From bf83737adbb2380119c7a4cae5fee4c00c9d5f22 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Fri, 21 Aug 2015 21:04:21 +0000
Subject: [PATCH] [x86] enable machine combiner reassociations for 256-bit
 vector min/max

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@245735 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86InstrInfo.cpp      |  4 ++
 test/CodeGen/X86/machine-combiner.ll | 68 ++++++++++++++++++++++++++++
 2 files changed, 72 insertions(+)

diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 9d799d8e780..c3e862b09ed 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -6409,10 +6409,14 @@ static bool isAssociativeAndCommutative(const MachineInstr &Inst) {
   case X86::MINCSSrr:
   case X86::VMAXCPDrr:
   case X86::VMAXCPSrr:
+  case X86::VMAXCPDYrr:
+  case X86::VMAXCPSYrr:
   case X86::VMAXCSDrr:
   case X86::VMAXCSSrr:
   case X86::VMINCPDrr:
   case X86::VMINCPSrr:
+  case X86::VMINCPDYrr:
+  case X86::VMINCPSYrr:
   case X86::VMINCSDrr:
   case X86::VMINCSSrr:
     return true;
diff --git a/test/CodeGen/X86/machine-combiner.ll b/test/CodeGen/X86/machine-combiner.ll
index 371747479e5..b4340b34cc0 100644
--- a/test/CodeGen/X86/machine-combiner.ll
+++ b/test/CodeGen/X86/machine-combiner.ll
@@ -550,3 +550,71 @@ define <2 x double> @reassociate_maxs_v2f64(<2 x double> %x0, <2 x double> %x1,
   ret <2 x double> %sel2
 }
 
+; Verify that AVX 256-bit vector single-precision minimum ops are reassociated.
+
+define <8 x float> @reassociate_mins_v8f32(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, <8 x float> %x3) {
+; AVX-LABEL: reassociate_mins_v8f32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vminps %ymm3, %ymm2, %ymm1
+; AVX-NEXT:    vminps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    retq
+  %t0 = fadd <8 x float> %x0, %x1
+  %cmp1 = fcmp olt <8 x float> %x2, %t0
+  %sel1 = select <8 x i1> %cmp1, <8 x float> %x2, <8 x float> %t0
+  %cmp2 = fcmp olt <8 x float> %x3, %sel1
+  %sel2 = select <8 x i1> %cmp2, <8 x float> %x3, <8 x float> %sel1
+  ret <8 x float> %sel2
+}
+
+; Verify that AVX 256-bit vector single-precision maximum ops are reassociated.
+
+define <8 x float> @reassociate_maxs_v8f32(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, <8 x float> %x3) {
+; AVX-LABEL: reassociate_maxs_v8f32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vmaxps %ymm3, %ymm2, %ymm1
+; AVX-NEXT:    vmaxps %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    retq
+  %t0 = fadd <8 x float> %x0, %x1
+  %cmp1 = fcmp ogt <8 x float> %x2, %t0
+  %sel1 = select <8 x i1> %cmp1, <8 x float> %x2, <8 x float> %t0
+  %cmp2 = fcmp ogt <8 x float> %x3, %sel1
+  %sel2 = select <8 x i1> %cmp2, <8 x float> %x3, <8 x float> %sel1
+  ret <8 x float> %sel2
+}
+
+; Verify that AVX 256-bit vector double-precision minimum ops are reassociated.
+
+define <4 x double> @reassociate_mins_v4f64(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, <4 x double> %x3) {
+; AVX-LABEL: reassociate_mins_v4f64:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vminpd %ymm3, %ymm2, %ymm1
+; AVX-NEXT:    vminpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    retq
+  %t0 = fadd <4 x double> %x0, %x1
+  %cmp1 = fcmp olt <4 x double> %x2, %t0
+  %sel1 = select <4 x i1> %cmp1, <4 x double> %x2, <4 x double> %t0
+  %cmp2 = fcmp olt <4 x double> %x3, %sel1
+  %sel2 = select <4 x i1> %cmp2, <4 x double> %x3, <4 x double> %sel1
+  ret <4 x double> %sel2
+}
+
+; Verify that AVX 256-bit vector double-precision maximum ops are reassociated.
+
+define <4 x double> @reassociate_maxs_v4f64(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, <4 x double> %x3) {
+; AVX-LABEL: reassociate_maxs_v4f64:
+; AVX:       # BB#0:
+; AVX-NEXT:    vaddpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    vmaxpd %ymm3, %ymm2, %ymm1
+; AVX-NEXT:    vmaxpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT:    retq
+  %t0 = fadd <4 x double> %x0, %x1
+  %cmp1 = fcmp ogt <4 x double> %x2, %t0
+  %sel1 = select <4 x i1> %cmp1, <4 x double> %x2, <4 x double> %t0
+  %cmp2 = fcmp ogt <4 x double> %x3, %sel1
+  %sel2 = select <4 x i1> %cmp2, <4 x double> %x3, <4 x double> %sel1
+  ret <4 x double> %sel2
+}
+
-- 
2.34.1