @test_cvtsi2ss(i32 %a0, i32 *%a1) {
 ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsi2ss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtsi2ssl (%rsi), %xmm1, %xmm1 # sched: [8:1.00]
+; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsi2ss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsi2ssl %edi, %xmm0, %xmm0 # sched: [3:1.00]
@@ -528,6 +586,13 @@ define float @test_cvtsi2ssq(i64 %a0, i64 *%a1) {
 ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsi2ssq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vcvtsi2ssq (%rsi), %xmm1, %xmm1 # sched: [8:1.00]
+; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsi2ssq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsi2ssq %rdi, %xmm0, %xmm0 # sched: [3:1.00]
@@ -584,6 +649,13 @@ define i32 @test_cvtss2si(float %a0, float *%a1) {
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtss2si:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtss2si %xmm0, %ecx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtss2si (%rdi), %eax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtss2si:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtss2si (%rdi), %eax # sched: [8:1.00]
@@ -643,6 +715,13 @@ define i64 @test_cvtss2siq(float %a0, float *%a1) {
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtss2siq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtss2si %xmm0, %rcx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtss2si (%rdi), %rax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtss2siq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtss2si (%rdi), %rax # sched: [8:1.00]
@@ -702,6 +781,13 @@ define i32 @test_cvttss2si(float %a0, float *%a1) {
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttss2si:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttss2si %xmm0, %ecx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvttss2si (%rdi), %eax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttss2si:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttss2si (%rdi), %eax # sched: [8:1.00]
@@ -758,6 +844,13 @@ define i64 @test_cvttss2siq(float %a0, float *%a1) {
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttss2siq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttss2si %xmm0, %rcx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvttss2si (%rdi), %rax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttss2siq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttss2si (%rdi), %rax # sched: [8:1.00]
@@ -809,6 +902,12 @@ define <4 x float> @test_divps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [13:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [13:1.00]
+; SKYLAKE-NEXT:    vdivps (%rdi), %xmm0, %xmm0 # sched: [13:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivps %xmm1, %xmm0, %xmm0 # sched: [19:19.00]
@@ -857,6 +956,12 @@ define float @test_divss(float %a0, float %a1, float *%a2) {
 ; HASWELL-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [13:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [13:1.00]
+; SKYLAKE-NEXT:    vdivss (%rdi), %xmm0, %xmm0 # sched: [13:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivss %xmm1, %xmm0, %xmm0 # sched: [19:19.00]
@@ -905,6 +1010,12 @@ define void @test_ldmxcsr(i32 %a0) {
 ; HASWELL-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_ldmxcsr:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
+; SKYLAKE-NEXT:    vldmxcsr -{{[0-9]+}}(%rsp) # sched: [2:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_ldmxcsr:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
@@ -955,6 +1066,12 @@ define <4 x float> @test_maxps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1004,6 +1121,12 @@ define <4 x float> @test_maxss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1053,6 +1176,12 @@ define <4 x float> @test_minps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1102,6 +1231,12 @@ define <4 x float> @test_minss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1156,6 +1291,13 @@ define void @test_movaps(<4 x float> *%a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movaps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovaps (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovaps %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movaps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovaps (%rdi), %xmm0 # sched: [5:1.00]
@@ -1209,6 +1351,11 @@ define <4 x float> @test_movhlps(<4 x float> %a0, <4 x float> %a1) {
 ; HASWELL-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movhlps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movhlps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:0.50]
@@ -1262,6 +1409,13 @@ define void @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {
 ; HASWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movhps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movhps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
@@ -1318,6 +1472,12 @@ define <4 x float> @test_movlhps(<4 x float> %a0, <4 x float> %a1) {
 ; HASWELL-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movlhps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movlhps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50]
@@ -1370,6 +1530,13 @@ define void @test_movlps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {
 ; HASWELL-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movlps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovlps %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movlps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00]
@@ -1421,6 +1588,11 @@ define i32 @test_movmskps(<4 x float> %a0) {
 ; HASWELL-NEXT:    vmovmskps %xmm0, %eax # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movmskps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovmskps %xmm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movmskps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovmskps %xmm0, %eax # sched: [1:0.50]
@@ -1467,6 +1639,11 @@ define void @test_movntps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovntps %xmm0, (%rdi) # sched: [1:1.00]
@@ -1516,6 +1693,13 @@ define void @test_movss_mem(float* %a0, float* %a1) {
 ; HASWELL-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movss_mem:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddss %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovss %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movss_mem:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:1.00]
@@ -1567,6 +1751,11 @@ define <4 x float> @test_movss_reg(<4 x float> %a0, <4 x float> %a1) {
 ; HASWELL-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movss_reg:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movss_reg:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50]
@@ -1616,6 +1805,13 @@ define void @test_movups(<4 x float> *%a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movups:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovups (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovups %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movups:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovups (%rdi), %xmm0 # sched: [5:1.00]
@@ -1666,6 +1862,12 @@ define <4 x float> @test_mulps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulps (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulps %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -1714,6 +1916,12 @@ define float @test_mulss(float %a0, float %a1, float *%a2) {
 ; HASWELL-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulss (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulss %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -1766,6 +1974,12 @@ define <4 x float> @test_orps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2
 ; HASWELL-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_orps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_orps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1819,6 +2033,11 @@ define void @test_prefetchnta(i8* %a0) {
 ; HASWELL-NEXT:    prefetchnta (%rdi) # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_prefetchnta:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    prefetchnta (%rdi) # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_prefetchnta:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    prefetchnta (%rdi) # sched: [5:1.00]
@@ -1871,6 +2090,13 @@ define <4 x float> @test_rcpps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rcpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrcpps %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vrcpps (%rdi), %xmm1 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rcpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vrcpps (%rdi), %xmm1 # sched: [7:1.00]
@@ -1935,6 +2161,14 @@ define <4 x float> @test_rcpss(float %a0, float *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rcpss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrcpss %xmm0, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vrcpss %xmm1, %xmm1, %xmm1 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rcpss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00]
@@ -1998,6 +2232,13 @@ define <4 x float> @test_rsqrtps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rsqrtps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrsqrtps %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vrsqrtps (%rdi), %xmm1 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rsqrtps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vrsqrtps (%rdi), %xmm1 # sched: [7:1.00]
@@ -2062,6 +2303,14 @@ define <4 x float> @test_rsqrtss(float %a0, float *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rsqrtss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vrsqrtss %xmm0, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vrsqrtss %xmm1, %xmm1, %xmm1 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_rsqrtss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00]
@@ -2119,6 +2368,11 @@ define void @test_sfence() {
 ; HASWELL-NEXT:    sfence # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sfence:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    sfence # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sfence:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    sfence # sched: [1:1.00]
@@ -2168,6 +2422,12 @@ define <4 x float> @test_shufps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%
 ; HASWELL-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shufps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_shufps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:0.50]
@@ -2222,6 +2482,13 @@ define <4 x float> @test_sqrtps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtps %xmm0, %xmm0 # sched: [14:1.00]
+; SKYLAKE-NEXT:    vsqrtps (%rdi), %xmm1 # sched: [14:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsqrtps (%rdi), %xmm1 # sched: [26:21.00]
@@ -2286,6 +2553,14 @@ define <4 x float> @test_sqrtss(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtss %xmm0, %xmm0, %xmm0 # sched: [14:1.00]
+; SKYLAKE-NEXT:    vmovaps (%rdi), %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vsqrtss %xmm1, %xmm1, %xmm1 # sched: [14:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovaps (%rdi), %xmm1 # sched: [5:1.00]
@@ -2340,6 +2615,12 @@ define i32 @test_stmxcsr() {
 ; HASWELL-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_stmxcsr:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [1:1.00]
+; SKYLAKE-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_stmxcsr:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vstmxcsr -{{[0-9]+}}(%rsp) # sched: [1:1.00]
@@ -2390,6 +2671,12 @@ define <4 x float> @test_subps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -2438,6 +2725,12 @@ define float @test_subss(float %a0, float %a1, float *%a2) {
 ; HASWELL-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubss (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -2526,6 +2819,20 @@ define i32 @test_ucomiss(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2) {
 ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_ucomiss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %cl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %cl # sched: [1:0.25]
+; SKYLAKE-NEXT:    vucomiss (%rdi), %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %dl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    orb %cl, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_ucomiss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vucomiss %xmm1, %xmm0 # sched: [3:1.00]
