AMDGPU/SI: Don't emit group segment global variables

[• tstellarAMD]

Summary

Only global or readonly segment variables should appear in object files.

[arsenm]

lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

129

I think you can get the Triple from the TargetMachine and don't have to construct one here

[• tstellarAMD]

Get Triple from Targetmachine.

[arsenm]

LGTM

[aizatsky]

I'm not yet sure how, but it seems that this change broke a build:

http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/8259/steps/check-llvm%20msan/logs/stdio

29897==WARNING: MemorySanitizer: use-of-uninitialized-value

#0 0x4396b49 in llvm::raw_ostream::operator<<(unsigned long) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/Support/raw_ostream.cpp:110:7
#1 0xe2f586 in llvm::AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(amd_kernel_code_s const&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp:187:40
#2 0xbbedbb in llvm::AMDGPUAsmPrinter::EmitAmdKernelCodeT(llvm::MachineFunction const&, llvm::AMDGPUAsmPrinter::SIProgramInfo const&) const /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:596:3
#3 0xbbca50 in llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:99:5
#4 0x29bf8fb in llvm::AsmPrinter::EmitFunctionBody() /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp:875:3
#5 0xbc0400 in llvm::AMDGPUAsmPrinter::runOnMachineFunction(llvm::MachineFunction&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:167:3
#6 0x2db7343 in llvm::MachineFunctionPass::runOnFunction(llvm::Function&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/CodeGen/MachineFunctionPass.cpp:43:10
#7 0x36c4bff in llvm::FPPassManager::runOnFunction(llvm::Function&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/IR/LegacyPassManager.cpp:1521:23
#8 0x36c5297 in llvm::FPPassManager::runOnModule(llvm::Module&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/IR/LegacyPassManager.cpp:1542:16
#9 0x36c65d8 in runOnModule /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/IR/LegacyPassManager.cpp:1598:23
#10 0x36c65d8 in llvm::legacy::PassManagerImpl::run(llvm::Module&) /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/lib/IR/LegacyPassManager.cpp:1701
#11 0x6a09af in compileModule /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/tools/llc/llc.cpp:381:5
#12 0x6a09af in main /mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm/tools/llc/llc.cpp:204
#13 0x7ffb1bd0aec4 in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x21ec4)
#14 0x60dd83 in _start (/mnt/b/sanitizer-buildbot3/sanitizer-x86_64-linux-fast/build/llvm_build_msan/bin/llc+0x60dd83)

lib/Target/AMDGPU/AMDGPUAsmPrinter.h

public:
   void EmitInstruction(const MachineInstr *MI) override;
 
   void EmitFunctionBodyStart() override;
 
   void EmitEndOfAsmFile(Module &M) override;
 
   void EmitFunctionEntryLabel() override;
 
+  void EmitGlobalVariable(const GlobalVariable *GV) override;
+
   bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
                        unsigned AsmVariant, const char *ExtraCode,
                        raw_ostream &O) override;
 
 protected:
   std::vector<std::string> DisasmLines, HexLines;
   size_t DisasmLineMaxLen;
 };
 
 } // End anonymous llvm
 
 #endif

lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

AMDGPUTargetStreamer *TS =
         static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
     TS->EmitAMDGPUSymbolType(CurrentFnSym->getName(),
                              ELF::STT_AMDGPU_HSA_KERNEL);
   }
 
   AsmPrinter::EmitFunctionEntryLabel();
 }
 
+void AMDGPUAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
+
+  const Module *M = GV->getParent();
+  if (Triple(M->getTargetTriple()).getOS() != Triple::AMDHSA ||
+      !AMDGPU::isGroupSegment(GV))
+    return AsmPrinter::EmitGlobalVariable(GV);
+}
+
 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
 
   // The starting address of all shader programs must be 256 bytes aligned.
   MF.setAlignment(8);
 
   SetupMachineFunction(MF);
 
   MCContext &Context = getObjFileLowering().getContext();

lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

 #ifndef LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H
 #define LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H
 
 #include "AMDKernelCodeT.h"
 
 namespace llvm {
 
 class FeatureBitset;
+class GlobalValue;
 class MCContext;
 class MCSection;
 
 namespace AMDGPU {
 
 struct IsaVersion {
   unsigned Major;
   unsigned Minor;
   unsigned Stepping;
 };
 
 IsaVersion getIsaVersion(const FeatureBitset &Features);
 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header,
                                const FeatureBitset &Features);
 MCSection *getHSATextSection(MCContext &Ctx);
 
+bool isGroupSegment(const GlobalValue *GV);
+
 } // end namespace AMDGPU
 } // end namespace llvm
 
 #endif

lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

 //===-- AMDGPUBaseInfo.cpp - AMDGPU Base encoding information--------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 #include "AMDGPUBaseInfo.h"
+#include "AMDGPU.h"
+#include "llvm/IR/GlobalValue.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCSectionELF.h"
 #include "llvm/MC/SubtargetFeature.h"
 
 #define GET_SUBTARGETINFO_ENUM
 #include "AMDGPUGenSubtargetInfo.inc"
 #undef GET_SUBTARGETINFO_ENUM
 

 MCSection *getHSATextSection(MCContext &Ctx) {
   return Ctx.getELFSection(".hsatext", ELF::SHT_PROGBITS,
                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
                            ELF::SHF_EXECINSTR |
                            ELF::SHF_AMDGPU_HSA_AGENT |
                            ELF::SHF_AMDGPU_HSA_CODE);
 }
 
+bool isGroupSegment(const GlobalValue *GV) {
+  return GV->getType()->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
+}
+
 } // End namespace AMDGPU
 } // End namespace llvm

test/CodeGen/AMDGPU/hsa-group-segment.ll

+; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=kaveri | FileCheck --check-prefix=HSA %s
+
+@internal_group = internal addrspace(3) global i32 undef
+@external_group = addrspace(3) global i32 undef
+
+define void @test() {
+entry:
+  store i32 0, i32 addrspace(3)* @internal_group
+  store i32 0, i32 addrspace(3)* @external_group
+  ret void
+}
+
+; HSA-NOT: internal_group:
+; HSA-NOT: external_group: