[llvm] e4c454b - [X86] Add a few more shuffles to hasUndefRegUpdate.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Sun May 10 12:32:50 PDT 2020
Author: Craig Topper
Date: 2020-05-10T12:30:16-07:00
New Revision: e4c454b065be3b884b09fe7c0043f4445bdbd585
URL: https://github.com/llvm/llvm-project/commit/e4c454b065be3b884b09fe7c0043f4445bdbd585
DIFF: https://github.com/llvm/llvm-project/commit/e4c454b065be3b884b09fe7c0043f4445bdbd585.diff
LOG: [X86] Add a few more shuffles to hasUndefRegUpdate.
Mostly found by asserting on tests that have undef operands. I'm
sure this isn't an exhaustive list.
Added:
Modified:
llvm/lib/Target/X86/X86InstrInfo.cpp
llvm/test/CodeGen/X86/mmx-build-vector.ll
llvm/test/CodeGen/X86/vector-idiv-v2i32.ll
llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index e761183974c0..502c138089b9 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -4838,6 +4838,13 @@ static bool hasUndefRegUpdate(unsigned Opcode, unsigned OpNum,
bool ForLoadFold = false) {
// Set the OpNum parameter to the first source operand.
switch (Opcode) {
+ case X86::MMX_PUNPCKHBWirr:
+ case X86::MMX_PUNPCKHWDirr:
+ case X86::MMX_PUNPCKHDQirr:
+ case X86::MMX_PUNPCKLBWirr:
+ case X86::MMX_PUNPCKLWDirr:
+ case X86::MMX_PUNPCKLDQirr:
+ case X86::MOVHLPSrr:
case X86::PACKSSWBrr:
case X86::PACKUSWBrr:
case X86::PACKSSDWrr:
@@ -4850,6 +4857,8 @@ static bool hasUndefRegUpdate(unsigned Opcode, unsigned OpNum,
case X86::PUNPCKLDQrr:
case X86::PUNPCKHQDQrr:
case X86::PUNPCKLQDQrr:
+ case X86::SHUFPDrri:
+ case X86::SHUFPSrri:
// These instructions are sometimes used with an undef first or second
// source. Return true here so BreakFalseDeps will assign this source to the
// same register as the first source to avoid a false dependency.
@@ -4857,6 +4866,8 @@ static bool hasUndefRegUpdate(unsigned Opcode, unsigned OpNum,
// VEX counterparts.
return OpNum == 2 && !ForLoadFold;
+ case X86::VMOVLHPSrr:
+ case X86::VMOVLHPSZrr:
case X86::VPACKSSWBrr:
case X86::VPACKUSWBrr:
case X86::VPACKSSDWrr:
@@ -4865,6 +4876,16 @@ static bool hasUndefRegUpdate(unsigned Opcode, unsigned OpNum,
case X86::VPACKUSWBZ128rr:
case X86::VPACKSSDWZ128rr:
case X86::VPACKUSDWZ128rr:
+ case X86::VPERM2F128rr:
+ case X86::VPERM2I128rr:
+ case X86::VSHUFF32X4Z256rri:
+ case X86::VSHUFF32X4Zrri:
+ case X86::VSHUFF64X2Z256rri:
+ case X86::VSHUFF64X2Zrri:
+ case X86::VSHUFI32X4Z256rri:
+ case X86::VSHUFI32X4Zrri:
+ case X86::VSHUFI64X2Z256rri:
+ case X86::VSHUFI64X2Zrri:
case X86::VPUNPCKHBWrr:
case X86::VPUNPCKLBWrr:
case X86::VPUNPCKHBWYrr:
diff --git a/llvm/test/CodeGen/X86/mmx-build-vector.ll b/llvm/test/CodeGen/X86/mmx-build-vector.ll
index dace3cdc7b23..2ce69a626126 100644
--- a/llvm/test/CodeGen/X86/mmx-build-vector.ll
+++ b/llvm/test/CodeGen/X86/mmx-build-vector.ll
@@ -455,7 +455,7 @@ define void @build_v8i8_0u2345z7(x86_mmx *%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i
; X64-NEXT: movd %ecx, %mm2
; X64-NEXT: punpcklbw %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1],mm2[2],mm1[2],mm2[3],mm1[3]
; X64-NEXT: movd %esi, %mm1
-; X64-NEXT: punpcklbw %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1],mm1[2],mm0[2],mm1[3],mm0[3]
+; X64-NEXT: punpcklbw %mm1, %mm1 # mm1 = mm1[0,0,1,1,2,2,3,3]
; X64-NEXT: punpcklwd %mm2, %mm1 # mm1 = mm1[0],mm2[0],mm1[1],mm2[1]
; X64-NEXT: punpckldq %mm0, %mm1 # mm1 = mm1[0],mm0[0]
; X64-NEXT: paddd %mm1, %mm1
@@ -488,7 +488,7 @@ define void @build_v8i8_0123zzzu(x86_mmx *%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i
; X86-NEXT: punpcklwd %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1]
; X86-NEXT: pxor %mm0, %mm0
; X86-NEXT: pxor %mm1, %mm1
-; X86-NEXT: punpcklbw %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1],mm1[2],mm0[2],mm1[3],mm0[3]
+; X86-NEXT: punpcklbw %mm1, %mm1 # mm1 = mm1[0,0,1,1,2,2,3,3]
; X86-NEXT: punpcklbw %mm0, %mm0 # mm0 = mm0[0,0,1,1,2,2,3,3]
; X86-NEXT: punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1]
; X86-NEXT: punpckldq %mm0, %mm2 # mm2 = mm2[0],mm0[0]
@@ -507,7 +507,7 @@ define void @build_v8i8_0123zzzu(x86_mmx *%p0, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i
; X64-NEXT: punpcklwd %mm1, %mm2 # mm2 = mm2[0],mm1[0],mm2[1],mm1[1]
; X64-NEXT: pxor %mm0, %mm0
; X64-NEXT: pxor %mm1, %mm1
-; X64-NEXT: punpcklbw %mm0, %mm1 # mm1 = mm1[0],mm0[0],mm1[1],mm0[1],mm1[2],mm0[2],mm1[3],mm0[3]
