maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

X86: Shave off one shuffle from the pcmpeqq sequence for SSE2 by making use of and commutativity. 

llvm-svn: 171064
This commit is contained in:
Benjamin Kramer 2012-12-25 13:09:08 +00:00
parent df4af41b9b
commit 81b5a8fd2e
2 changed files with 6 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -9173,7 +9173,7 @@ SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
return SDValue(); return SDValue();
if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) { if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
// If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
// pcmpeqd + 2 shuffles + pand. // pcmpeqd + pshufd + pand.
assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!"); assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
// First cast everything to the right type, // First cast everything to the right type,
@ -9184,11 +9184,9 @@ SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1); SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
// Make sure the lower and upper halves are both all-ones. // Make sure the lower and upper halves are both all-ones.
const int Mask1[] = { 0, 0, 2, 2 }; const int Mask[] = { 1, 0, 3, 2 };
SDValue S1 = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask1); SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
const int Mask2[] = { 1, 1, 3, 3 }; Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
SDValue S2 = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask2);
Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, S1, S2);
if (Invert) if (Invert)
Result = DAG.getNOT(dl, Result, MVT::v4i32); Result = DAG.getNOT(dl, Result, MVT::v4i32);

6
llvm/test/CodeGen/X86/vec_compare.ll
View File

@ -45,8 +45,7 @@ define <4 x i32> @test4(<4 x i32> %A, <4 x i32> %B) nounwind {
define <2 x i64> @test5(<2 x i64> %A, <2 x i64> %B) nounwind { define <2 x i64> @test5(<2 x i64> %A, <2 x i64> %B) nounwind {
; CHECK: test5: ; CHECK: test5:
; CHECK: pcmpeqd ; CHECK: pcmpeqd
; CHECK: pshufd $-11 ; CHECK: pshufd $-79
; CHECK: pshufd $-96
; CHECK: pand ; CHECK: pand
; CHECK: ret ; CHECK: ret
%C = icmp eq <2 x i64> %A, %B %C = icmp eq <2 x i64> %A, %B
@ -57,8 +56,7 @@ define <2 x i64> @test5(<2 x i64> %A, <2 x i64> %B) nounwind {
define <2 x i64> @test6(<2 x i64> %A, <2 x i64> %B) nounwind { define <2 x i64> @test6(<2 x i64> %A, <2 x i64> %B) nounwind {
; CHECK: test6: ; CHECK: test6:
; CHECK: pcmpeqd ; CHECK: pcmpeqd
; CHECK: pshufd $-11 ; CHECK: pshufd $-79
; CHECK: pshufd $-96
; CHECK: pand ; CHECK: pand
; CHECK: pcmpeqd ; CHECK: pcmpeqd
; CHECK: pxor ; CHECK: pxor