[llvm] r322574 - [X86][MMX] Accept UNDEF upper bits for MOVD GR32->MMX
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 16 09:01:32 PST 2018
Author: rksimon
Date: Tue Jan 16 09:01:31 2018
New Revision: 322574
URL: http://llvm.org/viewvc/llvm-project?rev=322574&view=rev
Log:
[X86][MMX] Accept UNDEF upper bits for MOVD GR32->MMX
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/mmx-schedule.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=322574&r1=322573&r2=322574&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Tue Jan 16 09:01:31 2018
@@ -30424,10 +30424,11 @@ static SDValue combineBitcast(SDNode *N,
}
// Detect bitcasts between i32 to x86mmx low word.
- if (N0.getOpcode() == ISD::BUILD_VECTOR && SrcVT == MVT::v2i32 &&
- isNullConstant(N0.getOperand(1))) {
+ if (N0.getOpcode() == ISD::BUILD_VECTOR && SrcVT == MVT::v2i32) {
SDValue N00 = N0.getOperand(0);
- if (N00.getValueType() == MVT::i32)
+ SDValue N01 = N0.getOperand(1);
+ if (N00.getValueType() == MVT::i32 &&
+ (N01.getOpcode() == ISD::UNDEF || isNullConstant(N01)))
return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
}
Modified: llvm/trunk/test/CodeGen/X86/mmx-schedule.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/mmx-schedule.ll?rev=322574&r1=322573&r2=322574&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/mmx-schedule.ll (original)
+++ llvm/trunk/test/CodeGen/X86/mmx-schedule.ll Tue Jan 16 09:01:31 2018
@@ -636,26 +636,20 @@ declare void @llvm.x86.mmx.maskmovq(x86_
define i32 @test_movd(x86_mmx %a0, i32 %a1, i32 *%a2) {
; GENERIC-LABEL: test_movd:
; GENERIC: # %bb.0:
-; GENERIC-NEXT: movd %edi, %xmm0 # sched: [1:1.00]
-; GENERIC-NEXT: movq %xmm0, -{{[0-9]+}}(%rsp) # sched: [5:1.00]
-; GENERIC-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [4:0.50]
-; GENERIC-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [6:0.50]
-; GENERIC-NEXT: movlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [5:1.00]
-; GENERIC-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [7:1.00]
-; GENERIC-NEXT: paddd %mm1, %mm0 # sched: [3:1.00]
-; GENERIC-NEXT: movd %mm1, %ecx # sched: [1:0.33]
+; GENERIC-NEXT: movd %edi, %mm1 # sched: [1:0.33]
+; GENERIC-NEXT: movd (%rsi), %mm2 # sched: [4:0.50]
+; GENERIC-NEXT: paddd %mm1, %mm2 # sched: [3:1.00]
+; GENERIC-NEXT: paddd %mm2, %mm0 # sched: [3:1.00]
+; GENERIC-NEXT: movd %mm2, %ecx # sched: [1:0.33]
; GENERIC-NEXT: movd %mm0, %eax # sched: [1:0.33]
; GENERIC-NEXT: movl %ecx, (%rsi) # sched: [5:1.00]
; GENERIC-NEXT: retq # sched: [1:1.00]
;
; ATOM-LABEL: test_movd:
; ATOM: # %bb.0:
-; ATOM-NEXT: movd %edi, %xmm0 # sched: [1:1.00]
-; ATOM-NEXT: movq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; ATOM-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [1:1.00]
-; ATOM-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [1:1.00]
-; ATOM-NEXT: movlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; ATOM-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [1:1.00]
+; ATOM-NEXT: movd (%rsi), %mm1 # sched: [1:1.00]
+; ATOM-NEXT: movd %edi, %mm2 # sched: [1:1.00]
+; ATOM-NEXT: paddd %mm2, %mm1 # sched: [1:0.50]
; ATOM-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
; ATOM-NEXT: movd %mm1, %ecx # sched: [3:3.00]
; ATOM-NEXT: movd %mm0, %eax # sched: [3:3.00]
@@ -664,112 +658,88 @@ define i32 @test_movd(x86_mmx %a0, i32 %
;
; SLM-LABEL: test_movd:
; SLM: # %bb.0:
-; SLM-NEXT: movd %edi, %xmm0 # sched: [1:0.50]
-; SLM-NEXT: movq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; SLM-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [3:1.00]
-; SLM-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [3:1.00]
-; SLM-NEXT: movlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; SLM-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [4:1.00]
