From 8b960d22ad4bc724b89e17f9ab0be9e6a468dd53 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Sat, 12 Sep 2015 19:47:50 +0000
Subject: [PATCH] [x86] enable machine combiner reassociations for 128-bit
 vector logical integer insts (2nd try)

The changes in:
test/CodeGen/X86/machine-cp.ll
are just due to scheduling differences after some logic instructions were reassociated.

llvm-svn: 247516
---
 llvm/lib/Target/X86/X86InstrInfo.cpp              |  6 ++
 llvm/test/CodeGen/X86/machine-combiner-int-vec.ll | 68 +++++++++++++++++++++++
 llvm/test/CodeGen/X86/machine-cp.ll               | 18 +++---
 3 files changed, 81 insertions(+), 11 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/machine-combiner-int-vec.ll

diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 8b88316..446d4bc 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -6408,6 +6408,12 @@ static bool isAssociativeAndCommutative(const MachineInstr &Inst) {
   case X86::IMUL16rr:
   case X86::IMUL32rr:
   case X86::IMUL64rr:
+  case X86::PANDrr:
+  case X86::PORrr:
+  case X86::PXORrr:
+  case X86::VPANDrr:
+  case X86::VPORrr:
+  case X86::VPXORrr:
   // 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/machine-combiner-int-vec.ll b/llvm/test/CodeGen/X86/machine-combiner-int-vec.ll
new file mode 100644
index 0000000..8316f66
--- /dev/null
+++ b/llvm/test/CodeGen/X86/machine-combiner-int-vec.ll
@@ -0,0 +1,68 @@
+; RUN: llc -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=sse < %s | FileCheck %s --check-prefix=SSE
+; RUN: llc -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=avx < %s | FileCheck %s --check-prefix=AVX
+
+; Verify that 128-bit vector logical ops are reassociated.
+
+define <4 x i32> @reassociate_and_v4i32(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, <4 x i32> %x3) {
+; SSE-LABEL: reassociate_and_v4i32:
+; SSE:       # BB#0:
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pand %xmm3, %xmm2
+; SSE-NEXT:    pand %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_and_v4i32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpand %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+
+  %t0 = add <4 x i32> %x0, %x1
+  %t1 = and <4 x i32> %x2, %t0
+  %t2 = and <4 x i32> %x3, %t1
+  ret <4 x i32> %t2
+}
+
+define <4 x i32> @reassociate_or_v4i32(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, <4 x i32> %x3) {
+; SSE-LABEL: reassociate_or_v4i32:
+; SSE:       # BB#0:
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    por %xmm3, %xmm2
+; SSE-NEXT:    por %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_or_v4i32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpor %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+
+  %t0 = add <4 x i32> %x0, %x1
+  %t1 = or <4 x i32> %x2, %t0
+  %t2 = or <4 x i32> %x3, %t1
+  ret <4 x i32> %t2
+}
+
+define <4 x i32> @reassociate_xor_v4i32(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, <4 x i32> %x3) {
+; SSE-LABEL: reassociate_xor_v4i32:
+; SSE:       # BB#0:
+; SSE-NEXT:    paddd %xmm1, %xmm0
+; SSE-NEXT:    pxor %xmm3, %xmm2
+; SSE-NEXT:    pxor %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: reassociate_xor_v4i32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpxor %xmm3, %xmm2, %xmm1
+; AVX-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    retq
+
+  %t0 = add <4 x i32> %x0, %x1
+  %t1 = xor <4 x i32> %x2, %t0
+  %t2 = xor <4 x i32> %x3, %t1
+  ret <4 x i32> %t2
+}
+
diff --git a/llvm/test/CodeGen/X86/machine-cp.ll b/llvm/test/CodeGen/X86/machine-cp.ll
index 768b9ca..d3b305b 100644
--- a/llvm/test/CodeGen/X86/machine-cp.ll
+++ b/llvm/test/CodeGen/X86/machine-cp.ll
@@ -73,22 +73,18 @@ while.end:                                        ; preds = %while.body, %entry
 ; Machine propagation used to delete the first copy as the
 ; first few uses were <undef>.
 ; CHECK-NEXT: movdqa [[SRC]], [[CPY1:%xmm[0-9]+]]
-; CHECK-NEXT: movdqa [[SRC]], [[CPY2:%xmm[0-9]+]]
-; CHECK-NEXT: punpckhbw [[SRC]],
-; Check that CPY1 is not redefined.
-; CHECK-NOT: , [[CPY1]]
-; undef use, we do not care.
-; CHECK: punpcklwd [[CPY1]],
-; Check that CPY1 is not redefined.
-; CHECK-NOT: , [[CPY1]]
+; CHECK: punpcklbw [[CPY1]], [[CPY1]]
+; CHECK-NEXT: punpcklwd [[CPY1]], [[CPY1]]
+; CHECK-NEXT: pslld $31, [[CPY1]]
+; CHECK: movdqa [[SRC]], [[CPY2:%xmm[0-9]+]]
 ; CHECK: punpcklbw [[CPY2]], [[CPY2]]
 ; CHECK-NEXT: punpckhwd [[CPY2]], [[CPY2]]
 ; CHECK-NEXT: pslld $31, [[CPY2]]
+; CHECK: punpckhbw [[SRC]],
 ; Check that CPY1 is not redefined.
 ; CHECK-NOT: , [[CPY1]]
-; CHECK: punpcklbw [[CPY1]], [[CPY1]]
-; CHECK-NEXT: punpcklwd [[CPY1]], [[CPY1]]
-; CHECK-NEXT: pslld $31, [[CPY1]]
+; undef use, we do not care.
+; CHECK: punpcklwd [[CPY1]],
 define <16 x float> @foo(<16 x float> %x) {
 bb:
   %v3 = icmp slt <16 x i32> undef, zeroinitializer
-- 
2.7.4