[llvm] r350830 - [x86] fix remaining miscompile bug in horizontal binop matching (PR40243)

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Thu Jan 10 07:27:23 2019
New Revision: 350830

URL: http://llvm.org/viewvc/llvm-project?rev=350830&view=rev
Log:
[x86] fix remaining miscompile bug in horizontal binop matching (PR40243)

When we use the partial-matching function on a 128-bit chunk, we must 
account for the possibility that we've matched undef halves of the
original source vectors, so the outputs may need to be reset.

This should allow closing PR40243:
https://bugs.llvm.org/show_bug.cgi?id=40243


Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=350830&r1=350829&r2=350830&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Thu Jan 10 07:27:23 2019
@@ -8370,7 +8370,6 @@ static SDValue LowerToHorizontalOp(const
   SDLoc DL(BV);
   SDValue InVec0, InVec1;
   if (VT == MVT::v8i32 || VT == MVT::v16i16) {
-    // Try to match an AVX2 horizontal add/sub of signed integers.
     SDValue InVec2, InVec3;
     unsigned X86Opcode;
     bool CanFold = true;
@@ -8397,12 +8396,16 @@ static SDValue LowerToHorizontalOp(const
       if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
         return SDValue();
 
-      // Convert this build_vector into a pair of horizontal binop followed by
-      // a concat vector.
+      // Convert this build_vector into a pair of horizontal binops followed by
+      // a concat vector. We must adjust the outputs from the partial horizontal
+      // matching calls above to account for undefined vector halves.
+      SDValue V0 = InVec0.isUndef() ? InVec2 : InVec0;
+      SDValue V1 = InVec1.isUndef() ? InVec3 : InVec1;
+      assert((!V0.isUndef() || !V1.isUndef()) && "Horizontal-op of undefs?");
       bool isUndefLO = NumUndefsLO == Half;
       bool isUndefHI = NumUndefsHI == Half;
-      return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
-                                   isUndefLO, isUndefHI);
+      return ExpandHorizontalBinOp(V0, V1, DL, DAG, X86Opcode, false, isUndefLO,
+                                   isUndefHI);
     }
   }
 

Modified: llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll?rev=350830&r1=350829&r2=350830&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll (original)
+++ llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll Thu Jan 10 07:27:23 2019
@@ -101,6 +101,10 @@ define <8 x i32> @PR40243_alt(<8 x i32>
 ;
 ; AVX1-LABEL: PR40243_alt:
 ; AVX1:       # %bb.0:
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: PR40243_alt: