[LegalizeTypes][X86] When splitting a vselect for type legalization, don't split...

Summary: The VSELECT splitting code tries to split a setcc input as well. But on avx512 where mask registers are well supported it should be better to just split the mask and use a single compare.

Reviewers: RKSimon, spatel, efriedma

Reviewed By: spatel

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68359

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@373863 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp b/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp
index 560b5729e3def001c3c4810b4703fc1ea996dd8c..5562f400b6e1da954dbc170901bcd589c02a7009 100644 (file)
--- a/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp
@@ -521,9 +521,18 @@ void DAGTypeLegalizer::SplitRes_SELECT(SDNode *N, SDValue &Lo, SDValue &Hi) {
       GetSplitVector(Cond, CL, CH);
     // It seems to improve code to generate two narrow SETCCs as opposed to
     // splitting a wide result vector.
-    else if (Cond.getOpcode() == ISD::SETCC)
-      SplitVecRes_SETCC(Cond.getNode(), CL, CH);
-    else
+    else if (Cond.getOpcode() == ISD::SETCC) {
+      // If the condition is a vXi1 vector, and the LHS of the setcc is a legal
+      // type and the setcc result type is the same vXi1, then leave the setcc
+      // alone.
+      EVT CondLHSVT = Cond.getOperand(0).getValueType();
+      if (Cond.getValueType().getVectorElementType() == MVT::i1 &&
+          isTypeLegal(CondLHSVT) &&
+          getSetCCResultType(CondLHSVT) == Cond.getValueType())
+        std::tie(CL, CH) = DAG.SplitVector(Cond, dl);
+      else
+        SplitVecRes_SETCC(Cond.getNode(), CL, CH);
+    } else
       std::tie(CL, CH) = DAG.SplitVector(Cond, dl);
   }
 

diff --git a/test/CodeGen/X86/avx512-vselect.ll b/test/CodeGen/X86/avx512-vselect.ll
index 7ee4e6674e0e754f64f879da724c9e838219517b..07e5aeac015bc8835ff5e9f80cf3b706de738ce3 100644 (file)
--- a/test/CodeGen/X86/avx512-vselect.ll
+++ b/test/CodeGen/X86/avx512-vselect.ll
@@ -51,10 +51,9 @@ entry:
 define <16 x i64> @test3(<16 x i8> %x, <16 x i64> %a, <16 x i64> %b) {
 ; CHECK-SKX-LABEL: test3:
 ; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vpshufd {{.*#+}} xmm5 = xmm0[2,3,0,1]
-; CHECK-SKX-NEXT:    vptestnmb %xmm5, %xmm5, %k1
-; CHECK-SKX-NEXT:    vptestnmb %xmm0, %xmm0, %k2
-; CHECK-SKX-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k2}
+; CHECK-SKX-NEXT:    vptestnmb %xmm0, %xmm0, %k1
+; CHECK-SKX-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-SKX-NEXT:    kshiftrw $8, %k1, %k1
 ; CHECK-SKX-NEXT:    vpblendmq %zmm2, %zmm4, %zmm1 {%k1}
 ; CHECK-SKX-NEXT:    retq
 ;
@@ -76,10 +75,9 @@ define <16 x i64> @test3(<16 x i8> %x, <16 x i64> %a, <16 x i64> %b) {
 define <16 x i64> @test4(<16 x i16> %x, <16 x i64> %a, <16 x i64> %b) {
 ; CHECK-SKX-LABEL: test4:
 ; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vextracti128 $1, %ymm0, %xmm5
-; CHECK-SKX-NEXT:    vptestnmw %xmm5, %xmm5, %k1
-; CHECK-SKX-NEXT:    vptestnmw %xmm0, %xmm0, %k2
-; CHECK-SKX-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k2}
+; CHECK-SKX-NEXT:    vptestnmw %ymm0, %ymm0, %k1
+; CHECK-SKX-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-SKX-NEXT:    kshiftrw $8, %k1, %k1
 ; CHECK-SKX-NEXT:    vpblendmq %zmm2, %zmm4, %zmm1 {%k1}
 ; CHECK-SKX-NEXT:    retq
 ;
@@ -99,23 +97,13 @@ define <16 x i64> @test4(<16 x i16> %x, <16 x i64> %a, <16 x i64> %b) {
 }
 
