AVX-512: Fixed a crash during legalization of <3 x i8> type

The compiler fails with assertion during legalization of SETCC for <3 x i8> operands.
The result is extended to <4 x i8> and then truncated <4 x i1>. It does not happen on AVX2, because the final result of SETCC is <4 x i32>.

Differential Revision: https://reviews.llvm.org/D34503

llvm-svn: 306242
This commit is contained in:
Elena Demikhovsky 2017-06-25 13:36:20 +00:00
parent 70f7512add
commit 72f991cded
2 changed files with 32 additions and 2 deletions

View File

@ -615,9 +615,8 @@ SDValue DAGTypeLegalizer::PromoteIntRes_SETCC(SDNode *N) {
SDValue SetCC = DAG.getNode(N->getOpcode(), dl, SVT, LHS, RHS,
N->getOperand(2));
assert(NVT.bitsLE(SVT) && "Integer type overpromoted?");
// Convert to the expected type.
return DAG.getNode(ISD::TRUNCATE, dl, NVT, SetCC);
return DAG.getSExtOrTrunc(SetCC, dl, NVT);
}
SDValue DAGTypeLegalizer::PromoteIntRes_SHL(SDNode *N) {

View File

@ -0,0 +1,31 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mcpu=skx -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
; This test crashed during type legalization of SETCC result type.
define <3 x i8 > @foo(<3 x i8>%x, <3 x i8>%a, <3 x i8>%b) {
; CHECK-LABEL: foo:
; CHECK: # BB#0:
; CHECK-NEXT: vmovd %edi, %xmm0
; CHECK-NEXT: vpinsrd $1, %esi, %xmm0, %xmm0
; CHECK-NEXT: vpinsrd $2, %edx, %xmm0, %xmm0
; CHECK-NEXT: vpslld $24, %xmm0, %xmm0
; CHECK-NEXT: vpsrad $24, %xmm0, %xmm0
; CHECK-NEXT: vmovd %ecx, %xmm1
; CHECK-NEXT: vpinsrd $1, %r8d, %xmm1, %xmm1
; CHECK-NEXT: vpinsrd $2, %r9d, %xmm1, %xmm1
; CHECK-NEXT: vpslld $24, %xmm1, %xmm1
; CHECK-NEXT: vpsrad $24, %xmm1, %xmm1
; CHECK-NEXT: vpcmpgtd %xmm0, %xmm1, %k0
; CHECK-NEXT: vpmovm2d %k0, %xmm0
; CHECK-NEXT: vpextrb $0, %xmm0, %eax
; CHECK-NEXT: vpextrb $4, %xmm0, %edx
; CHECK-NEXT: vpextrb $8, %xmm0, %ecx
; CHECK-NEXT: # kill: %AL<def> %AL<kill> %EAX<kill>
; CHECK-NEXT: # kill: %DL<def> %DL<kill> %EDX<kill>
; CHECK-NEXT: # kill: %CL<def> %CL<kill> %ECX<kill>
; CHECK-NEXT: retq
%cmp.i = icmp slt <3 x i8> %x, %a
%res = sext <3 x i1> %cmp.i to <3 x i8>
ret <3 x i8> %res
}