From 22d51014afb3b29e2918bf7cf8c17b04944da6b0 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sat, 29 Sep 2018 14:17:32 +0000
Subject: [PATCH] [X86] getTargetConstantBitsFromNode - add support for peeking
 through ISD::EXTRACT_SUBVECTOR

llvm-svn: 343375
---
 llvm/lib/Target/X86/X86ISelLowering.cpp        | 15 +++++++++++++++
 llvm/test/CodeGen/X86/pr38639.ll               | 11 +++++------
 llvm/test/CodeGen/X86/vector-shuffle-avx512.ll |  8 ++++----
 3 files changed, 24 insertions(+), 10 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 58a1865..a71c327 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5742,6 +5742,21 @@ static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
     return CastBitData(UndefSrcElts, SrcEltBits);
   }
 
+  // Extract constant bits from a subvector's source.
+  if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+      isa<ConstantSDNode>(Op.getOperand(1))) {
+    if (getTargetConstantBitsFromNode(Op.getOperand(0), EltSizeInBits,
+                                      UndefElts, EltBits, AllowWholeUndefs,
+                                      AllowPartialUndefs)) {
+      unsigned NumSubElts = VT.getVectorNumElements();
+      unsigned BaseIdx = Op.getConstantOperandVal(1);
+      UndefElts = UndefElts.extractBits(NumSubElts, BaseIdx);
+      EltBits.erase(EltBits.begin() + BaseIdx + NumSubElts, EltBits.end());
+      EltBits.erase(EltBits.begin(), EltBits.begin() + BaseIdx);
+      return true;
+    }
+  }
+
   return false;
 }
 
diff --git a/llvm/test/CodeGen/X86/pr38639.ll b/llvm/test/CodeGen/X86/pr38639.ll
index c877568..502cf43 100644
--- a/llvm/test/CodeGen/X86/pr38639.ll
+++ b/llvm/test/CodeGen/X86/pr38639.ll
@@ -4,12 +4,11 @@
 define <8 x double> @test(<4 x double> %a, <4 x double> %b) {
 ; CHECK-LABEL: test:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovaps {{.*#+}} ymm2 = <u,0.82071743224100002,0.82071743224100002,0.82071743224100002>
-; CHECK-NEXT:    vextractf128 $1, %ymm2, %xmm4
-; CHECK-NEXT:    vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; CHECK-NEXT:    vblendps {{.*#+}} ymm3 = ymm0[0,1],ymm2[2,3],ymm0[4,5,6,7]
-; CHECK-NEXT:    vblendps {{.*#+}} xmm2 = xmm4[0,1],xmm2[2,3]
-; CHECK-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm3[1],ymm1[1],ymm3[3],ymm1[3]
+; CHECK-NEXT:    vmovaps {{.*#+}} ymm1 = <u,0.82071743224100002,0.82071743224100002,0.82071743224100002>
+; CHECK-NEXT:    vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3],ymm1[4,5,6,7]
+; CHECK-NEXT:    vblendps {{.*#+}} ymm1 = ymm0[0,1],ymm1[2,3],ymm0[4,5,6,7]
+; CHECK-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]
+; CHECK-NEXT:    vmovaps {{.*#+}} xmm2 = [8.207174e-01,8.207174e-01]
 ; CHECK-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
 ; CHECK-NEXT:    retq
   %1 = shufflevector <4 x double> %a, <4 x double> <double undef, double 0x3FEA435134576E1C, double 0x3FEA435134576E1C, double 0x3FEA435134576E1C>, <8 x i32> <i32 6, i32 5, i32 2, i32 3, i32 5, i32 1, i32 3, i32 7>
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll b/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
index 70d3065..de1f51b 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
@@ -541,7 +541,7 @@ define <8 x float> @expand14(<4 x float> %a) {
 define <8 x float> @expand15(<4 x float> %a) {
 ; SKX64-LABEL: expand15:
 ; SKX64:       # %bb.0:
-; SKX64-NEXT:    vpermilps {{.*#+}} xmm1 = mem[0,1,0,0]
+; SKX64-NEXT:    vmovaps {{.*#+}} xmm1 = [0.000000e+00,2.000000e+00,0.000000e+00,0.000000e+00]
 ; SKX64-NEXT:    vpermilps {{.*#+}} xmm2 = xmm0[0,1,1,3]
 ; SKX64-NEXT:    vmovaps {{.*#+}} ymm0 = [0,1,8,3,10,3,2,3]
 ; SKX64-NEXT:    vpermi2ps %ymm2, %ymm1, %ymm0
@@ -549,7 +549,7 @@ define <8 x float> @expand15(<4 x float> %a) {
 ;
 ; KNL64-LABEL: expand15:
 ; KNL64:       # %bb.0:
-; KNL64-NEXT:    vpermilps {{.*#+}} xmm1 = mem[0,1,0,0]
+; KNL64-NEXT:    vmovaps {{.*#+}} xmm1 = [0.000000e+00,2.000000e+00,0.000000e+00,0.000000e+00]
 ; KNL64-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,1,1]
 ; KNL64-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
 ; KNL64-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
@@ -558,7 +558,7 @@ define <8 x float> @expand15(<4 x float> %a) {
 ;
 ; SKX32-LABEL: expand15:
 ; SKX32:       # %bb.0:
-; SKX32-NEXT:    vpermilps {{.*#+}} xmm1 = mem[0,1,0,0]
+; SKX32-NEXT:    vmovaps {{.*#+}} xmm1 = [0.000000e+00,2.000000e+00,0.000000e+00,0.000000e+00]
 ; SKX32-NEXT:    vpermilps {{.*#+}} xmm2 = xmm0[0,1,1,3]
 ; SKX32-NEXT:    vmovaps {{.*#+}} ymm0 = [0,1,8,3,10,3,2,3]
 ; SKX32-NEXT:    vpermi2ps %ymm2, %ymm1, %ymm0
@@ -566,7 +566,7 @@ define <8 x float> @expand15(<4 x float> %a) {
 ;
 ; KNL32-LABEL: expand15:
 ; KNL32:       # %bb.0:
-; KNL32-NEXT:    vpermilps {{.*#+}} xmm1 = mem[0,1,0,0]
+; KNL32-NEXT:    vmovaps {{.*#+}} xmm1 = [0.000000e+00,2.000000e+00,0.000000e+00,0.000000e+00]
 ; KNL32-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[0,1,1,1]
 ; KNL32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
 ; KNL32-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
-- 
2.7.4