From 23a0d7597f1eca70f5b2c64b7f260f2ba01ee09c Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Tue, 13 Jun 2017 07:13:47 +0000
Subject: [PATCH] [AVX-512] Mark masked version of vpcmpeq as being commutable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@305275 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86InstrAVX512.td             |  1 +
 test/CodeGen/X86/stack-folding-int-avx512.ll | 14 ++++++++++++++
 2 files changed, 15 insertions(+)

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index d8702693884..56ae8f6cc79 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1631,6 +1631,7 @@ multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
              [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
                                      (_.VT (bitconvert (_.LdFrag addr:$src2)))))],
              IIC_SSE_ALU_F32P_RM>, EVEX_4V;
+  let isCommutable = IsCommutable in
   def rrk : AVX512BI<opc, MRMSrcReg,
               (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
               !strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
diff --git a/test/CodeGen/X86/stack-folding-int-avx512.ll b/test/CodeGen/X86/stack-folding-int-avx512.ll
index eaa8219a2bc..7b51b55630f 100644
--- a/test/CodeGen/X86/stack-folding-int-avx512.ll
+++ b/test/CodeGen/X86/stack-folding-int-avx512.ll
@@ -518,6 +518,20 @@ define i16 @stack_fold_pcmpeqd_mask(<16 x i32> %a0, <16 x i32> %a1, <16 x i32>*
   ret i16 %7
 }
 
+define i16 @stack_fold_pcmpeqd_mask_commuted(<16 x i32> %a0, <16 x i32> %a1, <16 x i32>* %a2, i16 %mask) {
+  ;CHECK-LABEL: stack_fold_pcmpeqd_mask_commuted
+  ;CHECK:       vpcmpeqd {{-?[0-9]*}}(%rsp), {{%zmm[0-9][0-9]*}}, {{%k[0-7]}} {{{%k[0-7]}}} {{.*#+}} 64-byte Folded Reload
+  %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()
+  ; load and add are here to keep the operations below the side effecting block and to avoid folding the wrong load
+  %2 = load <16 x i32>, <16 x i32>* %a2
+  %3 = add <16 x i32> %a1, %2
+  %4 = bitcast i16 %mask to <16 x i1>
+  %5 = icmp eq <16 x i32> %a0, %3
+  %6 = and <16 x i1> %4, %5
+  %7 = bitcast <16 x i1> %6 to i16
+  ret i16 %7
+}
+
 define i16 @stack_fold_pcmpled_mask(<16 x i32> %a0, <16 x i32> %a1, <16 x i32>* %a2, i16 %mask) {
   ;CHECK-LABEL: stack_fold_pcmpled_mask
   ;CHECK:       vpcmpled {{-?[0-9]*}}(%rsp), {{%zmm[0-9][0-9]*}}, {{%k[0-7]}} {{{%k[0-7]}}} {{.*#+}} 64-byte Folded Reload
-- 
2.50.1