 define <16 x i64> @test5(<16 x i32> %x, <16 x i64> %a, <16 x i64> %b) {
-; CHECK-SKX-LABEL: test5:
-; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vextracti64x4 $1, %zmm0, %ymm5
-; CHECK-SKX-NEXT:    vptestnmd %ymm5, %ymm5, %k1
-; CHECK-SKX-NEXT:    vptestnmd %ymm0, %ymm0, %k2
-; CHECK-SKX-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k2}
-; CHECK-SKX-NEXT:    vpblendmq %zmm2, %zmm4, %zmm1 {%k1}
-; CHECK-SKX-NEXT:    retq
-;
-; CHECK-KNL-LABEL: test5:
-; CHECK-KNL:       # %bb.0:
-; CHECK-KNL-NEXT:    vextracti64x4 $1, %zmm0, %ymm5
-; CHECK-KNL-NEXT:    vptestnmd %zmm5, %zmm5, %k1
-; CHECK-KNL-NEXT:    vptestnmd %zmm0, %zmm0, %k2
-; CHECK-KNL-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k2}
-; CHECK-KNL-NEXT:    vpblendmq %zmm2, %zmm4, %zmm1 {%k1}
-; CHECK-KNL-NEXT:    retq
+; CHECK-LABEL: test5:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vptestnmd %zmm0, %zmm0, %k1
+; CHECK-NEXT:    vpblendmq %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-NEXT:    kshiftrw $8, %k1, %k1
+; CHECK-NEXT:    vpblendmq %zmm2, %zmm4, %zmm1 {%k1}
+; CHECK-NEXT:    retq
   %c = icmp eq <16 x i32> %x, zeroinitializer
   %ret = select <16 x i1> %c, <16 x i64> %a, <16 x i64> %b
   ret <16 x i64> %ret
@@ -124,10 +112,9 @@ define <16 x i64> @test5(<16 x i32> %x, <16 x i64> %a, <16 x i64> %b) {
 define <32 x i32> @test6(<32 x i8> %x, <32 x i32> %a, <32 x i32> %b) {
 ; CHECK-SKX-LABEL: test6:
 ; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vextracti128 $1, %ymm0, %xmm5
-; CHECK-SKX-NEXT:    vptestnmb %xmm5, %xmm5, %k1
-; CHECK-SKX-NEXT:    vptestnmb %xmm0, %xmm0, %k2
-; CHECK-SKX-NEXT:    vpblendmd %zmm1, %zmm3, %zmm0 {%k2}
+; CHECK-SKX-NEXT:    vptestnmb %ymm0, %ymm0, %k1
+; CHECK-SKX-NEXT:    vpblendmd %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-SKX-NEXT:    kshiftrd $16, %k1, %k1
 ; CHECK-SKX-NEXT:    vpblendmd %zmm2, %zmm4, %zmm1 {%k1}
 ; CHECK-SKX-NEXT:    retq
 ;
@@ -151,10 +138,9 @@ define <32 x i32> @test6(<32 x i8> %x, <32 x i32> %a, <32 x i32> %b) {
 define <32 x i32> @test7(<32 x i16> %x, <32 x i32> %a, <32 x i32> %b) {
 ; CHECK-SKX-LABEL: test7:
 ; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vextracti64x4 $1, %zmm0, %ymm5
-; CHECK-SKX-NEXT:    vptestnmw %ymm5, %ymm5, %k1
-; CHECK-SKX-NEXT:    vptestnmw %ymm0, %ymm0, %k2
-; CHECK-SKX-NEXT:    vpblendmd %zmm1, %zmm3, %zmm0 {%k2}
+; CHECK-SKX-NEXT:    vptestnmw %zmm0, %zmm0, %k1
+; CHECK-SKX-NEXT:    vpblendmd %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-SKX-NEXT:    kshiftrd $16, %k1, %k1
 ; CHECK-SKX-NEXT:    vpblendmd %zmm2, %zmm4, %zmm1 {%k1}
 ; CHECK-SKX-NEXT:    retq
 ;
@@ -179,10 +165,9 @@ define <32 x i32> @test7(<32 x i16> %x, <32 x i32> %a, <32 x i32> %b) {
 define <64 x i16> @test8(<64 x i8> %x, <64 x i16> %a, <64 x i16> %b) {
 ; CHECK-SKX-LABEL: test8:
 ; CHECK-SKX:       # %bb.0:
-; CHECK-SKX-NEXT:    vextracti64x4 $1, %zmm0, %ymm5
-; CHECK-SKX-NEXT:    vptestnmb %ymm5, %ymm5, %k1
-; CHECK-SKX-NEXT:    vptestnmb %ymm0, %ymm0, %k2
-; CHECK-SKX-NEXT:    vpblendmw %zmm1, %zmm3, %zmm0 {%k2}
+; CHECK-SKX-NEXT:    vptestnmb %zmm0, %zmm0, %k1
+; CHECK-SKX-NEXT:    vpblendmw %zmm1, %zmm3, %zmm0 {%k1}
+; CHECK-SKX-NEXT:    kshiftrq $32, %k1, %k1
 ; CHECK-SKX-NEXT:    vpblendmw %zmm2, %zmm4, %zmm1 {%k1}
 ; CHECK-SKX-NEXT:    retq
 ;

