[llvm] [AMDGPU] Fixed llvm-debuginfo-analyzer for AMDGPU. (PR #145125)

Adam Yang via llvm-commits llvm-commits at lists.llvm.org
Fri Jun 20 17:08:56 PDT 2025


https://github.com/adam-yang created https://github.com/llvm/llvm-project/pull/145125

Constructing Target triple with `ObjectFile::makeTriple` instead of just with `Arch` and leaving the rest unknown. Also creating the subtarget with the `CPU`. AMDGPU needs the full triple and `CPU` to disassemble correctly.

To run a full test, also fixed a failure with the `$noreg` operand in `DBG_VALUE`, and writing relocation with `AMDPAL`.

>From 472c5f1352bdc2a85876afe64ca97bc38cca1cd2 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Fri, 20 Jun 2025 16:52:37 -0700
Subject: [PATCH] Made llvm-debuginfo-analyzer work for AMDGPU. A few changes
 to generate DWARF correctly in AMDGPU

---
 .../LogicalView/Readers/LVBinaryReader.h      |   3 +-
 .../LogicalView/Readers/LVBinaryReader.cpp    |   4 +-
 .../LogicalView/Readers/LVCodeViewReader.cpp  |  10 +-
 .../LogicalView/Readers/LVDWARFReader.cpp     |  12 +-
 .../AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp  |   3 +-
 .../Target/AMDGPU/SIPreAllocateWWMRegs.cpp    |   3 +
 .../llvm-debuginfo-analyzer/DWARF/amdgpu.ll   | 103 ++++++++++++++++++
 7 files changed, 127 insertions(+), 11 deletions(-)
 create mode 100644 llvm/test/tools/llvm-debuginfo-analyzer/DWARF/amdgpu.ll

diff --git a/llvm/include/llvm/DebugInfo/LogicalView/Readers/LVBinaryReader.h b/llvm/include/llvm/DebugInfo/LogicalView/Readers/LVBinaryReader.h
index 1847fa8323480..1f8b884bc1b5d 100644
--- a/llvm/include/llvm/DebugInfo/LogicalView/Readers/LVBinaryReader.h
+++ b/llvm/include/llvm/DebugInfo/LogicalView/Readers/LVBinaryReader.h
@@ -159,7 +159,8 @@ class LVBinaryReader : public LVReader {
   LVAddress WasmCodeSectionOffset = 0;
 
   // Loads all info for the architecture of the provided object file.
-  Error loadGenericTargetInfo(StringRef TheTriple, StringRef TheFeatures);
+  Error loadGenericTargetInfo(StringRef TheTriple, StringRef TheFeatures,
+                              StringRef CPU);
 
   virtual void mapRangeAddress(const object::ObjectFile &Obj) {}
   virtual void mapRangeAddress(const object::ObjectFile &Obj,
diff --git a/llvm/lib/DebugInfo/LogicalView/Readers/LVBinaryReader.cpp b/llvm/lib/DebugInfo/LogicalView/Readers/LVBinaryReader.cpp
index 80b4185b7c600..414f0f3efc82d 100644
--- a/llvm/lib/DebugInfo/LogicalView/Readers/LVBinaryReader.cpp
+++ b/llvm/lib/DebugInfo/LogicalView/Readers/LVBinaryReader.cpp
@@ -275,7 +275,8 @@ void LVBinaryReader::mapVirtualAddress(const object::COFFObjectFile &COFFObj) {
 }
 
 Error LVBinaryReader::loadGenericTargetInfo(StringRef TheTriple,
-                                            StringRef TheFeatures) {
+                                            StringRef TheFeatures,
+                                            StringRef CPU) {
   std::string TargetLookupError;
   const Target *TheTarget =
       TargetRegistry::lookupTarget(TheTriple, TargetLookupError);
@@ -298,7 +299,6 @@ Error LVBinaryReader::loadGenericTargetInfo(StringRef TheTriple,
   MAI.reset(AsmInfo);
 