+; X64-NEXT: punpcklbw %mm1, %mm1 # mm1 = mm1[0,0,1,1,2,2,3,3]
; X64-NEXT: punpcklbw %mm0, %mm0 # mm0 = mm0[0,0,1,1,2,2,3,3]
; X64-NEXT: punpcklwd %mm1, %mm0 # mm0 = mm0[0],mm1[0],mm0[1],mm1[1]
; X64-NEXT: punpckldq %mm0, %mm2 # mm2 = mm2[0],mm0[0]
diff --git a/llvm/test/CodeGen/X86/vector-idiv-v2i32.ll b/llvm/test/CodeGen/X86/vector-idiv-v2i32.ll
index b3accf80744f..032c9e293b0a 100644
--- a/llvm/test/CodeGen/X86/vector-idiv-v2i32.ll
+++ b/llvm/test/CodeGen/X86/vector-idiv-v2i32.ll
@@ -31,7 +31,7 @@ define void @test_udiv7_v2i32(<2 x i32>* %x, <2 x i32>* %y) nounwind {
; X86-NEXT: pmuludq %xmm1, %xmm2
; X86-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; X86-NEXT: movdqa %xmm0, %xmm3
-; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
; X86-NEXT: pmuludq %xmm1, %xmm3
; X86-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,3,2,3]
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
@@ -81,7 +81,7 @@ define void @test_urem7_v2i32(<2 x i32>* %x, <2 x i32>* %y) nounwind {
; X86-NEXT: pmuludq %xmm1, %xmm2
; X86-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; X86-NEXT: movdqa %xmm0, %xmm3
-; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
; X86-NEXT: pmuludq %xmm1, %xmm3
; X86-NEXT: pshufd {{.*#+}} xmm1 = xmm3[1,3,2,3]
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
@@ -137,7 +137,7 @@ define void @test_sdiv7_v2i32(<2 x i32>* %x, <2 x i32>* %y) nounwind {
; X86-NEXT: pmuludq %xmm1, %xmm2
; X86-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; X86-NEXT: movdqa %xmm0, %xmm3
-; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
; X86-NEXT: pmuludq %xmm1, %xmm3
; X86-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
@@ -198,7 +198,7 @@ define void @test_srem7_v2i32(<2 x i32>* %x, <2 x i32>* %y) nounwind {
; X86-NEXT: pmuludq %xmm1, %xmm2
; X86-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3]
; X86-NEXT: movdqa %xmm0, %xmm3
-; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
; X86-NEXT: pmuludq %xmm1, %xmm3
; X86-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,3,2,3]
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
@@ -358,7 +358,7 @@ define void @test_udiv_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
; X86-NEXT: movd %eax, %xmm2
; X86-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X86-NEXT: movd %xmm0, %eax
-; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1,2,3]
; X86-NEXT: movd %xmm1, %esi
; X86-NEXT: xorl %edx, %edx
; X86-NEXT: divl %esi
@@ -412,7 +412,7 @@ define void @test_urem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
; X86-NEXT: movd %edx, %xmm2
; X86-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X86-NEXT: movd %xmm0, %eax
-; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1,2,3]
; X86-NEXT: movd %xmm1, %esi
; X86-NEXT: xorl %edx, %edx
; X86-NEXT: divl %esi
@@ -465,7 +465,7 @@ define void @test_sdiv_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
; X86-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X86-NEXT: movd %xmm0, %eax
; X86-NEXT: movd %xmm1, %edi
-; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1,2,3]
; X86-NEXT: movd %xmm1, %ebx
; X86-NEXT: cltd
; X86-NEXT: idivl %ebx
@@ -524,7 +524,7 @@ define void @test_srem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
; X86-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
; X86-NEXT: movd %xmm0, %eax
; X86-NEXT: movd %xmm1, %edi
-; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[2,3]
+; X86-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1,2,3]
; X86-NEXT: movd %xmm1, %ebx
; X86-NEXT: cltd
; X86-NEXT: idivl %ebx
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll b/llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll
index 1638cb7f4c58..739ea6bae535 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll
@@ -1869,7 +1869,7 @@ define <8 x double> @shuffle_v8f64_2301uuuu(<8 x double> %a0, <8 x double> %a1)
define <8 x double> @shuffle_v8f64_uuu2301(<8 x double> %a0, <8 x double> %a1) {
; ALL-LABEL: shuffle_v8f64_uuu2301:
; ALL: # %bb.0:
-; ALL-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,0,1],zmm1[2,3,0,1]
+; ALL-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1,2,3,0,1]
; ALL-NEXT: ret{{[l|q]}}
%1 = shufflevector <8 x double> %a1, <8 x double> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 2, i32 3, i32 0, i32 1>
ret <8 x double> %1
More information about the llvm-commits
mailing list