Enable execution dependency fix pass for YMM registers when AVX2 is enabled. Add AVX2 logical operations to list of replaceable instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144179 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Craig Topper 2011-11-09 09:37:21 +00:00
parent 94d80da4a0
commit b80ada98c5
3 changed files with 58 additions and 4 deletions

View File

@ -3367,6 +3367,14 @@ static const unsigned ReplaceableInstrs[][3] = {
{ X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
{ X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
{ X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr },
{ X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
{ X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
{ X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
{ X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
{ X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
{ X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
{ X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
{ X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
};
// FIXME: Some shuffle and unpack instructions have equivalents in different

View File

@ -135,10 +135,18 @@ bool X86TargetMachine::addPostRegAlloc(PassManagerBase &PM,
bool X86TargetMachine::addPreEmitPass(PassManagerBase &PM,
CodeGenOpt::Level OptLevel) {
bool ShouldPrint = false;
if (OptLevel != CodeGenOpt::None &&
(Subtarget.hasSSE2() || Subtarget.hasAVX())) {
PM.add(createExecutionDependencyFixPass(&X86::VR128RegClass));
ShouldPrint = true;
if (OptLevel != CodeGenOpt::None) {
if (Subtarget.hasXMMInt()) {
PM.add(createExecutionDependencyFixPass(&X86::VR128RegClass));
ShouldPrint = true;
}
if (Subtarget.hasAVX2()) {
// FIXME this should be turned on for just AVX, but the pass doesn't check
// that instructions are valid before replacing them and there are AVX2
// integer instructions in the table.
PM.add(createExecutionDependencyFixPass(&X86::VR256RegClass));
ShouldPrint = true;
}
}
if (Subtarget.hasAVX() && UseVZeroUpper) {

View File

@ -0,0 +1,38 @@
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core-avx2 -mattr=+avx2 | FileCheck %s
; CHECK: vpandn %ymm
define <4 x i64> @vpandn(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
entry:
; Force the execution domain with an add.
%a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
%y = xor <4 x i64> %a2, <i64 -1, i64 -1, i64 -1, i64 -1>
%x = and <4 x i64> %a, %y
ret <4 x i64> %x
}
; CHECK: vpand %ymm
define <4 x i64> @vpand(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
entry:
; Force the execution domain with an add.
%a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
%x = and <4 x i64> %a2, %b
ret <4 x i64> %x
}
; CHECK: vpor %ymm
define <4 x i64> @vpor(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
entry:
; Force the execution domain with an add.
%a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
%x = or <4 x i64> %a2, %b
ret <4 x i64> %x
}
; CHECK: vpxor %ymm
define <4 x i64> @vpxor(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
entry:
; Force the execution domain with an add.
%a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
%x = xor <4 x i64> %a2, %b
ret <4 x i64> %x
}