-; SLM-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; SLM-NEXT: movd %mm1, %ecx # sched: [1:0.50]
+; SLM-NEXT: movd (%rsi), %mm2 # sched: [3:1.00]
+; SLM-NEXT: movd %edi, %mm1 # sched: [1:0.50]
+; SLM-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; SLM-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; SLM-NEXT: movd %mm2, %ecx # sched: [1:0.50]
; SLM-NEXT: movd %mm0, %eax # sched: [1:0.50]
; SLM-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; SLM-NEXT: retq # sched: [4:1.00]
;
; SANDY-LABEL: test_movd:
; SANDY: # %bb.0:
-; SANDY-NEXT: vmovd %edi, %xmm0 # sched: [1:1.00]
-; SANDY-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [5:1.00]
-; SANDY-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [4:0.50]
-; SANDY-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [6:0.50]
-; SANDY-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [5:1.00]
-; SANDY-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [7:1.00]
-; SANDY-NEXT: paddd %mm1, %mm0 # sched: [3:1.00]
-; SANDY-NEXT: movd %mm1, %ecx # sched: [1:0.33]
+; SANDY-NEXT: movd %edi, %mm1 # sched: [1:0.33]
+; SANDY-NEXT: movd (%rsi), %mm2 # sched: [4:0.50]
+; SANDY-NEXT: paddd %mm1, %mm2 # sched: [3:1.00]
+; SANDY-NEXT: paddd %mm2, %mm0 # sched: [3:1.00]
+; SANDY-NEXT: movd %mm2, %ecx # sched: [1:0.33]
; SANDY-NEXT: movd %mm0, %eax # sched: [1:0.33]
; SANDY-NEXT: movl %ecx, (%rsi) # sched: [5:1.00]
; SANDY-NEXT: retq # sched: [1:1.00]
;
; HASWELL-LABEL: test_movd:
; HASWELL: # %bb.0:
-; HASWELL-NEXT: vmovd %edi, %xmm0 # sched: [1:1.00]
-; HASWELL-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; HASWELL-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:0.50]
-; HASWELL-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50]
-; HASWELL-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; HASWELL-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:0.50]
-; HASWELL-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; HASWELL-NEXT: movd %mm1, %ecx # sched: [1:1.00]
+; HASWELL-NEXT: movd %edi, %mm1 # sched: [1:1.00]
+; HASWELL-NEXT: movd (%rsi), %mm2 # sched: [5:0.50]
+; HASWELL-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; HASWELL-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; HASWELL-NEXT: movd %mm2, %ecx # sched: [1:1.00]
; HASWELL-NEXT: movd %mm0, %eax # sched: [1:1.00]
; HASWELL-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; HASWELL-NEXT: retq # sched: [7:1.00]
;
; BROADWELL-LABEL: test_movd:
; BROADWELL: # %bb.0:
-; BROADWELL-NEXT: vmovd %edi, %xmm0 # sched: [1:1.00]
-; BROADWELL-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; BROADWELL-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:0.50]
-; BROADWELL-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50]
-; BROADWELL-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; BROADWELL-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:0.50]
-; BROADWELL-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; BROADWELL-NEXT: movd %mm1, %ecx # sched: [1:1.00]
+; BROADWELL-NEXT: movd %edi, %mm1 # sched: [1:1.00]
+; BROADWELL-NEXT: movd (%rsi), %mm2 # sched: [5:0.50]
+; BROADWELL-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; BROADWELL-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; BROADWELL-NEXT: movd %mm2, %ecx # sched: [1:1.00]
; BROADWELL-NEXT: movd %mm0, %eax # sched: [1:1.00]
; BROADWELL-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; BROADWELL-NEXT: retq # sched: [7:1.00]
;
; SKYLAKE-LABEL: test_movd:
; SKYLAKE: # %bb.0:
-; SKYLAKE-NEXT: vmovd %edi, %xmm0 # sched: [1:1.00]
