AMDGPU/GlobalISel: Fix RegBankSelect for unaligned, uniform constant loads

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@371416 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index 5414dd241aef4360251e9a3dde62d35035900092..c573111312bc9be7bdd4d903389c77b4b21a8639 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -320,12 +320,13 @@ AMDGPURegisterBankInfo::getInstrAlternativeMappingsIntrinsicWSideEffects(
   }
 }
 
-static bool isInstrUniformNonExtLoad(const MachineInstr &MI) {
+static bool isInstrUniformNonExtLoadAlign4(const MachineInstr &MI) {
   if (!MI.hasOneMemOperand())
     return false;
 
   const MachineMemOperand *MMO = *MI.memoperands_begin();
-  return MMO->getSize() >= 4 && AMDGPUInstrInfo::isUniformMMO(MMO);
+  return MMO->getSize() >= 4 && MMO->getAlignment() >= 4 &&
+         AMDGPUInstrInfo::isUniformMMO(MMO);
 }
 
 RegisterBankInfo::InstructionMappings
@@ -426,7 +427,7 @@ AMDGPURegisterBankInfo::getInstrAlternativeMappings(
     unsigned PtrSize = PtrTy.getSizeInBits();
     unsigned AS = PtrTy.getAddressSpace();
     LLT LoadTy = MRI.getType(MI.getOperand(0).getReg());
-    if (isInstrUniformNonExtLoad(MI) &&
+    if (isInstrUniformNonExtLoadAlign4(MI) &&
         (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS)) {
       const InstructionMapping &SSMapping = getInstructionMapping(
           1, 1, getOperandsMapping(
@@ -1482,7 +1483,7 @@ AMDGPURegisterBankInfo::getInstrMappingForLoad(const MachineInstr &MI) const {
   const ValueMapping *ValMapping;
   const ValueMapping *PtrMapping;
 
-  if (isInstrUniformNonExtLoad(MI) &&
+  if (isInstrUniformNonExtLoadAlign4(MI) &&
       (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS)) {
     // We have a uniform instruction so we want to use an SMRD load
     ValMapping = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size);

diff --git a/test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir b/test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
index 1370ed3fa5f9bdcbf249bd160173e9de9ae91e55..d129383817b3d902b088482a51bc455930e4005f 100644 (file)
--- a/test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
+++ b/test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
@@ -1,5 +1,5 @@
-# RUN: llc -march=amdgcn -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
-# RUN: llc -march=amdgcn -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
 
 --- |
   define amdgpu_kernel void @load_global_v8i32_non_uniform(<8 x i32> addrspace(1)* %in) {
@@ -65,6 +65,9 @@
   define amdgpu_kernel void @extload_global_i8_to_i32_uniform() { ret void }
   define amdgpu_kernel void @extload_constant_i16_to_i32_uniform() { ret void }
   define amdgpu_kernel void @extload_global_i16_to_i32_uniform() { ret void }
+  define amdgpu_kernel void @load_constant_i32_uniform_align4() {ret void}
+  define amdgpu_kernel void @load_constant_i32_uniform_align2() {ret void}
+  define amdgpu_kernel void @load_constant_i32_uniform_align1() {ret void}
 
   declare i32 @llvm.amdgcn.workitem.id.x() #0
   attributes #0 = { nounwind readnone }
@@ -586,3 +589,49 @@ body: |
     %0:_(p4) = COPY $sgpr0_sgpr1
     %1:_(s32) = G_LOAD %0 :: (load 2, addrspace 1, align 2)
 ...
+
+---
+name: load_constant_i32_uniform_align4
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1
+    ; CHECK-LABEL: name: load_constant_i32_uniform_align4
+    ; CHECK: %0:sgpr(p4) = COPY $sgpr0_sgpr1
+    ; CHECK: %1:sgpr(s32) = G_LOAD %0(p4) :: (load 4, addrspace 4)
+    %0:_(p4) = COPY $sgpr0_sgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, addrspace 4, align 4)
+...
+
+---
+name: load_constant_i32_uniform_align2
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1
+    ; CHECK-LABEL: name: load_constant_i32_uniform_align2
+    ; CHECK: %0:sgpr(p4) = COPY $sgpr0_sgpr1
+    ; CHECK: %2:vgpr(p4) = COPY %0(p4)
+    ; CHECK: %1:vgpr(s32) = G_LOAD %2(p4) :: (load 4, align 2, addrspace 4)
+
+    %0:_(p4) = COPY $sgpr0_sgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, addrspace 4, align 2)
+...
+
+---
+name: load_constant_i32_uniform_align1
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1
+
+    ; CHECK-LABEL: name: load_constant_i32_uniform_align1
+    ; CHECK: %0:sgpr(p4) = COPY $sgpr0_sgpr1
+    ; CHECK: %2:vgpr(p4) = COPY %0(p4)
+    ; CHECK: %1:vgpr(s32) = G_LOAD %2(p4) :: (load 4, align 1, addrspace 4)
+    %0:_(p4) = COPY $sgpr0_sgpr1
+    %1:_(s32) = G_LOAD %0 :: (load 4, addrspace 4, align 1)
+...