@@ -2596,6 +2903,12 @@ define <4 x float> @test_unpckhps(<4 x float> %a0, <4 x float> %a1, <4 x float>
 ; HASWELL-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpckhps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpckhps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50]
@@ -2648,6 +2961,12 @@ define <4 x float> @test_unpcklps(<4 x float> %a0, <4 x float> %a1, <4 x float>
 ; HASWELL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpcklps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpcklps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50]
@@ -2700,6 +3019,12 @@ define <4 x float> @test_xorps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a
 ; HASWELL-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_xorps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vxorps (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_xorps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vxorps %xmm1, %xmm0, %xmm0 # sched: [1:0.50]

diff --git a/test/CodeGen/X86/sse2-schedule.ll b/test/CodeGen/X86/sse2-schedule.ll
index 23533a7bf53c0f1ebde6a7934cc7e7fb660070f8..bd79776a040a47371d65c9837e8199f6bf30c6ee 100644 (file)
--- a/test/CodeGen/X86/sse2-schedule.ll
+++ b/test/CodeGen/X86/sse2-schedule.ll
@@ -5,7 +5,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -40,6 +40,12 @@ define <2 x double> @test_addpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vaddpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -88,6 +94,12 @@ define double @test_addsd(double %a0, double %a1, double *%a2) {
 ; HASWELL-NEXT:    vaddsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -141,6 +153,13 @@ define <2 x double> @test_andpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -201,6 +220,13 @@ define <2 x double> @test_andnotpd(<2 x double> %a0, <2 x double> %a1, <2 x doub
 ; HASWELL-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_andnotpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vandnpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vandnpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_andnotpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vandnpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -263,6 +289,13 @@ define <2 x double> @test_cmppd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vorpd %xmm0, %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmppd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vorpd %xmm0, %xmm1, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmppd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqpd %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
@@ -316,6 +349,12 @@ define double @test_cmpsd(double %a0, double %a1, double *%a2) {
 ; HASWELL-NEXT:    vcmpeqsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cmpsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcmpeqsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcmpeqsd (%rdi), %xmm0, %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cmpsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcmpeqsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -409,6 +448,20 @@ define i32 @test_comisd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2) {
 ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_comisd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcomisd %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %cl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %cl # sched: [1:0.25]
+; SKYLAKE-NEXT:    vcomisd (%rdi), %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %dl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    orb %cl, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_comisd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcomisd %xmm1, %xmm0 # sched: [3:1.00]
@@ -480,6 +533,13 @@ define <2 x double> @test_cvtdq2pd(<4 x i32> %a0, <4 x i32> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtdq2pd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtdq2pd %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtdq2pd (%rdi), %xmm1 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtdq2pd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtdq2pd (%rdi), %xmm1 # sched: [8:1.00]
@@ -539,6 +599,13 @@ define <4 x float> @test_cvtdq2ps(<4 x i32> %a0, <4 x i32> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtdq2ps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtdq2ps %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcvtdq2ps (%rdi), %xmm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtdq2ps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtdq2ps (%rdi), %xmm1 # sched: [8:1.00]
@@ -596,6 +663,13 @@ define <4 x i32> @test_cvtpd2dq(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtpd2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtpd2dq %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtpd2dqx (%rdi), %xmm1 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtpd2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtpd2dqx (%rdi), %xmm1 # sched: [8:1.00]
@@ -654,6 +728,13 @@ define <4 x float> @test_cvtpd2ps(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtpd2ps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtpd2ps %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtpd2psx (%rdi), %xmm1 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtpd2ps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtpd2psx (%rdi), %xmm1 # sched: [8:1.00]
@@ -712,6 +793,13 @@ define <4 x i32> @test_cvtps2dq(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtps2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtps2dq %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcvtps2dq (%rdi), %xmm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtps2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtps2dq (%rdi), %xmm1 # sched: [8:1.00]
@@ -770,6 +858,13 @@ define <2 x double> @test_cvtps2pd(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtps2pd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtps2pd %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vcvtps2pd (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtps2pd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtps2pd (%rdi), %xmm1 # sched: [8:1.00]
@@ -828,6 +923,13 @@ define i32 @test_cvtsd2si(double %a0, double *%a1) {
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsd2si:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsd2si %xmm0, %ecx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtsd2si (%rdi), %eax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsd2si:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsd2si (%rdi), %eax # sched: [8:1.00]
@@ -887,6 +989,13 @@ define i64 @test_cvtsd2siq(double %a0, double *%a1) {
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsd2siq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsd2si %xmm0, %rcx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtsd2si (%rdi), %rax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsd2siq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsd2si (%rdi), %rax # sched: [8:1.00]
@@ -952,6 +1061,14 @@ define float @test_cvtsd2ss(double %a0, double *%a1) {
 ; HASWELL-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsd2ss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsd2ss %xmm0, %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vcvtsd2ss %xmm1, %xmm1, %xmm1 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vaddss %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsd2ss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero sched: [5:1.00]
@@ -1010,6 +1127,13 @@ define double @test_cvtsi2sd(i32 %a0, i32 *%a1) {
 ; HASWELL-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsi2sd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsi2sdl %edi, %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtsi2sdl (%rsi), %xmm1, %xmm1 # sched: [8:1.00]
+; SKYLAKE-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsi2sd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsi2sdl %edi, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1066,6 +1190,13 @@ define double @test_cvtsi2sdq(i64 %a0, i64 *%a1) {
 ; HASWELL-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtsi2sdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtsi2sdq %rdi, %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvtsi2sdq (%rsi), %xmm1, %xmm1 # sched: [8:1.00]
+; SKYLAKE-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtsi2sdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvtsi2sdq %rdi, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1130,6 +1261,14 @@ define double @test_cvtss2sd(float %a0, float *%a1) {
 ; HASWELL-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvtss2sd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vcvtss2sd %xmm1, %xmm1, %xmm1 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vaddsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvtss2sd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00]
@@ -1189,6 +1328,13 @@ define <4 x i32> @test_cvttpd2dq(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttpd2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttpd2dq %xmm0, %xmm0 # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvttpd2dqx (%rdi), %xmm1 # sched: [7:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttpd2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttpd2dqx (%rdi), %xmm1 # sched: [8:1.00]
@@ -1248,6 +1394,13 @@ define <4 x i32> @test_cvttps2dq(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttps2dq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttps2dq %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vcvttps2dq (%rdi), %xmm1 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttps2dq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttps2dq (%rdi), %xmm1 # sched: [8:1.00]
@@ -1304,6 +1457,13 @@ define i32 @test_cvttsd2si(double %a0, double *%a1) {
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttsd2si:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttsd2si %xmm0, %ecx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvttsd2si (%rdi), %eax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttsd2si:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttsd2si (%rdi), %eax # sched: [8:1.00]
@@ -1360,6 +1520,13 @@ define i64 @test_cvttsd2siq(double %a0, double *%a1) {
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_cvttsd2siq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vcvttsd2si %xmm0, %rcx # sched: [4:1.00]
+; SKYLAKE-NEXT:    vcvttsd2si (%rdi), %rax # sched: [4:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_cvttsd2siq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vcvttsd2si (%rdi), %rax # sched: [8:1.00]
@@ -1411,6 +1578,12 @@ define <2 x double> @test_divpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vdivpd (%rdi), %xmm0, %xmm0 # sched: [20:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivpd %xmm1, %xmm0, %xmm0 # sched: [20:1.00]
+; SKYLAKE-NEXT:    vdivpd (%rdi), %xmm0, %xmm0 # sched: [20:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivpd %xmm1, %xmm0, %xmm0 # sched: [19:19.00]
@@ -1459,6 +1632,12 @@ define double @test_divsd(double %a0, double %a1, double *%a2) {
 ; HASWELL-NEXT:    vdivsd (%rdi), %xmm0, %xmm0 # sched: [20:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_divsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdivsd %xmm1, %xmm0, %xmm0 # sched: [20:1.00]
+; SKYLAKE-NEXT:    vdivsd (%rdi), %xmm0, %xmm0 # sched: [20:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_divsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdivsd %xmm1, %xmm0, %xmm0 # sched: [19:19.00]
@@ -1508,6 +1687,11 @@ define void @test_lfence() {
 ; HASWELL-NEXT:    lfence # sched: [2:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_lfence:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    lfence # sched: [2:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_lfence:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    lfence # sched: [1:1.00]
@@ -1554,6 +1738,11 @@ define void @test_mfence() {
 ; HASWELL-NEXT:    mfence # sched: [2:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mfence:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    mfence # sched: [2:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mfence:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    mfence # sched: [1:1.00]
