AMDGPU/SI: Emit constant arrays in the .hsrodata_readonly_agent section

[• tstellarAMD]

Summary

This is done only when target HSA.

[arsenm]

test/CodeGen/AMDGPU/hsa-globals.ll

9

Should there be tests with a variety of alignments?

[• tstellarAMD]

Rebase after changes to D13806

[• tstellarAMD]

test/CodeGen/AMDGPU/hsa-globals.ll

9

Maybe, I'm not sure exactly how the output would be different.

[arsenm]

LGTM

[• tstellarAMD]

Rebased on top of recent changes.

lib/Target/AMDGPU/AMDGPUHSATargetObjectFile.h

 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
 #include "llvm/Target/TargetMachine.h"
 
 namespace llvm {
 
 class AMDGPUHSATargetObjectFile final : public TargetLoweringObjectFileELF {
 private:
   MCSection *DataGlobalAgentSection;
+  MCSection *RodataReadonlyAgentSection;
 
 public:
   void Initialize(MCContext &Ctx, const TargetMachine &TM) override;
 
   MCSection *SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind,
                                     Mangler &Mang,
                                     const TargetMachine &TM) const override;
 };
 
 } // end namespace llvm
 
 #endif

lib/Target/AMDGPU/AMDGPUHSATargetObjectFile.cpp

 void AMDGPUHSATargetObjectFile::Initialize(MCContext &Ctx,
                                            const TargetMachine &TM){
   TargetLoweringObjectFileELF::Initialize(Ctx, TM);
   InitializeELF(TM.Options.UseInitArray);
 
   TextSection = AMDGPU::getHSATextSection(Ctx);
 
   DataGlobalAgentSection = AMDGPU::getHSADataGlobalAgentSection(Ctx);
+
+  RodataReadonlyAgentSection = Ctx.getELFSection(".hsarodata_readonly_agent",
+                                                 ELF::SHT_PROGBITS,
+                                                 ELF::SHF_ALLOC |
+                                                 ELF::SHF_AMDGPU_HSA_READONLY |
+                                                 ELF::SHF_AMDGPU_HSA_AGENT);
 }
 
 
 static bool isModuleLinkage(const GlobalValue *GV) {
   return GV->getLinkage() == GlobalValue::CommonLinkage ||
          GV->getLinkage() == GlobalValue::InternalLinkage;
 }
 
 MCSection *AMDGPUHSATargetObjectFile::SelectSectionForGlobal(
                                         const GlobalValue *GV, SectionKind Kind,
                                         Mangler &Mang,
                                         const TargetMachine &TM) const {
   if (Kind.isText() && !GV->hasComdat())
     return getTextSection();
 
+  if (Kind.isReadOnly() && isModuleLinkage(GV))
+    return RodataReadonlyAgentSection;
+
   // isCommon means globally writeable with a zero initializer.
   if (isModuleLinkage(GV))
     return DataGlobalAgentSection;
 
   return TargetLoweringObjectFileELF::SelectSectionForGlobal(GV, Kind, Mang, TM);
 }

test/CodeGen/AMDGPU/hsa-globals.ll

+; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri | FileCheck --check-prefix=HSA-CI --check-prefix=HSA %s
+; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=carrizo | FileCheck --check-prefix=HSA-VI --check-prefix=HSA %s
+; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri -filetype=obj  | llvm-readobj -symbols -s -sd | FileCheck --check-prefix=ELF %s
+
+; FIXME: Add assembler support for globals
+; FIXME-XUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri | llvm-mc -filetype=obj -triple amdgcn--amdhsa -mcpu=kaveri | llvm-readobj -symbols -s -sd | FileCheck %s --check-prefix=ELF
+
+@const_array = internal unnamed_addr addrspace(2) constant [64 x i8] c"abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789<>", align 1
+
+; ELF: Section {
+; ELF: Name: .hsarodata_readonly_agent
+; ELF: Type: SHT_PROGBITS (0x1)
+; ELF: Flags [ (0xA00002)
+; ELF: SHF_ALLOC (0x2)
+; ELF: SHF_AMDGPU_HSA_AGENT (0x800000)
+; ELF: SHF_AMDGPU_HSA_READONLY (0x200000)
+; ELF: ]
+
+; ELF: Symbol {
+; ELF: Name: const_array
+; ELF: Binding: Local (0x0)
+; ELF: Type: Object (0x1)
+; ELF: Section: .hsarodata_readonly_agent (0x6)
+; ELF: }
+
+; HSA: .amdgpu_hsa_module_global const_array
+; HSA: .hsarodata_readonly_agent
+; HSA: const_array:
+; HSA: .ascii  "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789<>"
+
+
+define void @test(i8 addrspace(1) *%out, i32 %offset) {
+entry:
+  %ptr = getelementptr [64 x i8], [64 x i8] addrspace(2)* @const_array, i32 0, i32 %offset
+;  %0 = getelementptr inbounds [2 x i32], [2 x i32] addrspace(3)* @local_mem, i32 0, i32 %index
+  %val = load i8, i8 addrspace(2)* %ptr
+  store i8 %val, i8 addrspace(1)* %out
+  ret void
+}