   // Target subtargets.
-  StringRef CPU;
   MCSubtargetInfo *SubtargetInfo(
       TheTarget->createMCSubtargetInfo(TheTriple, CPU, TheFeatures));
   if (!SubtargetInfo)
diff --git a/llvm/lib/DebugInfo/LogicalView/Readers/LVCodeViewReader.cpp b/llvm/lib/DebugInfo/LogicalView/Readers/LVCodeViewReader.cpp
index e5895516b5e77..2ff70816b4bf1 100644
--- a/llvm/lib/DebugInfo/LogicalView/Readers/LVCodeViewReader.cpp
+++ b/llvm/lib/DebugInfo/LogicalView/Readers/LVCodeViewReader.cpp
@@ -1190,7 +1190,12 @@ Error LVCodeViewReader::loadTargetInfo(const ObjectFile &Obj) {
     FeaturesValue = SubtargetFeatures();
   }
   FeaturesValue = *Features;
-  return loadGenericTargetInfo(TT.str(), FeaturesValue.getString());
+
+  StringRef CPU;
+  if (auto OptCPU = Obj.tryGetCPUName())
+    CPU = *OptCPU;
+
+  return loadGenericTargetInfo(TT.str(), FeaturesValue.getString(), CPU);
 }
 
 Error LVCodeViewReader::loadTargetInfo(const PDBFile &Pdb) {
@@ -1200,8 +1205,9 @@ Error LVCodeViewReader::loadTargetInfo(const PDBFile &Pdb) {
   TT.setOS(Triple::Win32);
 
   StringRef TheFeature = "";
+  StringRef TheCPU = "";
 
-  return loadGenericTargetInfo(TT.str(), TheFeature);
+  return loadGenericTargetInfo(TT.str(), TheFeature, TheCPU);
 }
 
 std::string LVCodeViewReader::getRegisterName(LVSmall Opcode,
diff --git a/llvm/lib/DebugInfo/LogicalView/Readers/LVDWARFReader.cpp b/llvm/lib/DebugInfo/LogicalView/Readers/LVDWARFReader.cpp
index 7ff96ae90a7fd..bbaff31bf622e 100644
--- a/llvm/lib/DebugInfo/LogicalView/Readers/LVDWARFReader.cpp
+++ b/llvm/lib/DebugInfo/LogicalView/Readers/LVDWARFReader.cpp
@@ -949,10 +949,7 @@ LVElement *LVDWARFReader::getElementForOffset(LVOffset Offset,
 Error LVDWARFReader::loadTargetInfo(const ObjectFile &Obj) {
   // Detect the architecture from the object file. We usually don't need OS
   // info to lookup a target and create register info.
-  Triple TT;
-  TT.setArch(Triple::ArchType(Obj.getArch()));
-  TT.setVendor(Triple::UnknownVendor);
-  TT.setOS(Triple::UnknownOS);
+  Triple TT = Obj.makeTriple();
 
   // Features to be passed to target/subtarget
   Expected<SubtargetFeatures> Features = Obj.getFeatures();
@@ -962,7 +959,12 @@ Error LVDWARFReader::loadTargetInfo(const ObjectFile &Obj) {
     FeaturesValue = SubtargetFeatures();
   }
   FeaturesValue = *Features;
-  return loadGenericTargetInfo(TT.str(), FeaturesValue.getString());
+
+  StringRef CPU;
+  if (auto OptCPU = Obj.tryGetCPUName())
+    CPU = *OptCPU;
+
+  return loadGenericTargetInfo(TT.str(), FeaturesValue.getString(), CPU);
 }
 
 void LVDWARFReader::mapRangeAddress(const ObjectFile &Obj) {
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
index 78bd8de07306f..3f7953acef2df 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
@@ -231,7 +231,8 @@ class ELFAMDGPUAsmBackend : public AMDGPUAsmBackend {
 public:
   ELFAMDGPUAsmBackend(const Target &T, const Triple &TT)
       : AMDGPUAsmBackend(T), Is64Bit(TT.isAMDGCN()),
-        HasRelocationAddend(TT.getOS() == Triple::AMDHSA) {
+        HasRelocationAddend(TT.getOS() == Triple::AMDHSA ||
+                            TT.getOS() == Triple::AMDPAL) {
     switch (TT.getOS()) {
     case Triple::AMDHSA:
       OSABI = ELF::ELFOSABI_AMDGPU_HSA;
diff --git a/llvm/lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp b/llvm/lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp
index 205a45a045a42..469a6525b4ac0 100644
--- a/llvm/lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp
+++ b/llvm/lib/Target/AMDGPU/SIPreAllocateWWMRegs.cpp
@@ -130,6 +130,9 @@ void SIPreAllocateWWMRegs::rewriteRegs(MachineFunction &MF) {
         if (VirtReg.isPhysical())
           continue;
 