@@ -1598,6 +1787,11 @@ define void @test_maskmovdqu(<16 x i8> %a0, <16 x i8> %a1, i8* %a2) {
 ; HASWELL-NEXT:    vmaskmovdqu %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maskmovdqu:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaskmovdqu %xmm1, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maskmovdqu:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaskmovdqu %xmm1, %xmm0 # sched: [1:1.00]
@@ -1643,6 +1837,12 @@ define <2 x double> @test_maxpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1692,6 +1892,12 @@ define <2 x double> @test_maxsd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_maxsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmaxsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_maxsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmaxsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1741,6 +1947,12 @@ define <2 x double> @test_minpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1790,6 +2002,12 @@ define <2 x double> @test_minsd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_minsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vminsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_minsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vminsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -1844,6 +2062,13 @@ define void @test_movapd(<2 x double> *%a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vmovapd %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movapd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovapd (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovapd %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movapd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovapd (%rdi), %xmm0 # sched: [5:1.00]
@@ -1899,6 +2124,13 @@ define void @test_movdqa(<2 x i64> *%a0, <2 x i64> *%a1) {
 ; HASWELL-NEXT:    vmovdqa %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movdqa:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovdqa (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovdqa %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movdqa:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovdqa (%rdi), %xmm0 # sched: [5:1.00]
@@ -1954,6 +2186,13 @@ define void @test_movdqu(<2 x i64> *%a0, <2 x i64> *%a1) {
 ; HASWELL-NEXT:    vmovdqu %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movdqu:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovdqu (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovdqu %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movdqu:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovdqu (%rdi), %xmm0 # sched: [5:1.00]
@@ -2024,6 +2263,16 @@ define i32 @test_movd(<4 x i32> %a0, i32 %a1, i32 *%a2) {
 ; HASWELL-NEXT:    vmovd %xmm1, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovd %edi, %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddd %xmm2, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovd %xmm0, %eax # sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovd %xmm1, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:1.00]
@@ -2105,6 +2354,16 @@ define i64 @test_movd_64(<2 x i64> %a0, i64 %a1, i64 *%a2) {
 ; HASWELL-NEXT:    vmovq %xmm1, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movd_64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovq %rdi, %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm2, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovq %xmm0, %rax # sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovq %xmm1, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movd_64:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero sched: [5:1.00]
@@ -2171,6 +2430,13 @@ define void @test_movhpd(<2 x double> %a0, <2 x double> %a1, x86_mmx *%a2) {
 ; HASWELL-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movhpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movhpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
@@ -2229,6 +2495,13 @@ define void @test_movlpd(<2 x double> %a0, <2 x double> %a1, x86_mmx *%a2) {
 ; HASWELL-NEXT:    vmovlpd %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movlpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovlpd %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movlpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [6:1.00]
@@ -2279,6 +2552,11 @@ define i32 @test_movmskpd(<2 x double> %a0) {
 ; HASWELL-NEXT:    vmovmskpd %xmm0, %eax # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movmskpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovmskpd %xmm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movmskpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovmskpd %xmm0, %eax # sched: [1:0.50]
@@ -2326,6 +2604,12 @@ define void @test_movntdqa(<2 x i64> %a0, <2 x i64> *%a1) {
 ; HASWELL-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntdqa:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovntdq %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntdqa:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddq %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2373,6 +2657,12 @@ define void @test_movntpd(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovntpd %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
@@ -2425,6 +2715,13 @@ define <2 x i64> @test_movq_mem(<2 x i64> %a0, i64 *%a1) {
 ; HASWELL-NEXT:    vmovq %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movq_mem:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vmovq %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movq_mem:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero sched: [5:1.00]
@@ -2479,6 +2776,12 @@ define <2 x i64> @test_movq_reg(<2 x i64> %a0, <2 x i64> %a1) {
 ; HASWELL-NEXT:    vpaddq %xmm0, %xmm1, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movq_reg:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.33]
+; SKYLAKE-NEXT:    vpaddq %xmm0, %xmm1, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movq_reg:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.50]
@@ -2531,6 +2834,13 @@ define void @test_movsd_mem(double* %a0, double* %a1) {
 ; HASWELL-NEXT:    vmovsd %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movsd_mem:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddsd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovsd %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movsd_mem:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero sched: [5:1.00]
@@ -2583,6 +2893,11 @@ define <2 x double> @test_movsd_reg(<2 x double> %a0, <2 x double> %a1) {
 ; HASWELL-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm1[0],xmm0[0] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movsd_reg:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm1[0],xmm0[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movsd_reg:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm1[0],xmm0[0] sched: [1:0.50]
@@ -2632,6 +2947,13 @@ define void @test_movupd(<2 x double> *%a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vmovupd %xmm0, (%rsi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movupd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovupd (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vmovupd %xmm0, (%rsi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movupd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovupd (%rdi), %xmm0 # sched: [5:1.00]
@@ -2682,6 +3004,12 @@ define <2 x double> @test_mulpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulpd (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulpd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -2730,6 +3058,12 @@ define double @test_mulsd(double %a0, double %a1, double *%a2) {
 ; HASWELL-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    vmulsd (%rdi), %xmm0, %xmm0 # sched: [5:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mulsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmulsd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -2783,6 +3117,13 @@ define <2 x double> @test_orpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_orpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vorpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vorpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_orpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vorpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2842,6 +3183,12 @@ define <8 x i16> @test_packssdw(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpackssdw (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_packssdw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpackssdw (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_packssdw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2896,6 +3243,12 @@ define <16 x i8> @test_packsswb(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpacksswb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_packsswb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpacksswb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_packsswb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -2950,6 +3303,12 @@ define <16 x i8> @test_packuswb(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpackuswb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_packuswb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpackuswb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_packuswb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3004,6 +3363,12 @@ define <16 x i8> @test_paddb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpaddb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3056,6 +3421,12 @@ define <4 x i32> @test_paddd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpaddd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3104,6 +3475,12 @@ define <2 x i64> @test_paddq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3156,6 +3533,12 @@ define <16 x i8> @test_paddsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpaddsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3209,6 +3592,12 @@ define <8 x i16> @test_paddsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpaddsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3262,6 +3651,12 @@ define <16 x i8> @test_paddusb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpaddusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddusb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddusb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3315,6 +3710,12 @@ define <8 x i16> @test_paddusw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpaddusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddusw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddusw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3368,6 +3769,12 @@ define <8 x i16> @test_paddw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpaddw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_paddw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_paddw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3421,6 +3828,13 @@ define <2 x i64> @test_pand(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pand:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpand %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpand (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pand:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpand %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3483,6 +3897,13 @@ define <2 x i64> @test_pandn(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pandn:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpandn %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpandn (%rdi), %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pandn:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpandn %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3540,6 +3961,12 @@ define <16 x i8> @test_pavgb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpavgb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pavgb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpavgb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpavgb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pavgb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpavgb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3593,6 +4020,12 @@ define <8 x