commit 84a1f07363405c25304e3fd3df55a216676e9984
author Craig Topper <craig.topper@intel.com> Tue Jul 09 19:55:28 2019 +0000
committer Craig Topper <craig.topper@intel.com> Tue Jul 09 19:55:28 2019 +0000
tree da18767d1cb035014d1b7ae007b82ca8a62e9521
parent c236eeaf7d8a95a3457c4658600ee5b79052b7a2

[X86][AMDGPU][DAGCombiner] Move call to allowsMemoryAccess into isLoadBitCastBeneficial/isStoreBitCastBeneficial to allow X86 to bypass it

Basically the problem is that X86 doesn't set the Fast flag from
allowsMemoryAccess on certain CPUs due to slow unaligned memory
subtarget features. This prevents bitcasts from being folded into
loads and stores. But all vector loads and stores of the same width
are the same cost on X86.

This patch merges the allowsMemoryAccess call into isLoadBitCastBeneficial to allow X86 to skip it.

Differential Revision: https://reviews.llvm.org/D64295

llvm-svn: 365549

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 766294d..0ccd58d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -719,8 +719,9 @@
   return (OldSize < 32);
 }
 
-bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
-                                                   EVT CastTy) const {
+bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy, EVT CastTy,
+                                                   const SelectionDAG &DAG,
+                                                   const MachineMemOperand &MMO) const {
 
   assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
 
@@ -730,8 +731,12 @@
   unsigned LScalarSize = LoadTy.getScalarSizeInBits();
   unsigned CastScalarSize = CastTy.getScalarSizeInBits();
 
-  return (LScalarSize < CastScalarSize) ||
-         (CastScalarSize >= 32);
+  if ((LScalarSize >= CastScalarSize) && (CastScalarSize < 32))
+    return false;
+
+  bool Fast = false;
+  return allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), CastTy,
+                            MMO, &Fast) && Fast;
 }
 
 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 9723fc3..40ff24f 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -182,7 +182,8 @@
                              ISD::LoadExtType ExtType,
                              EVT ExtVT) const override;
 
-  bool isLoadBitCastBeneficial(EVT, EVT) const final;
+  bool isLoadBitCastBeneficial(EVT, EVT, const SelectionDAG &DAG,
+                               const MachineMemOperand &MMO) const final;
 
   bool storeOfVectorConstantIsCheap(EVT MemVT,
                                     unsigned NumElem,

llvm/lib/Target/X86/X86ISelLowering.cpp

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 370ecef..3cab44b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -4941,8 +4941,9 @@
   return Subtarget.hasLZCNT();
 }
 
-bool X86TargetLowering::isLoadBitCastBeneficial(EVT LoadVT,
-                                                EVT BitcastVT) const {
+bool X86TargetLowering::isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT,
+                                                const SelectionDAG &DAG,
+                                                const MachineMemOperand &MMO) const {
   if (!Subtarget.hasAVX512() && !LoadVT.isVector() && BitcastVT.isVector() &&
       BitcastVT.getVectorElementType() == MVT::i1)
     return false;
@@ -4950,7 +4951,12 @@
   if (!Subtarget.hasDQI() && BitcastVT == MVT::v8i1 && LoadVT == MVT::i8)
     return false;
 
-  return TargetLowering::isLoadBitCastBeneficial(LoadVT, BitcastVT);
+  // If both types are legal vectors, it's always ok to convert them.
+  if (LoadVT.isVector() && BitcastVT.isVector() &&
+      isTypeLegal(LoadVT) && isTypeLegal(BitcastVT))
+    return true;
+
+  return TargetLowering::isLoadBitCastBeneficial(LoadVT, BitcastVT, DAG, MMO);
 }
 
 bool X86TargetLowering::canMergeStoresTo(unsigned AddressSpace, EVT MemVT,

llvm/lib/Target/X86/X86ISelLowering.h

diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index 0631fb7..e0be03b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -1127,7 +1127,9 @@
       return NumElem > 2;
     }
 
-    bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT) const override;
+    bool isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT,
+                                 const SelectionDAG &DAG,
+                                 const MachineMemOperand &MMO) const override;
 
     /// Intel processors have a unified instruction and data cache
     const char * getClearCacheBuiltinName() const override {