diff --git a/test/CodeGen/X86/min-legal-vector-width.ll b/test/CodeGen/X86/min-legal-vector-width.ll
index 46e73c1f8542bf8e5f0c1f7b1048f6ae7239ddf7..88329600b2336ac09a4607ee656b6349282d6ddb 100644 (file)
--- a/test/CodeGen/X86/min-legal-vector-width.ll
+++ b/test/CodeGen/X86/min-legal-vector-width.ll
@@ -1013,9 +1013,7 @@ define void @vselect_split_v8i16_setcc(<8 x i16> %s, <8 x i16> %t, <8 x i64>* %p
 ; CHECK-NEXT:    vmovdqa (%rsi), %ymm2
 ; CHECK-NEXT:    vmovdqa 32(%rsi), %ymm3
 ; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k1
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
-; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k2
+; CHECK-NEXT:    kshiftrb $4, %k1, %k2
 ; CHECK-NEXT:    vmovdqa64 32(%rdi), %ymm3 {%k2}
 ; CHECK-NEXT:    vmovdqa64 (%rdi), %ymm2 {%k1}
 ; CHECK-NEXT:    vmovdqa %ymm2, (%rdx)
@@ -1035,10 +1033,8 @@ define void @vselect_split_v8i32_setcc(<8 x i32> %s, <8 x i32> %t, <8 x i64>* %p
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vmovdqa (%rsi), %ymm2
 ; CHECK-NEXT:    vmovdqa 32(%rsi), %ymm3
-; CHECK-NEXT:    vpcmpeqd %xmm1, %xmm0, %k1
-; CHECK-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; CHECK-NEXT:    vpcmpeqd %xmm1, %xmm0, %k2
+; CHECK-NEXT:    vpcmpeqd %ymm1, %ymm0, %k1
+; CHECK-NEXT:    kshiftrb $4, %k1, %k2
 ; CHECK-NEXT:    vmovdqa64 32(%rdi), %ymm3 {%k2}
 ; CHECK-NEXT:    vmovdqa64 (%rdi), %ymm2 {%k1}
 ; CHECK-NEXT:    vmovdqa %ymm2, (%rdx)
@@ -1059,9 +1055,7 @@ define void @vselect_split_v16i8_setcc(<16 x i8> %s, <16 x i8> %t, <16 x i32>* %
 ; CHECK-NEXT:    vmovdqa (%rsi), %ymm2
 ; CHECK-NEXT:    vmovdqa 32(%rsi), %ymm3
 ; CHECK-NEXT:    vpcmpeqb %xmm1, %xmm0, %k1
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
-; CHECK-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
-; CHECK-NEXT:    vpcmpeqb %xmm1, %xmm0, %k2
+; CHECK-NEXT:    kshiftrw $8, %k1, %k2
 ; CHECK-NEXT:    vmovdqa32 32(%rdi), %ymm3 {%k2}
 ; CHECK-NEXT:    vmovdqa32 (%rdi), %ymm2 {%k1}
 ; CHECK-NEXT:    vmovdqa %ymm2, (%rdx)
@@ -1081,10 +1075,8 @@ define void @vselect_split_v16i16_setcc(<16 x i16> %s, <16 x i16> %t, <16 x i32>
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vmovdqa (%rsi), %ymm2
 ; CHECK-NEXT:    vmovdqa 32(%rsi), %ymm3
-; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k1
-; CHECK-NEXT:    vextracti128 $1, %ymm1, %xmm1
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; CHECK-NEXT:    vpcmpeqw %xmm1, %xmm0, %k2
+; CHECK-NEXT:    vpcmpeqw %ymm1, %ymm0, %k1
+; CHECK-NEXT:    kshiftrw $8, %k1, %k2
 ; CHECK-NEXT:    vmovdqa32 32(%rdi), %ymm3 {%k2}
 ; CHECK-NEXT:    vmovdqa32 (%rdi), %ymm2 {%k1}
 ; CHECK-NEXT:    vmovdqa %ymm2, (%rdx)