i16> @test_pavgw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpavgw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pavgw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpavgw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpavgw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pavgw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpavgw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -3649,6 +4082,13 @@ define <16 x i8> @test_pcmpeqb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpeqb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpeqb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpeqb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3708,6 +4148,13 @@ define <4 x i32> @test_pcmpeqd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpeqd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpeqd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3767,6 +4214,13 @@ define <8 x i16> @test_pcmpeqw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpeqw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpeqw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpeqw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3827,6 +4281,13 @@ define <16 x i8> @test_pcmpgtb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpgtb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpgtb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpgtb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpgtb %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3887,6 +4348,13 @@ define <4 x i32> @test_pcmpgtd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpgtd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpeqd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpgtd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3947,6 +4415,13 @@ define <8 x i16> @test_pcmpgtw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpgtw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpgtw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm0, %xmm1, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpgtw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpgtw %xmm1, %xmm0, %xmm1 # sched: [1:0.50]
@@ -3999,6 +4474,12 @@ define i16 @test_pextrw(<8 x i16> %a0) {
 ; HASWELL-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pextrw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpextrw $6, %xmm0, %eax # sched: [2:1.00]
+; SKYLAKE-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pextrw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpextrw $6, %xmm0, %eax # sched: [1:0.50]
@@ -4049,6 +4530,12 @@ define <8 x i16> @test_pinsrw(<8 x i16> %a0, i16 %a1, i16 *%a2) {
 ; HASWELL-NEXT:    vpinsrw $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pinsrw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpinsrw $1, %edi, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vpinsrw $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pinsrw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpinsrw $1, %edi, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4105,6 +4592,12 @@ define <4 x i32> @test_pmaddwd(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmaddwd (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaddwd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaddwd %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmaddwd (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaddwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaddwd %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4159,6 +4652,12 @@ define <8 x i16> @test_pmaxsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmaxsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4212,6 +4711,12 @@ define <16 x i8> @test_pmaxub(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpmaxub (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxub:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxub %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxub (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxub:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxub %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4265,6 +4770,12 @@ define <8 x i16> @test_pminsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpminsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4318,6 +4829,12 @@ define <16 x i8> @test_pminub(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpminub (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminub:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminub %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminub (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminub:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminub %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4364,6 +4881,11 @@ define i32 @test_pmovmskb(<16 x i8> %a0) {
 ; HASWELL-NEXT:    vpmovmskb %xmm0, %eax # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovmskb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovmskb %xmm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovmskb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovmskb %xmm0, %eax # sched: [1:0.50]
@@ -4409,6 +4931,12 @@ define <8 x i16> @test_pmulhuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmulhuw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmulhuw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmulhuw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmulhuw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmulhuw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmulhuw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4458,6 +4986,12 @@ define <8 x i16> @test_pmulhw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmulhw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmulhw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmulhw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmulhw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmulhw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmulhw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4507,6 +5041,12 @@ define <8 x i16> @test_pmullw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmullw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmullw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmullw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmullw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmullw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmullw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4563,6 +5103,12 @@ define <2 x i64> @test_pmuludq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmuludq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmuludq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmuludq %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmuludq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmuludq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmuludq %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4618,6 +5164,13 @@ define <2 x i64> @test_por(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_por:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpor (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_por:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4677,6 +5230,12 @@ define <2 x i64> @test_psadbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsadbw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psadbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpsadbw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psadbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsadbw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -4734,6 +5293,13 @@ define <4 x i32> @test_pshufd(<4 x i32> %a0, <4 x i32> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pshufd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,3,2] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpshufd {{.*#+}} xmm1 = mem[3,2,1,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pshufd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpshufd {{.*#+}} xmm1 = mem[3,2,1,0] sched: [6:1.00]
@@ -4792,6 +5358,13 @@ define <8 x i16> @test_pshufhw(<8 x i16> %a0, <8 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pshufhw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpshufhw {{.*#+}} xmm1 = mem[0,1,2,3,7,6,5,4] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pshufhw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpshufhw {{.*#+}} xmm1 = mem[0,1,2,3,7,6,5,4] sched: [6:1.00]
@@ -4850,6 +5423,13 @@ define <8 x i16> @test_pshuflw(<8 x i16> %a0, <8 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pshuflw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpshuflw {{.*#+}} xmm1 = mem[3,2,1,0,4,5,6,7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pshuflw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpshuflw {{.*#+}} xmm1 = mem[3,2,1,0,4,5,6,7] sched: [6:1.00]
@@ -4906,6 +5486,13 @@ define <4 x i32> @test_pslld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpslld $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pslld:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpslld %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpslld (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpslld $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pslld:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpslld %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -4960,6 +5547,11 @@ define <4 x i32> @test_pslldq(<4 x i32> %a0) {
 ; HASWELL-NEXT:    vpslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pslldq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pslldq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11] sched: [1:0.50]
@@ -5009,6 +5601,13 @@ define <2 x i64> @test_psllq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpsllq $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psllq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsllq %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsllq (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsllq $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psllq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsllq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5067,6 +5666,13 @@ define <8 x i16> @test_psllw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsllw $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psllw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsllw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsllw (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsllw $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psllw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsllw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5125,6 +5731,13 @@ define <4 x i32> @test_psrad(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpsrad $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psrad:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsrad %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrad (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrad $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psrad:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsrad %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5183,6 +5796,13 @@ define <8 x i16> @test_psraw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsraw $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psraw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsraw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsraw (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsraw $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psraw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsraw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5241,6 +5861,13 @@ define <4 x i32> @test_psrld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpsrld $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psrld:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsrld %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrld (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrld $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psrld:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsrld %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5295,6 +5922,11 @@ define <4 x i32> @test_psrldq(<4 x i32> %a0) {
 ; HASWELL-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psrldq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psrldq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero sched: [1:0.50]
