From: Craig Topper <craig.topper@gmail.com>
Date: Mon, 18 Jul 2016 06:14:50 +0000 (+0000)
Subject: [X86] Add floating point packed logical ops to X86InstrInfo::isAssociativeAndCommutative.
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=ba9b93d7f2badb8b7686f8bf3ca9aa983b016eaf;p=platform%2Fupstream%2Fllvm.git

[X86] Add floating point packed logical ops to X86InstrInfo::isAssociativeAndCommutative.

llvm-svn: 275768
---

diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 22a3951..5209818 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -7643,6 +7643,12 @@ bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
   case X86::PANDrr:
   case X86::PORrr:
   case X86::PXORrr:
+  case X86::ANDPDrr:
+  case X86::ANDPSrr:
+  case X86::ORPDrr:
+  case X86::ORPSrr:
+  case X86::XORPDrr:
+  case X86::XORPSrr:
   case X86::VPANDrr:
   case X86::VPANDYrr:
   case X86::VPANDDZ128rr:
@@ -7667,6 +7673,36 @@ bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const {
   case X86::VPXORQZ128rr:
   case X86::VPXORQZ256rr:
   case X86::VPXORQZrr:
+  case X86::VANDPDrr:
+  case X86::VANDPSrr:
+  case X86::VANDPDYrr:
+  case X86::VANDPSYrr:
+  case X86::VANDPDZ128rr:
+  case X86::VANDPSZ128rr:
+  case X86::VANDPDZ256rr:
+  case X86::VANDPSZ256rr:
+  case X86::VANDPDZrr:
+  case X86::VANDPSZrr:
+  case X86::VORPDrr:
+  case X86::VORPSrr:
+  case X86::VORPDYrr:
+  case X86::VORPSYrr:
+  case X86::VORPDZ128rr:
+  case X86::VORPSZ128rr:
+  case X86::VORPDZ256rr:
+  case X86::VORPSZ256rr:
+  case X86::VORPDZrr:
+  case X86::VORPSZrr:
+  case X86::VXORPDrr:
+  case X86::VXORPSrr:
+  case X86::VXORPDYrr:
+  case X86::VXORPSYrr:
+  case X86::VXORPDZ128rr:
+  case X86::VXORPSZ128rr:
+  case X86::VXORPDZ256rr:
+  case X86::VXORPSZ256rr:
+  case X86::VXORPDZrr:
+  case X86::VXORPSZrr:
   // Normal min/max instructions are not commutative because of NaN and signed
   // zero semantics, but these are. Thus, there's no need to check for global
   // relaxed math; the instructions themselves have the properties we need.
diff --git a/llvm/test/CodeGen/X86/WidenArith.ll b/llvm/test/CodeGen/X86/WidenArith.ll
index cc5fcba..cdd1a28 100644
--- a/llvm/test/CodeGen/X86/WidenArith.ll
+++ b/llvm/test/CodeGen/X86/WidenArith.ll
@@ -9,8 +9,8 @@ define <8 x i32> @test(<8 x float> %a, <8 x float> %b) {
 ; CHECK-NEXT:    vsubps %ymm2, %ymm1, %ymm3
 ; CHECK-NEXT:    vcmpltps %ymm1, %ymm0, %ymm0
 ; CHECK-NEXT:    vcmpltps %ymm3, %ymm2, %ymm1
+; CHECK-NEXT:    vandps {{.*}}(%rip), %ymm1, %ymm1
 ; CHECK-NEXT:    vandps %ymm1, %ymm0, %ymm0
-; CHECK-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm0
 ; CHECK-NEXT:    retq
  %c1 = fadd <8 x float> %a, %b
  %b1 = fmul <8 x float> %b, %a
diff --git a/llvm/test/CodeGen/X86/v8i1-masks.ll b/llvm/test/CodeGen/X86/v8i1-masks.ll
index d5c3150..0135832 100644
--- a/llvm/test/CodeGen/X86/v8i1-masks.ll
+++ b/llvm/test/CodeGen/X86/v8i1-masks.ll
@@ -13,8 +13,8 @@ define void @and_masks(<8 x float>* %a, <8 x float>* %b, <8 x float>* %c) nounwi
 ; X32-NEXT:    vcmpltps %ymm0, %ymm1, %ymm1
 ; X32-NEXT:    vmovups (%eax), %ymm2
 ; X32-NEXT:    vcmpltps %ymm0, %ymm2, %ymm0
+; X32-NEXT:    vandps LCPI0_0, %ymm1, %ymm1
 ; X32-NEXT:    vandps %ymm1, %ymm0, %ymm0
-; X32-NEXT:    vandps LCPI0_0, %ymm0, %ymm0
 ; X32-NEXT:    vmovaps %ymm0, (%eax)
 ; X32-NEXT:    vzeroupper
 ; X32-NEXT:    retl
@@ -26,8 +26,8 @@ define void @and_masks(<8 x float>* %a, <8 x float>* %b, <8 x float>* %c) nounwi
 ; X64-NEXT:    vcmpltps %ymm0, %ymm1, %ymm1
 ; X64-NEXT:    vmovups (%rdx), %ymm2
 ; X64-NEXT:    vcmpltps %ymm0, %ymm2, %ymm0
+; X64-NEXT:    vandps {{.*}}(%rip), %ymm1, %ymm1
 ; X64-NEXT:    vandps %ymm1, %ymm0, %ymm0
-; X64-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm0
 ; X64-NEXT:    vmovaps %ymm0, (%rax)
 ; X64-NEXT:    vzeroupper
 ; X64-NEXT:    retq