-; SKYLAKE-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; SKYLAKE-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:0.50]
-; SKYLAKE-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50]
-; SKYLAKE-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; SKYLAKE-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:0.50]
-; SKYLAKE-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; SKYLAKE-NEXT: movd %mm1, %ecx # sched: [2:1.00]
+; SKYLAKE-NEXT: movd %edi, %mm1 # sched: [1:1.00]
+; SKYLAKE-NEXT: movd (%rsi), %mm2 # sched: [5:0.50]
+; SKYLAKE-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; SKYLAKE-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; SKYLAKE-NEXT: movd %mm2, %ecx # sched: [2:1.00]
; SKYLAKE-NEXT: movd %mm0, %eax # sched: [2:1.00]
; SKYLAKE-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; SKYLAKE-NEXT: retq # sched: [7:1.00]
;
; SKX-LABEL: test_movd:
; SKX: # %bb.0:
-; SKX-NEXT: vmovd %edi, %xmm0 # sched: [1:1.00]
-; SKX-NEXT: vpmovqd %xmm0, -{{[0-9]+}}(%rsp) # sched: [4:1.00]
-; SKX-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:0.50]
-; SKX-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:0.50]
-; SKX-NEXT: vpmovqd %xmm0, -{{[0-9]+}}(%rsp) # sched: [4:1.00]
-; SKX-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:0.50]
-; SKX-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; SKX-NEXT: movd %mm1, %ecx # sched: [2:1.00]
+; SKX-NEXT: movd %edi, %mm1 # sched: [1:1.00]
+; SKX-NEXT: movd (%rsi), %mm2 # sched: [5:0.50]
+; SKX-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; SKX-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; SKX-NEXT: movd %mm2, %ecx # sched: [2:1.00]
; SKX-NEXT: movd %mm0, %eax # sched: [2:1.00]
; SKX-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; SKX-NEXT: retq # sched: [7:1.00]
;
; BTVER2-LABEL: test_movd:
; BTVER2: # %bb.0:
-; BTVER2-NEXT: vmovd %edi, %xmm0 # sched: [1:0.50]
-; BTVER2-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; BTVER2-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [5:1.00]
-; BTVER2-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [5:1.00]
-; BTVER2-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
-; BTVER2-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [6:1.00]
-; BTVER2-NEXT: paddd %mm1, %mm0 # sched: [1:0.50]
-; BTVER2-NEXT: movd %mm1, %ecx # sched: [1:0.50]
+; BTVER2-NEXT: movd (%rsi), %mm2 # sched: [5:1.00]
+; BTVER2-NEXT: movd %edi, %mm1 # sched: [1:0.50]
+; BTVER2-NEXT: paddd %mm1, %mm2 # sched: [1:0.50]
+; BTVER2-NEXT: paddd %mm2, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT: movd %mm2, %ecx # sched: [1:0.50]
; BTVER2-NEXT: movd %mm0, %eax # sched: [1:0.50]
; BTVER2-NEXT: movl %ecx, (%rsi) # sched: [1:1.00]
; BTVER2-NEXT: retq # sched: [4:1.00]
;
; ZNVER1-LABEL: test_movd:
; ZNVER1: # %bb.0:
-; ZNVER1-NEXT: vmovd %edi, %xmm0 # sched: [3:1.00]
-; ZNVER1-NEXT: vmovq %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:0.50]
-; ZNVER1-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [8:0.50]
-; ZNVER1-NEXT: movq -{{[0-9]+}}(%rsp), %mm1 # sched: [8:0.50]
-; ZNVER1-NEXT: vmovlps %xmm0, -{{[0-9]+}}(%rsp) # sched: [1:0.50]
-; ZNVER1-NEXT: paddd -{{[0-9]+}}(%rsp), %mm1 # sched: [8:0.50]
-; ZNVER1-NEXT: paddd %mm1, %mm0 # sched: [1:0.25]
-; ZNVER1-NEXT: movd %mm1, %ecx # sched: [2:1.00]
+; ZNVER1-NEXT: movd (%rsi), %mm2 # sched: [8:0.50]
+; ZNVER1-NEXT: movd %edi, %mm1 # sched: [3:1.00]
+; ZNVER1-NEXT: paddd %mm1, %mm2 # sched: [1:0.25]
+; ZNVER1-NEXT: paddd %mm2, %mm0 # sched: [1:0.25]
+; ZNVER1-NEXT: movd %mm2, %ecx # sched: [2:1.00]
; ZNVER1-NEXT: movd %mm0, %eax # sched: [2:1.00]
; ZNVER1-NEXT: movl %ecx, (%rsi) # sched: [1:0.50]
; ZNVER1-NEXT: retq # sched: [1:0.50]
More information about the llvm-commits
mailing list