@@ -5344,6 +5976,13 @@ define <2 x i64> @test_psrlq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpsrlq $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psrlq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsrlq %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrlq (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrlq $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psrlq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsrlq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5402,6 +6041,13 @@ define <8 x i16> @test_psrlw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsrlw $2, %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psrlw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsrlw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrlw (%rdi), %xmm0, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpsrlw $2, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psrlw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsrlw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5459,6 +6105,12 @@ define <16 x i8> @test_psubb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsubb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5511,6 +6163,12 @@ define <4 x i32> @test_psubd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpsubd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5559,6 +6217,12 @@ define <2 x i64> @test_psubq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpsubq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5611,6 +6275,12 @@ define <16 x i8> @test_psubsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsubsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5664,6 +6334,12 @@ define <8 x i16> @test_psubsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsubsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubsw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5717,6 +6393,12 @@ define <16 x i8> @test_psubusb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsubusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubusb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubusb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubusb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubusb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5770,6 +6452,12 @@ define <8 x i16> @test_psubusw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsubusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubusw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubusw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubusw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubusw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5823,6 +6511,12 @@ define <8 x i16> @test_psubw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsubw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psubw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsubw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsubw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psubw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsubw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -5875,6 +6569,12 @@ define <16 x i8> @test_punpckhbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpckhbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpckhbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] sched: [1:0.50]
@@ -5930,6 +6630,13 @@ define <4 x i32> @test_punpckhdq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpckhdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpckhdq {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpckhdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50]
@@ -5986,6 +6693,13 @@ define <2 x i64> @test_punpckhqdq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2)
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpckhqdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpckhqdq {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpckhqdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:0.50]
@@ -6041,6 +6755,12 @@ define <8 x i16> @test_punpckhwd(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpckhwd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpckhwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:0.50]
@@ -6093,6 +6813,12 @@ define <16 x i8> @test_punpcklbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpcklbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpcklbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:0.50]
@@ -6148,6 +6874,13 @@ define <4 x i32> @test_punpckldq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpckldq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpckldq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:0.50]
@@ -6204,6 +6937,13 @@ define <2 x i64> @test_punpcklqdq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2)
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpcklqdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpcklqdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50]
@@ -6259,6 +6999,12 @@ define <8 x i16> @test_punpcklwd(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_punpcklwd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_punpcklwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:0.50]
@@ -6312,6 +7058,13 @@ define <2 x i64> @test_pxor(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pxor:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpxor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    vpxor (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pxor:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpxor %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -6368,6 +7121,13 @@ define <2 x double> @test_shufpd(<2 x double> %a0, <2 x double> %a1, <2 x double
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shufpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vshufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vshufpd {{.*#+}} xmm1 = xmm1[1],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_shufpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vshufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0] sched: [1:0.50]
@@ -6425,6 +7185,13 @@ define <2 x double> @test_sqrtpd(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtpd %xmm0, %xmm0 # sched: [21:1.00]
+; SKYLAKE-NEXT:    vsqrtpd (%rdi), %xmm1 # sched: [21:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsqrtpd (%rdi), %xmm1 # sched: [26:21.00]
@@ -6489,6 +7256,14 @@ define <2 x double> @test_sqrtsd(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sqrtsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsqrtsd %xmm0, %xmm0, %xmm0 # sched: [21:1.00]
+; SKYLAKE-NEXT:    vmovapd (%rdi), %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vsqrtsd %xmm1, %xmm1, %xmm1 # sched: [21:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_sqrtsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovapd (%rdi), %xmm1 # sched: [5:1.00]
@@ -6543,6 +7318,12 @@ define <2 x double> @test_subpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -6591,6 +7372,12 @@ define double @test_subsd(double %a0, double %a1, double *%a2) {
 ; HASWELL-NEXT:    vsubsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_subsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vsubsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vsubsd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_subsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vsubsd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -6679,6 +7466,20 @@ define i32 @test_ucomisd(<2 x double> %a0, <2 x double> %a1, <2 x double> *%a2)
 ; HASWELL-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_ucomisd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vucomisd %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %cl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %cl # sched: [1:0.25]
+; SKYLAKE-NEXT:    vucomisd (%rdi), %xmm0 # sched: [7:1.00]
+; SKYLAKE-NEXT:    setnp %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    sete %dl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    orb %cl, %dl # sched: [1:0.25]
+; SKYLAKE-NEXT:    movzbl %dl, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_ucomisd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vucomisd %xmm1, %xmm0 # sched: [3:1.00]
@@ -6750,6 +7551,13 @@ define <2 x double> @test_unpckhpd(<2 x double> %a0, <2 x double> %a1, <2 x doub
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpckhpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpckhpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:0.50]
@@ -6812,6 +7620,13 @@ define <2 x double> @test_unpcklpd(<2 x double> %a0, <2 x double> %a1, <2 x doub
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_unpcklpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vunpcklpd {{.*#+}} xmm1 = xmm0[0],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_unpcklpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:0.50]
@@ -6868,6 +7683,13 @@ define <2 x double> @test_xorpd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_xorpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vxorpd %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vxorpd (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_xorpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vxorpd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]

diff --git a/test/CodeGen/X86/sse3-schedule.ll b/test/CodeGen/X86/sse3-schedule.ll
index ee15f0e7df6feef8d91b5cb4c62c788af6d74c64..bb70599c2803ac08f2d2f94227c86278b198d131 100644 (file)
--- a/test/CodeGen/X86/sse3-schedule.ll
+++ b/test/CodeGen/X86/sse3-schedule.ll
@@ -5,7 +5,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -40,6 +40,12 @@ define <2 x double> @test_addsubpd(<2 x double> %a0, <2 x double> %a1, <2 x doub
 ; HASWELL-NEXT:    vaddsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addsubpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddsubpd (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addsubpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -89,6 +95,12 @@ define <4 x float> @test_addsubps(<4 x float> %a0, <4 x float> %a1, <4 x float>
 ; HASWELL-NEXT:    vaddsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_addsubps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vaddsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vaddsubps (%rdi), %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_addsubps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vaddsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -138,6 +150,12 @@ define <2 x double> @test_haddpd(<2 x double> %a0, <2 x double> %a1, <2 x double
 ; HASWELL-NEXT:    vhaddpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_haddpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhaddpd %xmm1, %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhaddpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_haddpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -187,6 +205,12 @@ define <4 x float> @test_haddps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%
 ; HASWELL-NEXT:    vhaddps (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_haddps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhaddps %xmm1, %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhaddps (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_haddps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -236,6 +260,12 @@ define <2 x double> @test_hsubpd(<2 x double> %a0, <2 x double> %a1, <2 x double
 ; HASWELL-NEXT:    vhsubpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_hsubpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhsubpd %xmm1, %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhsubpd (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_hsubpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhsubpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -285,6 +315,12 @@ define <4 x float> @test_hsubps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%
 ; HASWELL-NEXT:    vhsubps (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_hsubps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vhsubps %xmm1, %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    vhsubps (%rdi), %xmm0, %xmm0 # sched: [5:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_hsubps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vhsubps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -331,6 +367,11 @@ define <16 x i8> @test_lddqu(i8* %a0) {
