From: Craig Topper <craig.topper@gmail.com>
Date: Sun, 19 Feb 2017 21:32:15 +0000 (+0000)
Subject: [AVX-512] Disable peephole optimizations on the VPTERNLOG commute test. Add new patte... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=97295ca18160356c5c53091b4d85bc40f63d8773;p=llvm

[AVX-512] Disable peephole optimizations on the VPTERNLOG commute test. Add new patterns to enable isel to fold the loads on it own.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295616 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index dea8b2b49b5..cdeefb572ec 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -8915,6 +8915,17 @@ def VPTERNLOG213_imm8 : SDNodeXForm<imm, [{
   if (Imm & 0x20) NewImm |= 0x08;
   return getI8Imm(NewImm, SDLoc(N));
 }]>;
+def VPTERNLOG132_imm8 : SDNodeXForm<imm, [{
+  // Convert a VPTERNLOG immediate by swapping operand 1 and operand 2.
+  uint8_t Imm = N->getZExtValue();
+  // Swap bits 1/2 and 5/6.
+  uint8_t NewImm = Imm & 0x99;
+  if (Imm & 0x02) NewImm |= 0x04;
+  if (Imm & 0x04) NewImm |= 0x02;
+  if (Imm & 0x20) NewImm |= 0x40;
+  if (Imm & 0x40) NewImm |= 0x20;
+  return getI8Imm(NewImm, SDLoc(N));
+}]>;
 
 multiclass avx512_ternlog<bits<8> opc, string OpcodeStr, SDNode OpNode,
                           X86VectorVTInfo _>{
@@ -8958,6 +8969,45 @@ multiclass avx512_ternlog<bits<8> opc, string OpcodeStr, SDNode OpNode,
             (!cast<Instruction>(NAME#_.ZSuffix#rrik) _.RC:$src1, _.KRCWM:$mask,
              _.RC:$src2, _.RC:$src3, (VPTERNLOG213_imm8 imm:$src4))>;
   }
+
+  // Additional patterns for matching loads in other positions.
+  def : Pat<(_.VT (OpNode (bitconvert (_.LdFrag addr:$src3)),
+                          _.RC:$src2, _.RC:$src1, (i8 imm:$src4))),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmi) _.RC:$src1, _.RC:$src2,
+                                   addr:$src3, (VPTERNLOG321_imm8 imm:$src4))>;
+  def : Pat<(_.VT (OpNode _.RC:$src1,
+                          (bitconvert (_.LdFrag addr:$src3)),
+                          _.RC:$src2, (i8 imm:$src4))),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmi) _.RC:$src1, _.RC:$src2,
+                                   addr:$src3, (VPTERNLOG132_imm8 imm:$src4))>;
+
+  // Additional patterns for matching zero masking with loads in other
+  // positions.
+  let AddedComplexity = 30 in {
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src2, _.RC:$src1, (i8 imm:$src4)),
+                   _.ImmAllZerosV)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmikz) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG321_imm8 imm:$src4))>;
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode _.RC:$src1, (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src2, (i8 imm:$src4)),
+                   _.ImmAllZerosV)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmikz) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG132_imm8 imm:$src4))>;
+  }
+
+  // Additional patterns for matching masked loads with different
+  // operand orders.
+  let AddedComplexity = 20 in {
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode _.RC:$src1, (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src2, (i8 imm:$src4)),
+                   _.RC:$src1)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG132_imm8 imm:$src4))>;
+  }
 }
 
 multiclass avx512_common_ternlog<string OpcodeStr, AVX512VLVectorVTInfo _>{
diff --git a/test/CodeGen/X86/avx512-vpternlog-commute.ll b/test/CodeGen/X86/avx512-vpternlog-commute.ll
index 1f203936bcc..988b43d403b 100644
--- a/test/CodeGen/X86/avx512-vpternlog-commute.ll
+++ b/test/CodeGen/X86/avx512-vpternlog-commute.ll
@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx | FileCheck %s
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-apple-darwin -mcpu=skx | FileCheck %s
 
 ; These test cases demonstrate cases where vpternlog could benefit from being commuted.