From 9c32ef6fdbc10207c1bf08d5e1f1fcd0b9647598 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sat, 3 Jun 2017 11:12:57 +0000
Subject: [PATCH] [X86][SSE] Add SCALAR_TO_VECTOR(PEXTRW/PEXTRB) support to
 faux shuffle combining

Generalized existing SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT) code to support AssertZext + PEXTRW/PEXTRB cases as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@304659 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp            | 40 +++++++++++++----
 .../X86/clear_upper_vector_element_bits.ll    | 44 +++----------------
 2 files changed, 36 insertions(+), 48 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 63874767b71..5303d7a406a 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5848,17 +5848,39 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
     return true;
   }
   case ISD::SCALAR_TO_VECTOR: {
-    // Match against a scalar_to_vector of an extract from a similar vector.
+    // Match against a scalar_to_vector of an extract from a vector,
+    // for PEXTRW/PEXTRB we must handle the implicit zext of the scalar.
     SDValue N0 = N.getOperand(0);
-    if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
-        N0.getOperand(0).getValueType() != VT ||
-        !isa<ConstantSDNode>(N0.getOperand(1)) ||
-        NumElts <= N0.getConstantOperandVal(1) ||
-        !N->isOnlyUserOf(N0.getNode()))
+    SDValue SrcExtract;
+
+    if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+        N0.getOperand(0).getValueType() == VT) {
+      SrcExtract = N0;
+    } else if (N0.getOpcode() == ISD::AssertZext &&
+               N0.getOperand(0).getOpcode() == X86ISD::PEXTRW &&
+               cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i16) {
+      SrcExtract = N0.getOperand(0);
+      assert(SrcExtract.getOperand(0).getValueType() == MVT::v8i16);
+    } else if (N0.getOpcode() == ISD::AssertZext &&
+               N0.getOperand(0).getOpcode() == X86ISD::PEXTRB &&
+               cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i8) {
+      SrcExtract = N0.getOperand(0);
+      assert(SrcExtract.getOperand(0).getValueType() == MVT::v16i8);
+    }
+
+    if (!SrcExtract || !isa<ConstantSDNode>(SrcExtract.getOperand(1)) ||
+        NumElts <= SrcExtract.getConstantOperandVal(1))
       return false;
-    Ops.push_back(N0.getOperand(0));
-    Mask.push_back(N0.getConstantOperandVal(1));
-    Mask.append(NumElts - 1, SM_SentinelUndef);
+
+    SDValue SrcVec = SrcExtract.getOperand(0);
+    EVT SrcVT = SrcVec.getValueType();
+    unsigned NumSrcElts = SrcVT.getVectorNumElements();
+    unsigned NumZeros = (NumBitsPerElt / SrcVT.getScalarSizeInBits()) - 1;
+
+    Ops.push_back(SrcVec);
+    Mask.push_back(SrcExtract.getConstantOperandVal(1));
+    Mask.append(NumZeros, SM_SentinelZero);
+    Mask.append(NumSrcElts - Mask.size(), SM_SentinelUndef);
     return true;
   }
   case X86ISD::PINSRB:
diff --git a/test/CodeGen/X86/clear_upper_vector_element_bits.ll b/test/CodeGen/X86/clear_upper_vector_element_bits.ll
index ae0f4406ba0..1218b68b1be 100644
--- a/test/CodeGen/X86/clear_upper_vector_element_bits.ll
+++ b/test/CodeGen/X86/clear_upper_vector_element_bits.ll
@@ -405,12 +405,7 @@ define <16 x i8> @_clearupper16xi8a(<16 x i8>) nounwind {
 ;
 ; AVX-LABEL: _clearupper16xi8a:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX-NEXT:    vpextrb $1, %xmm0, %ecx
-; AVX-NEXT:    vmovd %eax, %xmm1
-; AVX-NEXT:    vpinsrb $1, %ecx, %xmm1, %xmm1
-; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3,4,5,6,7]
-; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    vandps {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %x0  = extractelement <16 x i8> %0, i32 0
   %x1  = extractelement <16 x i8> %0, i32 1
@@ -575,39 +570,10 @@ define <32 x i8> @_clearupper32xi8a(<32 x i8>) nounwind {
 ; SSE-NEXT:    pand %xmm2, %xmm1
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: _clearupper32xi8a:
-; AVX1:       # BB#0:
-; AVX1-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX1-NEXT:    vpextrb $1, %xmm0, %ecx
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vpextrb $0, %xmm1, %edx
-; AVX1-NEXT:    vpextrb $1, %xmm1, %esi
-; AVX1-NEXT:    vmovd %edx, %xmm2
-; AVX1-NEXT:    vpinsrb $1, %esi, %xmm2, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4,5,6,7]
-; AVX1-NEXT:    vmovd %eax, %xmm2
-; AVX1-NEXT:    vpinsrb $1, %ecx, %xmm2, %xmm2
-; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3,4,5,6,7]
-; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: _clearupper32xi8a:
-; AVX2:       # BB#0:
-; AVX2-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX2-NEXT:    vpextrb $1, %xmm0, %ecx
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT:    vpextrb $0, %xmm1, %edx
-; AVX2-NEXT:    vpextrb $1, %xmm1, %esi
-; AVX2-NEXT:    vmovd %edx, %xmm2
-; AVX2-NEXT:    vpinsrb $1, %esi, %xmm2, %xmm2
-; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1,2,3,4,5,6,7]
-; AVX2-NEXT:    vmovd %eax, %xmm2
-; AVX2-NEXT:    vpinsrb $1, %ecx, %xmm2, %xmm2
-; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3,4,5,6,7]
-; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpand {{.*}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: _clearupper32xi8a:
+; AVX:       # BB#0:
+; AVX-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm0
+; AVX-NEXT:    retq
   %x0  = extractelement <32 x i8> %0, i32 0
   %x1  = extractelement <32 x i8> %0, i32 1
   %x2  = extractelement <32 x i8> %0, i32 2
-- 
2.50.1