 ; HASWELL-NEXT:    vlddqu (%rdi), %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_lddqu:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vlddqu (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_lddqu:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vlddqu (%rdi), %xmm0 # sched: [5:1.00]
@@ -381,6 +422,13 @@ define void @test_monitor(i8* %a0, i32 %a1, i32 %a2) {
 ; HASWELL-NEXT:    monitor # sched: [100:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_monitor:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    leaq (%rdi), %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    movl %esi, %ecx # sched: [1:0.25]
+; SKYLAKE-NEXT:    monitor # sched: [100:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_monitor:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    leaq (%rdi), %rax # sched: [1:0.50]
@@ -436,6 +484,13 @@ define <2 x double> @test_movddup(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movddup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movddup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovddup {{.*#+}} xmm1 = mem[0,0] sched: [5:1.00]
@@ -493,6 +548,13 @@ define <4 x float> @test_movshdup(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movshdup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovshdup {{.*#+}} xmm1 = mem[1,1,3,3] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movshdup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovshdup {{.*#+}} xmm1 = mem[1,1,3,3] sched: [5:1.00]
@@ -550,6 +612,13 @@ define <4 x float> @test_movsldup(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movsldup:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2] sched: [1:1.00]
+; SKYLAKE-NEXT:    vmovsldup {{.*#+}} xmm1 = mem[0,0,2,2] sched: [1:0.50]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movsldup:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovsldup {{.*#+}} xmm1 = mem[0,0,2,2] sched: [5:1.00]
@@ -606,6 +675,13 @@ define void @test_mwait(i32 %a0, i32 %a1) {
 ; HASWELL-NEXT:    mwait # sched: [20:2.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mwait:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    movl %edi, %ecx # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl %esi, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    mwait # sched: [20:2.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mwait:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    movl %edi, %ecx # sched: [1:0.17]

diff --git a/test/CodeGen/X86/sse41-schedule.ll b/test/CodeGen/X86/sse41-schedule.ll
index 117fae1937f44fdae892a628c5604ebc10344ff1..88e7d69eab54f05193c20d2b255a9da60c25ebc0 100644 (file)
--- a/test/CodeGen/X86/sse41-schedule.ll
+++ b/test/CodeGen/X86/sse41-schedule.ll
@@ -4,7 +4,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -37,6 +37,13 @@ define <2 x double> @test_blendpd(<2 x double> %a0, <2 x double> %a1, <2 x doubl
 ; HASWELL-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],mem[1] sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1] sched: [1:0.33]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],mem[1] sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1] sched: [1:0.50]
@@ -82,6 +89,12 @@ define <4 x float> @test_blendps(<4 x float> %a0, <4 x float> %a1, <4 x float> *
 ; HASWELL-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2,3] sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2],xmm0[3] sched: [1:0.33]
+; SKYLAKE-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2,3] sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2],xmm0[3] sched: [1:0.50]
@@ -130,6 +143,12 @@ define <2 x double> @test_blendvpd(<2 x double> %a0, <2 x double> %a1, <2 x doub
 ; HASWELL-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendvpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vblendvpd %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendvpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendvpd %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -179,6 +198,12 @@ define <4 x float> @test_blendvps(<4 x float> %a0, <4 x float> %a1, <4 x float>
 ; HASWELL-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_blendvps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vblendvps %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_blendvps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -222,6 +247,12 @@ define <2 x double> @test_dppd(<2 x double> %a0, <2 x double> %a1, <2 x double>
 ; HASWELL-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [9:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_dppd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [9:1.00]
+; SKYLAKE-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [9:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_dppd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -265,6 +296,12 @@ define <4 x float> @test_dpps(<4 x float> %a0, <4 x float> %a1, <4 x float> *%a2
 ; HASWELL-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [14:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_dpps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [14:2.00]
+; SKYLAKE-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [14:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_dpps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
@@ -308,6 +345,12 @@ define <4 x float> @test_insertps(<4 x float> %a0, <4 x float> %a1, float *%a2)
 ; HASWELL-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_insertps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vinsertps {{.*#+}} xmm0 = zero,xmm1[0],xmm0[2,3] sched: [1:1.00]
+; SKYLAKE-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_insertps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vinsertps {{.*#+}} xmm0 = zero,xmm1[0],xmm0[2,3] sched: [1:0.50]
@@ -347,6 +390,11 @@ define <2 x i64> @test_movntdqa(i8* %a0) {
 ; HASWELL-NEXT:    vmovntdqa (%rdi), %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_movntdqa:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmovntdqa (%rdi), %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_movntdqa:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmovntdqa (%rdi), %xmm0 # sched: [5:1.00]
@@ -386,6 +434,12 @@ define <8 x i16> @test_mpsadbw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vmpsadbw $7, (%rdi), %xmm0, %xmm0 # sched: [7:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mpsadbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vmpsadbw $7, %xmm1, %xmm0, %xmm0 # sched: [7:2.00]
+; SKYLAKE-NEXT:    vmpsadbw $7, (%rdi), %xmm0, %xmm0 # sched: [7:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_mpsadbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vmpsadbw $7, %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
@@ -430,6 +484,12 @@ define <8 x i16> @test_packusdw(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpackusdw (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_packusdw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpackusdw %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpackusdw (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_packusdw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpackusdw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -480,6 +540,12 @@ define <16 x i8> @test_pblendvb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> %a2, <16
 ; HASWELL-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pblendvb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vpblendvb %xmm2, (%rdi), %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pblendvb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -523,6 +589,12 @@ define <8 x i16> @test_pblendw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5,6],mem[7] sched: [4:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pblendw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],mem[2,3],xmm0[4,5,6],mem[7] sched: [4:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pblendw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7] sched: [1:0.50]
@@ -565,6 +637,12 @@ define <2 x i64> @test_pcmpeqq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpeqq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpcmpeqq (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpeqq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpeqq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -609,6 +687,12 @@ define i32 @test_pextrb(<16 x i8> %a0, i8 *%a1) {
 ; HASWELL-NEXT:    vpextrb $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pextrb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpextrb $3, %xmm0, %eax # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpextrb $1, %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pextrb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpextrb $3, %xmm0, %eax # sched: [1:0.50]
@@ -652,6 +736,12 @@ define i32 @test_pextrd(<4 x i32> %a0, i32 *%a1) {
 ; HASWELL-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pextrd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpextrd $3, %xmm0, %eax # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pextrd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpextrd $3, %xmm0, %eax # sched: [1:0.50]
@@ -694,6 +784,12 @@ define i64 @test_pextrq(<2 x i64> %a0, <2 x i64> %a1, i64 *%a2) {
 ; HASWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pextrq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpextrq $1, %xmm0, %rax # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pextrq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpextrq $1, %xmm0, %rax # sched: [1:0.50]
@@ -736,6 +832,12 @@ define i32 @test_pextrw(<8 x i16> %a0, i16 *%a1) {
 ; HASWELL-NEXT:    vpextrw $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pextrw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpextrw $3, %xmm0, %eax # sched: [2:1.00]
+; SKYLAKE-NEXT:    vpextrw $1, %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pextrw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpextrw $3, %xmm0, %eax # sched: [1:0.50]
@@ -779,6 +881,12 @@ define <8 x i16> @test_phminposuw(<8 x i16> *%a0) {
 ; HASWELL-NEXT:    vphminposuw %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phminposuw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphminposuw (%rdi), %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vphminposuw %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phminposuw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphminposuw (%rdi), %xmm0 # sched: [7:1.00]
@@ -822,6 +930,12 @@ define <16 x i8> @test_pinsrb(<16 x i8> %a0, i8 %a1, i8 *%a2) {
 ; HASWELL-NEXT:    vpinsrb $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pinsrb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpinsrb $1, %edi, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vpinsrb $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pinsrb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpinsrb $1, %edi, %xmm0, %xmm0 # sched: [1:0.50]
@@ -864,6 +978,12 @@ define <4 x i32> @test_pinsrd(<4 x i32> %a0, i32 %a1, i32 *%a2) {
 ; HASWELL-NEXT:    vpinsrd $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pinsrd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vpinsrd $3, (%rsi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pinsrd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm0 # sched: [1:0.50]
@@ -910,6 +1030,13 @@ define <2 x i64> @test_pinsrq(<2 x i64> %a0, <2 x i64> %a1, i64 %a2, i64 *%a3) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pinsrq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpinsrq $1, %rdi, %xmm0, %xmm0 # sched: [2:2.00]
+; SKYLAKE-NEXT:    vpinsrq $1, (%rsi), %xmm1, %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pinsrq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpinsrq $1, (%rsi), %xmm1, %xmm1 # sched: [6:1.00]
