From 317fad5921b8e2b12ac88453f7fca189c2830590 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Wed, 30 Jan 2019 16:19:19 +0000
Subject: [PATCH] [X86][AVX] Prefer to combine shuffle to broadcasts whenever
 possible

This is the first step towards improving broadcast support on AVX1 targets.

llvm-svn: 352634
---
 llvm/lib/Target/X86/X86ISelLowering.cpp    | 25 ++++++++++++----------
 llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll | 34 ++++++++++++++++++++----------
 2 files changed, 37 insertions(+), 22 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 0fdbf76..fd42dcf 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -30503,7 +30503,7 @@ static bool matchUnaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
   // instructions are no slower than UNPCKLPD but has the option to
   // fold the input operand into even an unaligned memory load.
   if (MaskVT.is128BitVector() && Subtarget.hasSSE3() && AllowFloatDomain) {
-    if (!Subtarget.hasAVX2() && isTargetShuffleEquivalent(Mask, {0, 0})) {
+    if (isTargetShuffleEquivalent(Mask, {0, 0})) {
       Shuffle = X86ISD::MOVDDUP;
       SrcVT = DstVT = MVT::v2f64;
       return true;
@@ -30561,16 +30561,6 @@ static bool matchUnaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
     }
   }
 
-  // Attempt to match against broadcast-from-vector.
-  if (Subtarget.hasAVX2()) {
-    SmallVector<int, 64> BroadcastMask(NumMaskElts, 0);
-    if (isTargetShuffleEquivalent(Mask, BroadcastMask)) {
-      SrcVT = DstVT = MaskVT;
-      Shuffle = X86ISD::VBROADCAST;
-      return true;
-    }
-  }
-
   return false;
 }
 
@@ -31050,6 +31040,19 @@ static SDValue combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
       }
     }
 
+    // Attempt to match against broadcast-from-vector.
+    // TODO: Add (partial) AVX1 support.
+    if (Subtarget.hasAVX2() && (!IsEVEXShuffle || NumRootElts == NumMaskElts)) {
+      SmallVector<int, 64> BroadcastMask(NumMaskElts, 0);
+      if (isTargetShuffleEquivalent(Mask, BroadcastMask)) {
+        if (Depth == 1 && Root.getOpcode() == X86ISD::VBROADCAST)
+          return SDValue(); // Nothing to do!
+        Res = DAG.getBitcast(MaskVT, V1);
+        Res = DAG.getNode(X86ISD::VBROADCAST, DL, MaskVT, Res);
+        return DAG.getBitcast(RootVT, Res);
+      }
+    }
+
     SDValue NewV1 = V1; // Save operand in case early exit happens.
     if (matchUnaryShuffle(MaskVT, Mask, AllowFloatDomain, AllowIntDomain, NewV1,
                           DL, DAG, Subtarget, Shuffle, ShuffleSrcVT,
diff --git a/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll b/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
index 8041647..c43f293 100644
--- a/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
+++ b/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
@@ -403,17 +403,29 @@ define <4 x float> @test15(<4 x float> %A, <4 x float> %B) {
 ; SSE-NEXT:    movaps %xmm2, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: test15:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm1[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; AVX-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[3,1,2,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vmovsldup {{.*#+}} xmm1 = xmm2[0,0,2,2]
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
-; AVX-NEXT:    retq
+; AVX1-LABEL: test15:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; AVX1-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm1[1,1,3,3]
+; AVX1-NEXT:    vaddss %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[3,1,2,3]
+; AVX1-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vmovsldup {{.*#+}} xmm1 = xmm2[0,0,2,2]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: test15:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; AVX512-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm1[1,1,3,3]
+; AVX512-NEXT:    vaddss %xmm3, %xmm2, %xmm2
+; AVX512-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX512-NEXT:    vpermilps {{.*#+}} xmm1 = xmm1[3,1,2,3]
+; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vbroadcastss %xmm2, %xmm1
+; AVX512-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX512-NEXT:    retq
   %1 = extractelement <4 x float> %A, i32 1
   %2 = extractelement <4 x float> %B, i32 1
   %add = fadd float %1, %2
-- 
2.7.4