diff --git a/test/CodeGen/X86/pr34177.ll b/test/CodeGen/X86/pr34177.ll
index 056682bb2750b27d956a82f4ef0c0b6cc2a4e9cb..f8ead6352f1d3528917a5fd6e88e273b98735105 100644 (file)
--- a/test/CodeGen/X86/pr34177.ll
+++ b/test/CodeGen/X86/pr34177.ll
@@ -6,45 +6,88 @@ target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 define void @test(<4 x i64> %a, <4 x x86_fp80> %b, <8 x x86_fp80>* %c) local_unnamed_addr {
-; CHECK-LABEL: test:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovq %xmm0, %rax
-; CHECK-NEXT:    vpextrq $1, %xmm0, %rcx
-; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
-; CHECK-NEXT:    vmovq %xmm0, %rdx
-; CHECK-NEXT:    vpextrq $1, %xmm0, %rsi
-; CHECK-NEXT:    cmpq $3, %rsi
-; CHECK-NEXT:    fld1
-; CHECK-NEXT:    fldz
-; CHECK-NEXT:    fld %st(0)
-; CHECK-NEXT:    fcmove %st(2), %st
-; CHECK-NEXT:    cmpq $2, %rdx
-; CHECK-NEXT:    fld %st(1)
-; CHECK-NEXT:    fcmove %st(3), %st
-; CHECK-NEXT:    cmpq $1, %rcx
-; CHECK-NEXT:    fld %st(2)
-; CHECK-NEXT:    fcmove %st(4), %st
-; CHECK-NEXT:    testq %rax, %rax
-; CHECK-NEXT:    fxch %st(3)
-; CHECK-NEXT:    fcmove %st(4), %st
-; CHECK-NEXT:    fstp %st(4)
-; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    fstpt 70(%rdi)
-; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    fstpt 50(%rdi)
-; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    fstpt 30(%rdi)
-; CHECK-NEXT:    fldt {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    fstpt 10(%rdi)
-; CHECK-NEXT:    fxch %st(1)
-; CHECK-NEXT:    fadd %st, %st(0)
-; CHECK-NEXT:    fstpt 60(%rdi)
-; CHECK-NEXT:    fadd %st, %st(0)
-; CHECK-NEXT:    fstpt 40(%rdi)
-; CHECK-NEXT:    fadd %st, %st(0)
-; CHECK-NEXT:    fstpt 20(%rdi)
-; CHECK-NEXT:    fadd %st, %st(0)
-; CHECK-NEXT:    fstpt (%rdi)
+; AVX512F-LABEL: test:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vmovq %xmm0, %rax
+; AVX512F-NEXT:    vpextrq $1, %xmm0, %rcx
+; AVX512F-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX512F-NEXT:    vmovq %xmm0, %rdx
+; AVX512F-NEXT:    vpextrq $1, %xmm0, %rsi
+; AVX512F-NEXT:    cmpq $3, %rsi
+; AVX512F-NEXT:    fld1
+; AVX512F-NEXT:    fldz
+; AVX512F-NEXT:    fld %st(0)
+; AVX512F-NEXT:    fcmove %st(2), %st
+; AVX512F-NEXT:    cmpq $2, %rdx
+; AVX512F-NEXT:    fld %st(1)
+; AVX512F-NEXT:    fcmove %st(3), %st
+; AVX512F-NEXT:    cmpq $1, %rcx
+; AVX512F-NEXT:    fld %st(2)
+; AVX512F-NEXT:    fcmove %st(4), %st
+; AVX512F-NEXT:    testq %rax, %rax
+; AVX512F-NEXT:    fxch %st(3)
+; AVX512F-NEXT:    fcmove %st(4), %st
+; AVX512F-NEXT:    fstp %st(4)