@@ -955,6 +1082,12 @@ define <16 x i8> @test_pmaxsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpmaxsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -998,6 +1131,12 @@ define <4 x i32> @test_pmaxsd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmaxsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1041,6 +1180,12 @@ define <4 x i32> @test_pmaxud(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmaxud (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxud:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxud (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxud:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxud %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1084,6 +1229,12 @@ define <8 x i16> @test_pmaxuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmaxuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaxuw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpmaxuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaxuw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaxuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1127,6 +1278,12 @@ define <16 x i8> @test_pminsb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpminsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminsb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminsb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1170,6 +1327,12 @@ define <4 x i32> @test_pminsd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpminsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminsd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminsd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1213,6 +1376,12 @@ define <4 x i32> @test_pminud(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpminud (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminud:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminud %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminud (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminud:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminud %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1256,6 +1425,12 @@ define <8 x i16> @test_pminuw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpminuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pminuw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpminuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpminuw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pminuw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpminuw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -1304,6 +1479,13 @@ define <8 x i16> @test_pmovsxbw(<16 x i8> %a0, <8 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxbw %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxbw (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxbw (%rdi), %xmm1 # sched: [6:1.00]
@@ -1355,6 +1537,13 @@ define <4 x i32> @test_pmovsxbd(<16 x i8> %a0, <4 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxbd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxbd %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxbd (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxbd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxbd (%rdi), %xmm1 # sched: [6:1.00]
@@ -1406,6 +1595,13 @@ define <2 x i64> @test_pmovsxbq(<16 x i8> %a0, <2 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxbq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxbq %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxbq (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxbq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxbq (%rdi), %xmm1 # sched: [6:1.00]
@@ -1457,6 +1653,13 @@ define <2 x i64> @test_pmovsxdq(<4 x i32> %a0, <2 x i32> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxdq %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxdq (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxdq (%rdi), %xmm1 # sched: [6:1.00]
@@ -1508,6 +1711,13 @@ define <4 x i32> @test_pmovsxwd(<8 x i16> %a0, <4 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxwd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxwd %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxwd (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxwd (%rdi), %xmm1 # sched: [6:1.00]
@@ -1559,6 +1769,13 @@ define <2 x i64> @test_pmovsxwq(<8 x i16> %a0, <2 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovsxwq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovsxwq %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovsxwq (%rdi), %xmm1 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovsxwq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovsxwq (%rdi), %xmm1 # sched: [6:1.00]
@@ -1610,6 +1827,13 @@ define <8 x i16> @test_pmovzxbw(<16 x i8> %a0, <8 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxbw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxbw {{.*#+}} xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxbw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxbw {{.*#+}} xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero sched: [6:1.00]
@@ -1661,6 +1885,13 @@ define <4 x i32> @test_pmovzxbd(<16 x i8> %a0, <4 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxbd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxbd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero sched: [6:1.00]
@@ -1712,6 +1943,13 @@ define <2 x i64> @test_pmovzxbq(<16 x i8> %a0, <2 x i8> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxbq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxbq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxbq {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxbq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxbq {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero sched: [6:1.00]
@@ -1763,6 +2001,13 @@ define <2 x i64> @test_pmovzxdq(<4 x i32> %a0, <2 x i32> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxdq {{.*#+}} xmm1 = mem[0],zero,mem[1],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxdq {{.*#+}} xmm1 = mem[0],zero,mem[1],zero sched: [6:1.00]
@@ -1814,6 +2059,13 @@ define <4 x i32> @test_pmovzxwd(<8 x i16> %a0, <4 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxwd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxwd {{.*#+}} xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxwd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxwd {{.*#+}} xmm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero sched: [6:1.00]
@@ -1865,6 +2117,13 @@ define <2 x i64> @test_pmovzxwq(<8 x i16> %a0, <2 x i16> *%a1) {
 ; HASWELL-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmovzxwq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpmovzxwq {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero sched: [1:1.00]
+; SKYLAKE-NEXT:    vpaddq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmovzxwq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmovzxwq {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero sched: [6:1.00]
@@ -1911,6 +2170,12 @@ define <2 x i64> @test_pmuldq(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmuldq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmuldq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmuldq %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmuldq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmuldq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmuldq %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -1955,6 +2220,12 @@ define <4 x i32> @test_pmulld(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpmulld (%rdi), %xmm0, %xmm0 # sched: [10:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmulld:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmulld %xmm1, %xmm0, %xmm0 # sched: [10:2.00]
+; SKYLAKE-NEXT:    vpmulld (%rdi), %xmm0, %xmm0 # sched: [10:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmulld:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmulld %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -2013,6 +2284,16 @@ define i32 @test_ptest(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    movzbl %cl, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_ptest:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vptest %xmm1, %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    setb %al # sched: [1:0.50]
+; SKYLAKE-NEXT:    vptest (%rdi), %xmm0 # sched: [2:1.00]
+; SKYLAKE-NEXT:    setb %cl # sched: [1:0.50]
+; SKYLAKE-NEXT:    andb %al, %cl # sched: [1:0.25]
+; SKYLAKE-NEXT:    movzbl %cl, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_ptest:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vptest %xmm1, %xmm0 # sched: [1:0.50]
@@ -2070,6 +2351,13 @@ define <2 x double> @test_roundpd(<2 x double> %a0, <2 x double> *%a1) {
 ; HASWELL-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundpd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundpd $7, %xmm0, %xmm0 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundpd $7, (%rdi), %xmm1 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddpd %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundpd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundpd $7, (%rdi), %xmm1 # sched: [8:1.00]
@@ -2121,6 +2409,13 @@ define <4 x float> @test_roundps(<4 x float> %a0, <4 x float> *%a1) {
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundps:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundps $7, %xmm0, %xmm0 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundps $7, (%rdi), %xmm1 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundps:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundps $7, (%rdi), %xmm1 # sched: [8:1.00]
@@ -2173,6 +2468,13 @@ define <2 x double> @test_roundsd(<2 x double> %a0, <2 x double> %a1, <2 x doubl
 ; HASWELL-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundsd $7, %xmm1, %xmm0, %xmm1 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundsd $7, (%rdi), %xmm0, %xmm0 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundsd $7, %xmm1, %xmm0, %xmm1 # sched: [3:1.00]
@@ -2225,6 +2527,13 @@ define <4 x float> @test_roundss(<4 x float> %a0, <4 x float> %a1, <4 x float> *
 ; HASWELL-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_roundss:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vroundss $7, %xmm1, %xmm0, %xmm1 # sched: [5:1.25]
+; SKYLAKE-NEXT:    vroundss $7, (%rdi), %xmm0, %xmm0 # sched: [6:2.00]
+; SKYLAKE-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_roundss:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vroundss $7, %xmm1, %xmm0, %xmm1 # sched: [3:1.00]

diff --git a/test/CodeGen/X86/sse42-schedule.ll b/test/CodeGen/X86/sse42-schedule.ll
index d25be4b2abd148f7f07f92fb28509bb7650e6029..f02b041b8eb0938d7a1d44750f292bf29c1421d2 100644 (file)
--- a/test/CodeGen/X86/sse42-schedule.ll
+++ b/test/CodeGen/X86/sse42-schedule.ll
@@ -4,7 +4,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -37,6 +37,13 @@ define i32 @crc32_32_8(i32 %a0, i8 %a1, i8 *%a2) {
 ; HASWELL-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: crc32_32_8:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
+; SKYLAKE-NEXT:    crc32b (%rdx), %edi # sched: [7:1.00]
+; SKYLAKE-NEXT:    movl %edi, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: crc32_32_8:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
@@ -86,6 +93,13 @@ define i32 @crc32_32_16(i32 %a0, i16 %a1, i16 *%a2) {
 ; HASWELL-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: crc32_32_16:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    crc32w %si, %edi # sched: [3:1.00]
+; SKYLAKE-NEXT:    crc32w (%rdx), %edi # sched: [7:1.00]
+; SKYLAKE-NEXT:    movl %edi, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: crc32_32_16:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32w %si, %edi # sched: [3:1.00]
@@ -135,6 +149,13 @@ define i32 @crc32_32_32(i32 %a0, i32 %a1, i32 *%a2) {
 ; HASWELL-NEXT:    movl %edi, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: crc32_32_32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    crc32l %esi, %edi # sched: [3:1.00]
+; SKYLAKE-NEXT:    crc32l (%rdx), %edi # sched: [7:1.00]
+; SKYLAKE-NEXT:    movl %edi, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: crc32_32_32:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32l %esi, %edi # sched: [3:1.00]
@@ -184,6 +205,13 @@ define i64 @crc32_64_8(i64 %a0, i8 %a1, i8 *%a2) nounwind {
 ; HASWELL-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: crc32_64_8:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
+; SKYLAKE-NEXT:    crc32b (%rdx), %edi # sched: [7:1.00]
+; SKYLAKE-NEXT:    movq %rdi, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: crc32_64_8:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32b %sil, %edi # sched: [3:1.00]
@@ -233,6 +261,13 @@ define i64 @crc32_64_64(i64 %a0, i64 %a1, i64 *%a2) {
