From 1513e5ffeebd7a0da092f1d970d037b1a7a385cb Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Fri, 10 May 2019 15:46:04 +0000
Subject: [PATCH] [X86][SSE] Add getHopForBuildVector vector splitting

If we only use the lower xmm of a ymm hop, then extract the xmm's (for free), perform the xmm hop and then insert back into a ymm (for free).

Fixes some of the regressions noted in D61782

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@360435 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp | 16 +++++++++++
 test/CodeGen/X86/haddsub-undef.ll  |  4 +--
 test/CodeGen/X86/phaddsub-undef.ll | 56 +++++++++-----------------------------
 3 files changed, 31 insertions(+), 45 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 950e0f4c8e7..f7f3d38f68e 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -8581,6 +8581,22 @@ static SDValue getHopForBuildVector(const BuildVectorSDNode *BV,
   else if (V1.getValueSizeInBits() < Width)
     V1 = insertSubVector(DAG.getUNDEF(VT), V1, 0, DAG, SDLoc(BV), Width);
 
+  unsigned NumElts = VT.getVectorNumElements();
+  APInt DemandedElts = APInt::getAllOnesValue(NumElts);
+  for (unsigned i = 0; i != NumElts; ++i)
+    if (BV->getOperand(i).isUndef())
+      DemandedElts.clearBit(i);
+
+  // If we don't need the upper xmm, then perform as a xmm hop.
+  unsigned HalfNumElts = NumElts / 2;
+  if (VT.is256BitVector() && DemandedElts.lshr(HalfNumElts) == 0) {
+    MVT HalfVT = MVT::getVectorVT(VT.getScalarType(), HalfNumElts);
+    V0 = extractSubVector(V0, 0, DAG, SDLoc(BV), 128);
+    V1 = extractSubVector(V1, 0, DAG, SDLoc(BV), 128);
+    SDValue Half = DAG.getNode(HOpcode, SDLoc(BV), HalfVT, V0, V1);
+    return insertSubVector(DAG.getUNDEF(VT), Half, 0, DAG, SDLoc(BV), 256);
+  }
+
   return DAG.getNode(HOpcode, SDLoc(BV), VT, V0, V1);
 }
 
diff --git a/test/CodeGen/X86/haddsub-undef.ll b/test/CodeGen/X86/haddsub-undef.ll
index a5bfaaf566c..c8eb9d7d1c3 100644
--- a/test/CodeGen/X86/haddsub-undef.ll
+++ b/test/CodeGen/X86/haddsub-undef.ll
@@ -247,7 +247,7 @@ define <8 x float> @test10_undef(<8 x float> %a, <8 x float> %b) {
 ;
 ; AVX-LABEL: test10_undef:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %vecext = extractelement <8 x float> %a, i32 0
   %vecext1 = extractelement <8 x float> %a, i32 1
@@ -300,7 +300,7 @@ define <8 x float> @test12_undef(<8 x float> %a, <8 x float> %b) {
 ;
 ; AVX-LABEL: test12_undef:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vhaddps %ymm0, %ymm0, %ymm0
+; AVX-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %vecext = extractelement <8 x float> %a, i32 0
   %vecext1 = extractelement <8 x float> %a, i32 1
diff --git a/test/CodeGen/X86/phaddsub-undef.ll b/test/CodeGen/X86/phaddsub-undef.ll
index 6fffbefa427..b0be5c7eede 100644
--- a/test/CodeGen/X86/phaddsub-undef.ll
+++ b/test/CodeGen/X86/phaddsub-undef.ll
@@ -16,20 +16,10 @@ define <8 x i32> @test14_undef(<8 x i32> %a, <8 x i32> %b) {
 ; SSE-NEXT:    phaddd %xmm2, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test14_undef:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: test14_undef:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vphaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512-LABEL: test14_undef:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vphaddd %ymm1, %ymm0, %ymm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test14_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %vecext = extractelement <8 x i32> %a, i32 0
   %vecext1 = extractelement <8 x i32> %a, i32 1
   %add = add i32 %vecext, %vecext1
@@ -149,20 +139,10 @@ define <8 x i32> @test16_undef(<8 x i32> %a, <8 x i32> %b) {
 ; SSE-NEXT:    phaddd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test16_undef:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: test16_undef:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512-LABEL: test16_undef:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test16_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %vecext = extractelement <8 x i32> %a, i32 0
   %vecext1 = extractelement <8 x i32> %a, i32 1
   %add = add i32 %vecext, %vecext1
@@ -180,20 +160,10 @@ define <16 x i32> @test16_v16i32_undef(<16 x i32> %a, <16 x i32> %b) {
 ; SSE-NEXT:    phaddd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test16_v16i32_undef:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: test16_v16i32_undef:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
-; AVX2-NEXT:    retq
-;
-; AVX512-LABEL: test16_v16i32_undef:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test16_v16i32_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %vecext = extractelement <16 x i32> %a, i32 0
   %vecext1 = extractelement <16 x i32> %a, i32 1
   %add = add i32 %vecext, %vecext1
@@ -268,7 +238,7 @@ define <16 x i32> @test17_v16i32_undef(<16 x i32> %a, <16 x i32> %b) {
 ; AVX512-LABEL: test17_v16i32_undef:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX512-NEXT:    vphaddd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
   %vecext = extractelement <16 x i32> %a, i32 0
   %vecext1 = extractelement <16 x i32> %a, i32 1
-- 
2.11.0