+; AVX512F-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512F-NEXT:    fstpt 70(%rdi)
+; AVX512F-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512F-NEXT:    fstpt 50(%rdi)
+; AVX512F-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512F-NEXT:    fstpt 30(%rdi)
+; AVX512F-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512F-NEXT:    fstpt 10(%rdi)
+; AVX512F-NEXT:    fxch %st(1)
+; AVX512F-NEXT:    fadd %st, %st(0)
+; AVX512F-NEXT:    fstpt 60(%rdi)
+; AVX512F-NEXT:    fadd %st, %st(0)
+; AVX512F-NEXT:    fstpt 40(%rdi)
+; AVX512F-NEXT:    fadd %st, %st(0)
+; AVX512F-NEXT:    fstpt 20(%rdi)
+; AVX512F-NEXT:    fadd %st, %st(0)
+; AVX512F-NEXT:    fstpt (%rdi)
+;
+; AVX512VL-LABEL: test:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vpcmpeqq {{.*}}(%rip), %ymm0, %k0
+; AVX512VL-NEXT:    kshiftrb $2, %k0, %k1
+; AVX512VL-NEXT:    kshiftrb $1, %k0, %k2
+; AVX512VL-NEXT:    kmovd %k0, %eax
+; AVX512VL-NEXT:    testb $1, %al
+; AVX512VL-NEXT:    fld1
+; AVX512VL-NEXT:    fldz
+; AVX512VL-NEXT:    fld %st(0)
+; AVX512VL-NEXT:    fcmovne %st(2), %st
+; AVX512VL-NEXT:    kshiftrb $1, %k1, %k0
+; AVX512VL-NEXT:    kmovd %k0, %eax
+; AVX512VL-NEXT:    testb $1, %al
+; AVX512VL-NEXT:    fld %st(1)
+; AVX512VL-NEXT:    fcmovne %st(3), %st
+; AVX512VL-NEXT:    kmovd %k1, %eax
+; AVX512VL-NEXT:    testb $1, %al
+; AVX512VL-NEXT:    fld %st(2)
+; AVX512VL-NEXT:    fcmovne %st(4), %st
+; AVX512VL-NEXT:    kmovd %k2, %eax
+; AVX512VL-NEXT:    testb $1, %al
+; AVX512VL-NEXT:    fxch %st(3)
+; AVX512VL-NEXT:    fcmovne %st(4), %st
+; AVX512VL-NEXT:    fstp %st(4)
+; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512VL-NEXT:    fstpt 70(%rdi)
+; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512VL-NEXT:    fstpt 50(%rdi)
+; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512VL-NEXT:    fstpt 30(%rdi)
+; AVX512VL-NEXT:    fldt {{[0-9]+}}(%rsp)
+; AVX512VL-NEXT:    fstpt 10(%rdi)
+; AVX512VL-NEXT:    fxch %st(1)
+; AVX512VL-NEXT:    fadd %st, %st(0)
+; AVX512VL-NEXT:    fstpt (%rdi)
+; AVX512VL-NEXT:    fadd %st, %st(0)
+; AVX512VL-NEXT:    fstpt 60(%rdi)
+; AVX512VL-NEXT:    fadd %st, %st(0)
+; AVX512VL-NEXT:    fstpt 40(%rdi)
+; AVX512VL-NEXT:    fadd %st, %st(0)
+; AVX512VL-NEXT:    fstpt 20(%rdi)
   %1 = icmp eq <4 x i64> <i64 0, i64 1, i64 2, i64 3>, %a
   %2 = select <4 x i1> %1, <4 x x86_fp80> <x86_fp80 0xK3FFF8000000000000000, x86_fp80 0xK3FFF8000000000000000, x86_fp80 0xK3FFF8000000000000000, x86_fp80 0xK3FFF8000000000000000>, <4 x x86_fp80> zeroinitializer
   %3 = fadd <4 x x86_fp80> %2, %2