From 7c56bce996ff74a57e6536f2bb3c23612fd606fb Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Fri, 20 Jul 2018 15:24:12 +0000
Subject: [PATCH] [X86][AVX] Add support for 32/64 bits 256-bit vector
 horizontal op redundant shuffle removal

llvm-svn: 337561
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 16 +++++++++++-----
 llvm/test/CodeGen/X86/haddsub-shuf.ll   |  6 ------
 2 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d9b42f6..e1b7adf 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -31144,8 +31144,8 @@ static SDValue foldShuffleOfHorizOp(SDNode *N) {
   // lanes of each operand as:
   // v4X32: A[0] + A[1] , A[2] + A[3] , B[0] + B[1] , B[2] + B[3]
   // ...similarly for v2f64 and v8i16.
-  // TODO: 256-bit is not the same because...x86.
-  if (HOp.getOperand(0) != HOp.getOperand(1) || HOp.getValueSizeInBits() != 128)
+  // TODO: Handle UNDEF operands.
+  if (HOp.getOperand(0) != HOp.getOperand(1))
     return SDValue();
 
   // When the operands of a horizontal math op are identical, the low half of
@@ -31156,9 +31156,15 @@ static SDValue foldShuffleOfHorizOp(SDNode *N) {
   // TODO: Other mask possibilities like {1,1} and {1,0} could be added here,
   // but this should be tied to whatever horizontal op matching and shuffle
   // canonicalization are producing.
-  if (isTargetShuffleEquivalent(Mask, { 0, 0 }) ||
-      isTargetShuffleEquivalent(Mask, { 0, 1, 0, 1 }) ||
-      isTargetShuffleEquivalent(Mask, { 0, 1, 2, 3, 0, 1, 2, 3 }))
+  if (HOp.getValueSizeInBits() == 128 &&
+      (isTargetShuffleEquivalent(Mask, {0, 0}) ||
+       isTargetShuffleEquivalent(Mask, {0, 1, 0, 1}) ||
+       isTargetShuffleEquivalent(Mask, {0, 1, 2, 3, 0, 1, 2, 3})))
+    return HOp;
+
+  if (HOp.getValueSizeInBits() == 256 &&
+      (isTargetShuffleEquivalent(Mask, {0, 0, 2, 2}) ||
+       isTargetShuffleEquivalent(Mask, {0, 1, 0, 1, 4, 5, 4, 5})))
     return HOp;
 
   return SDValue();
diff --git a/llvm/test/CodeGen/X86/haddsub-shuf.ll b/llvm/test/CodeGen/X86/haddsub-shuf.ll
index 8ede703..1cb9bcc 100644
--- a/llvm/test/CodeGen/X86/haddsub-shuf.ll
+++ b/llvm/test/CodeGen/X86/haddsub-shuf.ll
@@ -63,7 +63,6 @@ define <8 x float> @hadd_v8f32b(<8 x float> %a) {
 ; AVX-LABEL: hadd_v8f32b:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vhaddps %ymm0, %ymm0, %ymm0
-; AVX-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <8 x float> %a, <8 x float> undef, <8 x i32> <i32 0, i32 2, i32 undef, i32 undef, i32 4, i32 6, i32 undef, i32 undef>
   %a1 = shufflevector <8 x float> %a, <8 x float> undef, <8 x i32> <i32 1, i32 3, i32 undef, i32 undef, i32 5, i32 7, i32 undef, i32 undef>
@@ -129,7 +128,6 @@ define <8 x float> @hsub_v8f32b(<8 x float> %a) {
 ; AVX-LABEL: hsub_v8f32b:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vhsubps %ymm0, %ymm0, %ymm0
-; AVX-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <8 x float> %a, <8 x float> undef, <8 x i32> <i32 0, i32 2, i32 undef, i32 undef, i32 4, i32 6, i32 undef, i32 undef>
   %a1 = shufflevector <8 x float> %a, <8 x float> undef, <8 x i32> <i32 1, i32 3, i32 undef, i32 undef, i32 5, i32 7, i32 undef, i32 undef>
@@ -165,7 +163,6 @@ define <4 x double> @hadd_v4f64(<4 x double> %a) {
 ; AVX-LABEL: hadd_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vhaddpd %ymm0, %ymm0, %ymm0
-; AVX-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <4 x double> %a, <4 x double> undef, <4 x i32> <i32 0, i32 undef, i32 2, i32 undef>
   %a1 = shufflevector <4 x double> %a, <4 x double> undef, <4 x i32> <i32 1, i32 undef, i32 3, i32 undef>
@@ -201,7 +198,6 @@ define <4 x double> @hsub_v4f64(<4 x double> %a) {
 ; AVX-LABEL: hsub_v4f64:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vhsubpd %ymm0, %ymm0, %ymm0
-; AVX-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <4 x double> %a, <4 x double> undef, <4 x i32> <i32 0, i32 undef, i32 2, i32 undef>
   %a1 = shufflevector <4 x double> %a, <4 x double> undef, <4 x i32> <i32 1, i32 undef, i32 3, i32 undef>
@@ -279,7 +275,6 @@ define <8 x i32> @hadd_v8i32b(<8 x i32> %a) {
 ; AVX2-LABEL: hadd_v8i32b:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,1,0,1,4,5,4,5]
 ; AVX2-NEXT:    retq
   %a0 = shufflevector <8 x i32> %a, <8 x i32> undef, <8 x i32> <i32 0, i32 2, i32 undef, i32 undef, i32 4, i32 6, i32 undef, i32 undef>
   %a1 = shufflevector <8 x i32> %a, <8 x i32> undef, <8 x i32> <i32 1, i32 3, i32 undef, i32 undef, i32 5, i32 7, i32 undef, i32 undef>
@@ -357,7 +352,6 @@ define <8 x i32> @hsub_v8i32b(<8 x i32> %a) {
 ; AVX2-LABEL: hsub_v8i32b:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vphsubd %ymm0, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,1,0,1,4,5,4,5]
 ; AVX2-NEXT:    retq
   %a0 = shufflevector <8 x i32> %a, <8 x i32> undef, <8 x i32> <i32 0, i32 2, i32 undef, i32 undef, i32 4, i32 6, i32 undef, i32 undef>
   %a1 = shufflevector <8 x i32> %a, <8 x i32> undef, <8 x i32> <i32 1, i32 3, i32 undef, i32 undef, i32 5, i32 7, i32 undef, i32 undef>
-- 
2.7.4