 ; HASWELL-NEXT:    movq %rdi, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: crc32_64_64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    crc32q %rsi, %rdi # sched: [3:1.00]
+; SKYLAKE-NEXT:    crc32q (%rdx), %rdi # sched: [7:1.00]
+; SKYLAKE-NEXT:    movq %rdi, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: crc32_64_64:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    crc32q %rsi, %rdi # sched: [3:1.00]
@@ -306,6 +341,19 @@ define i32 @test_pcmpestri(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    leal (%rcx,%rsi), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpestri:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    movl $7, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl $7, %edx # sched: [1:0.25]
+; SKYLAKE-NEXT:    vpcmpestri $7, %xmm1, %xmm0 # sched: [18:4.00]
+; SKYLAKE-NEXT:    movl %ecx, %esi # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl $7, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl $7, %edx # sched: [1:0.25]
+; SKYLAKE-NEXT:    vpcmpestri $7, (%rdi), %xmm0 # sched: [18:4.00]
+; SKYLAKE-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
+; SKYLAKE-NEXT:    leal (%rcx,%rsi), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpestri:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
@@ -380,6 +428,16 @@ define <16 x i8> @test_pcmpestrm(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpcmpestrm $7, (%rdi), %xmm0 # sched: [19:4.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpestrm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    movl $7, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl $7, %edx # sched: [1:0.25]
+; SKYLAKE-NEXT:    vpcmpestrm $7, %xmm1, %xmm0 # sched: [19:4.00]
+; SKYLAKE-NEXT:    movl $7, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    movl $7, %edx # sched: [1:0.25]
+; SKYLAKE-NEXT:    vpcmpestrm $7, (%rdi), %xmm0 # sched: [19:4.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpestrm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    movl $7, %eax # sched: [1:0.17]
@@ -443,6 +501,15 @@ define i32 @test_pcmpistri(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    leal (%rcx,%rax), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpistri:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpistri $7, %xmm1, %xmm0 # sched: [11:3.00]
+; SKYLAKE-NEXT:    movl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    vpcmpistri $7, (%rdi), %xmm0 # sched: [11:3.00]
+; SKYLAKE-NEXT:    # kill: %ECX<def> %ECX<kill> %RCX<def>
+; SKYLAKE-NEXT:    leal (%rcx,%rax), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpistri:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpistri $7, %xmm1, %xmm0 # sched: [6:1.00]
@@ -493,6 +560,12 @@ define <16 x i8> @test_pcmpistrm(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpcmpistrm $7, (%rdi), %xmm0 # sched: [11:3.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpistrm:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpistrm $7, %xmm1, %xmm0 # sched: [11:3.00]
+; SKYLAKE-NEXT:    vpcmpistrm $7, (%rdi), %xmm0 # sched: [11:3.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpistrm:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpistrm $7, %xmm1, %xmm0 # sched: [7:1.00]
@@ -536,6 +609,12 @@ define <2 x i64> @test_pcmpgtq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pcmpgtq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpcmpgtq (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pcmpgtq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -580,6 +659,12 @@ define <2 x i64> @test_pclmulqdq(<2 x i64> %a0, <2 x i64> %a1, <2 x i64> *%a2) {
 ; HASWELL-NEXT:    vpclmulqdq $0, (%rdi), %xmm0, %xmm0 # sched: [11:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pclmulqdq:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpclmulqdq $0, %xmm1, %xmm0, %xmm0 # sched: [11:2.00]
+; SKYLAKE-NEXT:    vpclmulqdq $0, (%rdi), %xmm0, %xmm0 # sched: [11:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pclmulqdq:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpclmulqdq $0, %xmm1, %xmm0, %xmm0 # sched: [2:1.00]

diff --git a/test/CodeGen/X86/ssse3-schedule.ll b/test/CodeGen/X86/ssse3-schedule.ll
index acbc2cf9917e975457e235dd44a864f95e29555c..1bacb90b24feba822356b371552dac8c8a7dc74c 100644 (file)
--- a/test/CodeGen/X86/ssse3-schedule.ll
+++ b/test/CodeGen/X86/ssse3-schedule.ll
@@ -5,7 +5,7 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=sandybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=ivybridge | FileCheck %s --check-prefix=CHECK --check-prefix=SANDY
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=btver2 | FileCheck %s --check-prefix=CHECK --check-prefix=BTVER2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
 
@@ -46,6 +46,13 @@ define <16 x i8> @test_pabsb(<16 x i8> %a0, <16 x i8> *%a1) {
 ; HASWELL-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsb %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsb (%rdi), %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pabsb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpabsb (%rdi), %xmm1 # sched: [6:1.00]
@@ -104,6 +111,13 @@ define <4 x i32> @test_pabsd(<4 x i32> %a0, <4 x i32> *%a1) {
 ; HASWELL-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsd %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsd (%rdi), %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pabsd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpabsd (%rdi), %xmm1 # sched: [6:1.00]
@@ -162,6 +176,13 @@ define <8 x i16> @test_pabsw(<8 x i16> %a0, <8 x i16> *%a1) {
 ; HASWELL-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pabsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpabsw %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpabsw (%rdi), %xmm1 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpor %xmm1, %xmm0, %xmm0 # sched: [1:0.33]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pabsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpabsw (%rdi), %xmm1 # sched: [6:1.00]
@@ -219,6 +240,12 @@ define <8 x i16> @test_palignr(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpalignr {{.*#+}} xmm0 = mem[14,15],xmm0[0,1,2,3,4,5,6,7,8,9,10,11,12,13] sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_palignr:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpalignr {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5] sched: [1:1.00]
+; SKYLAKE-NEXT:    vpalignr {{.*#+}} xmm0 = mem[14,15],xmm0[0,1,2,3,4,5,6,7,8,9,10,11,12,13] sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_palignr:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpalignr {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5] sched: [1:0.50]
@@ -267,6 +294,12 @@ define <4 x i32> @test_phaddd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vphaddd (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phaddd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphaddd %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphaddd (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phaddd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphaddd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -316,6 +349,12 @@ define <8 x i16> @test_phaddsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vphaddsw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phaddsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphaddsw %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphaddsw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phaddsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphaddsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -365,6 +404,12 @@ define <8 x i16> @test_phaddw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vphaddw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phaddw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphaddw %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphaddw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phaddw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphaddw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -414,6 +459,12 @@ define <4 x i32> @test_phsubd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vphsubd (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phsubd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphsubd %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphsubd (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phsubd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphsubd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -463,6 +514,12 @@ define <8 x i16> @test_phsubsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vphsubsw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phsubsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphsubsw %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphsubsw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phsubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphsubsw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -512,6 +569,12 @@ define <8 x i16> @test_phsubw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vphsubw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_phsubw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vphsubw %xmm1, %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    vphsubw (%rdi), %xmm0, %xmm0 # sched: [3:2.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_phsubw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vphsubw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -561,6 +624,12 @@ define <8 x i16> @test_pmaddubsw(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpmaddubsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmaddubsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmaddubsw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmaddubsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmaddubsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmaddubsw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -611,6 +680,12 @@ define <8 x i16> @test_pmulhrsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpmulhrsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pmulhrsw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpmulhrsw %xmm1, %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    vpmulhrsw (%rdi), %xmm0, %xmm0 # sched: [5:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pmulhrsw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpmulhrsw %xmm1, %xmm0, %xmm0 # sched: [2:1.00]
@@ -660,6 +735,12 @@ define <16 x i8> @test_pshufb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpshufb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pshufb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpshufb %xmm1, %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    vpshufb (%rdi), %xmm0, %xmm0 # sched: [1:1.00]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_pshufb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpshufb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -713,6 +794,12 @@ define <16 x i8> @test_psignb(<16 x i8> %a0, <16 x i8> %a1, <16 x i8> *%a2) {
 ; HASWELL-NEXT:    vpsignb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psignb:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsignb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsignb (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psignb:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsignb %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -766,6 +853,12 @@ define <4 x i32> @test_psignd(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> *%a2) {
 ; HASWELL-NEXT:    vpsignd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psignd:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsignd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsignd (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psignd:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsignd %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
@@ -819,6 +912,12 @@ define <8 x i16> @test_psignw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; HASWELL-NEXT:    vpsignw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_psignw:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    vpsignw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    vpsignw (%rdi), %xmm0, %xmm0 # sched: [1:0.50]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
 ; BTVER2-LABEL: test_psignw:
 ; BTVER2:       # BB#0:
 ; BTVER2-NEXT:    vpsignw %xmm1, %xmm0, %xmm0 # sched: [1:0.50]