From ac84850ea620df7f799c38f5a4f2c788eabbea78 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Thu, 27 Jul 2017 18:15:54 +0000
Subject: [PATCH] [SelectionDAG] Improve DAGTypeLegalizer::convertMask
 assertion (PR33960)

Improve DAGTypeLegalizer::convertMask's isSETCCorConvertedSETCC assertion to properly check for any mixture of SETCC or BUILD_VECTOR of constants, or a logical mask op of them.

llvm-svn: 309302
---
 .../CodeGen/SelectionDAG/LegalizeVectorTypes.cpp   | 21 +++++-------
 llvm/test/CodeGen/X86/pr33960.ll                   | 39 ++++++++++++++++++++++
 2 files changed, 48 insertions(+), 12 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/pr33960.ll

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index b926176..dc844fe 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -2965,7 +2965,12 @@ static inline bool isSETCCorConvertedSETCC(SDValue N) {
   else if (N.getOpcode() == ISD::SIGN_EXTEND)
     N = N.getOperand(0);
 
-  return (N.getOpcode() == ISD::SETCC);
+  if (isLogicalMaskOp(N.getOpcode()))
+    return isSETCCorConvertedSETCC(N.getOperand(0)) &&
+           isSETCCorConvertedSETCC(N.getOperand(1));
+
+  return (N.getOpcode() == ISD::SETCC ||
+          ISD::isBuildVectorOfConstantSDNodes(N.getNode()));
 }
 #endif
 
@@ -2973,28 +2978,20 @@ static inline bool isSETCCorConvertedSETCC(SDValue N) {
 // to ToMaskVT if needed with vector extension or truncation.
 SDValue DAGTypeLegalizer::convertMask(SDValue InMask, EVT MaskVT,
                                       EVT ToMaskVT) {
-  LLVMContext &Ctx = *DAG.getContext();
-
   // Currently a SETCC or a AND/OR/XOR with two SETCCs are handled.
-  unsigned InMaskOpc = InMask->getOpcode();
-
   // FIXME: This code seems to be too restrictive, we might consider
   // generalizing it or dropping it.
-  assert((InMaskOpc == ISD::SETCC ||
-          ISD::isBuildVectorOfConstantSDNodes(InMask.getNode()) ||
-          (isLogicalMaskOp(InMaskOpc) &&
-           isSETCCorConvertedSETCC(InMask->getOperand(0)) &&
-           isSETCCorConvertedSETCC(InMask->getOperand(1)))) &&
-         "Unexpected mask argument.");
+  assert(isSETCCorConvertedSETCC(InMask) && "Unexpected mask argument.");
 
   // Make a new Mask node, with a legal result VT.
   SmallVector<SDValue, 4> Ops;
   for (unsigned i = 0, e = InMask->getNumOperands(); i < e; ++i)
     Ops.push_back(InMask->getOperand(i));
-  SDValue Mask = DAG.getNode(InMaskOpc, SDLoc(InMask), MaskVT, Ops);
+  SDValue Mask = DAG.getNode(InMask->getOpcode(), SDLoc(InMask), MaskVT, Ops);
 
   // If MaskVT has smaller or bigger elements than ToMaskVT, a vector sign
   // extend or truncate is needed.
+  LLVMContext &Ctx = *DAG.getContext();
   unsigned MaskScalarBits = MaskVT.getScalarSizeInBits();
   unsigned ToMaskScalBits = ToMaskVT.getScalarSizeInBits();
   if (MaskScalarBits < ToMaskScalBits) {
diff --git a/llvm/test/CodeGen/X86/pr33960.ll b/llvm/test/CodeGen/X86/pr33960.ll
new file mode 100644
index 0000000..fb9236d
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr33960.ll
@@ -0,0 +1,39 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx | FileCheck %s --check-prefix=X86
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefix=X64
+
+@b = external local_unnamed_addr global i32, align 4
+
+define void @PR33960() {
+; X86-LABEL: PR33960:
+; X86:       # BB#0: # %entry
+; X86-NEXT:    movl $0, b
+; X86-NEXT:    retl
+;
+; X64-LABEL: PR33960:
+; X64:       # BB#0: # %entry
+; X64-NEXT:    movl $0, {{.*}}(%rip)
+; X64-NEXT:    retq
+entry:
+  %tmp = insertelement <4 x i32> <i32 undef, i32 -7, i32 -3, i32 undef>, i32 -2, i32 3
+  %predphi26 = insertelement <4 x i32> %tmp, i32 -7, i32 0
+  %tmp1 = trunc <4 x i32> %predphi26 to <4 x i16>
+  %tmp2 = icmp eq <4 x i16> %tmp1, zeroinitializer
+  %tmp3 = icmp eq <4 x i32> undef, zeroinitializer
+  %tmp4 = and <4 x i1> %tmp2, %tmp3
+  %predphi17 = select <4 x i1> %tmp4, <4 x i32> undef, <4 x i32> zeroinitializer
+  %tmp5 = shl <4 x i32> %predphi17, <i32 16, i32 16, i32 16, i32 16>
+  %tmp6 = ashr exact <4 x i32> %tmp5, <i32 16, i32 16, i32 16, i32 16>
+  %tmp7 = or <4 x i32> %tmp6, undef
+  %tmp8 = or <4 x i32> undef, %tmp7
+  %tmp9 = or <4 x i32> undef, %tmp8
+  %tmp10 = or <4 x i32> undef, %tmp9
+  %tmp11 = or <4 x i32> undef, %tmp10
+  %tmp12 = or <4 x i32> undef, %tmp11
+  %bin.rdx = or <4 x i32> %tmp12, undef
+  %bin.rdx19 = or <4 x i32> %bin.rdx, undef
+  %tmp13 = extractelement <4 x i32> %bin.rdx19, i32 0
+  %or = or i32 0, %tmp13
+  store i32 %or, i32* @b, align 4
+  ret void
+}
-- 
2.7.4