From e166f9ed844a67ec9e70bc425af72731f1b6bbac Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Sun, 13 Oct 2019 05:47:47 +0000
Subject: [PATCH] [X86] Enable v4i32->v4i16 and v8i16->v8i8 saturating
 truncates to use pack instructions with avx512.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@374705 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp      |  1 +
 test/CodeGen/X86/vector-trunc-packus.ll | 61 ++++---------------------
 test/CodeGen/X86/vector-trunc-ssat.ll   | 56 ++++-------------------
 3 files changed, 17 insertions(+), 101 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index c7d90ddfc9a..8d538198aef 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -39871,6 +39871,7 @@ static SDValue combineTruncateWithSat(SDValue In, EVT VT, const SDLoc &DL,
   // registers, we should go ahead and use the pack instructions if possible.
   bool PreferAVX512 = ((Subtarget.hasAVX512() && InSVT == MVT::i32) ||
                        (Subtarget.hasBWI() && InSVT == MVT::i16)) &&
+                      (InVT.getSizeInBits() > 128) &&
                       (Subtarget.hasVLX() || InVT.getSizeInBits() > 256) &&
                       !(!Subtarget.useAVX512Regs() && VT.getSizeInBits() >= 256);
 
diff --git a/test/CodeGen/X86/vector-trunc-packus.ll b/test/CodeGen/X86/vector-trunc-packus.ll
index 2f433d8b8f7..a95a81f435e 100644
--- a/test/CodeGen/X86/vector-trunc-packus.ll
+++ b/test/CodeGen/X86/vector-trunc-packus.ll
@@ -2392,37 +2392,13 @@ define <4 x i16> @trunc_packus_v4i32_v4i16(<4 x i32> %a0) {
 ; AVX-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
 ;
-; AVX512F-LABEL: trunc_packus_v4i32_v4i16:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: trunc_packus_v4i32_v4i16:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512VL-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512BW-LABEL: trunc_packus_v4i32_v4i16:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512BWVL-LABEL: trunc_packus_v4i32_v4i16:
-; AVX512BWVL:       # %bb.0:
-; AVX512BWVL-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BWVL-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    retq
+; AVX512-LABEL: trunc_packus_v4i32_v4i16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
+; AVX512-NEXT:    retq
 ;
 ; SKX-LABEL: trunc_packus_v4i32_v4i16:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; SKX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; SKX-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
 ; SKX-NEXT:    vpackusdw %xmm0, %xmm0, %xmm0
 ; SKX-NEXT:    retq
   %1 = icmp slt <4 x i32> %a0, <i32 65535, i32 65535, i32 65535, i32 65535>
@@ -5731,34 +5707,13 @@ define <8 x i8> @trunc_packus_v8i16_v8i8(<8 x i16> %a0) {
 ; AVX-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
 ;
-; AVX512F-LABEL: trunc_packus_v8i16_v8i8:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: trunc_packus_v8i16_v8i8:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512BW-LABEL: trunc_packus_v8i16_v8i8:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512BWVL-LABEL: trunc_packus_v8i16_v8i8:
-; AVX512BWVL:       # %bb.0:
-; AVX512BWVL-NEXT:    vpminsw {{.*}}(%rip), %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX512BWVL-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    retq
+; AVX512-LABEL: trunc_packus_v8i16_v8i8:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
+; AVX512-NEXT:    retq
 ;
 ; SKX-LABEL: trunc_packus_v8i16_v8i8:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vpminsw {{.*}}(%rip), %xmm0, %xmm0
-; SKX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; SKX-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
 ; SKX-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
 ; SKX-NEXT:    retq
   %1 = icmp slt <8 x i16> %a0, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
diff --git a/test/CodeGen/X86/vector-trunc-ssat.ll b/test/CodeGen/X86/vector-trunc-ssat.ll
index d15559e4031..d9d9c367cea 100644
--- a/test/CodeGen/X86/vector-trunc-ssat.ll
+++ b/test/CodeGen/X86/vector-trunc-ssat.ll
@@ -2380,34 +2380,13 @@ define <4 x i16> @trunc_ssat_v4i32_v4i16(<4 x i32> %a0) {
 ; AVX-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
 ;
-; AVX512F-LABEL: trunc_ssat_v4i32_v4i16:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: trunc_ssat_v4i32_v4i16:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpmaxsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512VL-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512BW-LABEL: trunc_ssat_v4i32_v4i16:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512BWVL-LABEL: trunc_ssat_v4i32_v4i16:
-; AVX512BWVL:       # %bb.0:
-; AVX512BWVL-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpmaxsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    retq
+; AVX512-LABEL: trunc_ssat_v4i32_v4i16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
+; AVX512-NEXT:    retq
 ;
 ; SKX-LABEL: trunc_ssat_v4i32_v4i16:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vpminsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; SKX-NEXT:    vpmaxsd {{.*}}(%rip){1to4}, %xmm0, %xmm0
 ; SKX-NEXT:    vpackssdw %xmm0, %xmm0, %xmm0
 ; SKX-NEXT:    retq
   %1 = icmp slt <4 x i32> %a0, <i32 32767, i32 32767, i32 32767, i32 32767>
@@ -5620,32 +5599,13 @@ define <8 x i8> @trunc_ssat_v8i16_v8i8(<8 x i16> %a0) {
 ; AVX-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
 ;
-; AVX512F-LABEL: trunc_ssat_v8i16_v8i8:
-; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: trunc_ssat_v8i16_v8i8:
-; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
-; AVX512VL-NEXT:    retq
-;
-; AVX512BW-LABEL: trunc_ssat_v8i16_v8i8:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
-; AVX512BW-NEXT:    retq
-;
-; AVX512BWVL-LABEL: trunc_ssat_v8i16_v8i8:
-; AVX512BWVL:       # %bb.0:
-; AVX512BWVL-NEXT:    vpminsw {{.*}}(%rip), %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpmaxsw {{.*}}(%rip), %xmm0, %xmm0
-; AVX512BWVL-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
-; AVX512BWVL-NEXT:    retq
+; AVX512-LABEL: trunc_ssat_v8i16_v8i8:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
+; AVX512-NEXT:    retq
 ;
 ; SKX-LABEL: trunc_ssat_v8i16_v8i8:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vpminsw {{.*}}(%rip), %xmm0, %xmm0
-; SKX-NEXT:    vpmaxsw {{.*}}(%rip), %xmm0, %xmm0
 ; SKX-NEXT:    vpacksswb %xmm0, %xmm0, %xmm0
 ; SKX-NEXT:    retq
   %1 = icmp slt <8 x i16> %a0, <i16 127, i16 127, i16 127, i16 127, i16 127, i16 127, i16 127, i16 127>
-- 
2.40.0