+        if (MI.isDebugInstr() && VirtReg == AMDGPU::NoRegister)
+          continue;
+
         if (!VRM->hasPhys(VirtReg))
           continue;
 
diff --git a/llvm/test/tools/llvm-debuginfo-analyzer/DWARF/amdgpu.ll b/llvm/test/tools/llvm-debuginfo-analyzer/DWARF/amdgpu.ll
new file mode 100644
index 0000000000000..1d031979309a6
--- /dev/null
+++ b/llvm/test/tools/llvm-debuginfo-analyzer/DWARF/amdgpu.ll
@@ -0,0 +1,103 @@
+; RUN: llc %s -o %t.o -mcpu=gfx1030 -filetype=obj -O0
+; RUN: llvm-debuginfo-analyzer %t.o --print=all --attribute=all | FileCheck %s
+
+; This test compiles this module with AMDGPU backend under -O0,
+; and makes sure llvm-debuginfo-analzyer works for it.
+
+; Simple checks to make sure llvm-debuginfo-analzyer didn't fail early.
+; CHECK: Logical View:
+; CHECK: {CompileUnit}
+; CHECK: {Code} 's_endpgm'
+
+source_filename = "module"
+target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-p9:192:256:256:32-p10:32:32-p11:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9-p32:32:32-v8:8-v16:16-v32:32-v48:32-v64:32-v80:32-v96:32-v112:32-v128:32-v144:32-v160:32-v176:32-v192:32-v208:32-v224:32-v240:32-v256:32-i1:32-i8:8-i16:16-i32:32-i64:32-f16:16-f32:32-f64:32"
+target triple = "amdgcn-amd-amdpal"
+
+%dx.types.ResRet.f32 = type { float, float, float, float, i32 }
+
+; Function Attrs: memory(readwrite)
+define dllexport amdgpu_cs void @_amdgpu_cs_main(i32 inreg noundef %globalTable, i32 inreg noundef %userdata4, <3 x i32> inreg noundef %WorkgroupId, i32 inreg noundef %MultiDispatchInfo, <3 x i32> noundef %LocalInvocationId) #0 !dbg !14 {
+  %LocalInvocationId.i0 = extractelement <3 x i32> %LocalInvocationId, i64 0, !dbg !28
+  %WorkgroupId.i0 = extractelement <3 x i32> %WorkgroupId, i64 0, !dbg !28
+  %1 = call i64 @llvm.amdgcn.s.getpc(), !dbg !28
+  %2 = shl i32 %WorkgroupId.i0, 6, !dbg !28
+  %3 = add i32 %LocalInvocationId.i0, %2, !dbg !28
+    #dbg_value(i32 %3, !29, !DIExpression(DW_OP_LLVM_fragment, 0, 32), !28)
+  %4 = and i64 %1, -4294967296, !dbg !30
+  %5 = zext i32 %userdata4 to i64, !dbg !30
+  %6 = or disjoint i64 %4, %5, !dbg !30
+  %7 = inttoptr i64 %6 to ptr addrspace(4), !dbg !30
+  call void @llvm.assume(i1 true) [ "align"(ptr addrspace(4) %7, i32 4), "dereferenceable"(ptr addrspace(4) %7, i32 -1) ], !dbg !30
+  %8 = load <4 x i32>, ptr addrspace(4) %7, align 4, !dbg !30, !invariant.load !2
+  %9 = call float @llvm.amdgcn.struct.buffer.load.format.f32(<4 x i32> %8, i32 %3, i32 0, i32 0, i32 0), !dbg !30
+    #dbg_value(%dx.types.ResRet.f32 poison, !31, !DIExpression(), !32)
+  %10 = fmul reassoc arcp contract afn float %9, 2.000000e+00, !dbg !33
+    #dbg_value(float %10, !34, !DIExpression(), !35)
+  call void @llvm.assume(i1 true) [ "align"(ptr addrspace(4) %7, i32 4), "dereferenceable"(ptr addrspace(4) %7, i32 -1) ], !dbg !36
+  %11 = getelementptr i8, ptr addrspace(4) %7, i64 32, !dbg !36
+  %.upto01 = insertelement <4 x float> poison, float %10, i64 0, !dbg !36
+  %12 = shufflevector <4 x float> %.upto01, <4 x float> poison, <4 x i32> zeroinitializer, !dbg !36
+  %13 = load <4 x i32>, ptr addrspace(4) %11, align 4, !dbg !36, !invariant.load !2
+  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %12, <4 x i32> %13, i32 %3, i32 0, i32 0, i32 0), !dbg !36
+  ret void, !dbg !37
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare noundef i64 @llvm.amdgcn.s.getpc() #1
+
+; Function Attrs: nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: write)
+declare void @llvm.assume(i1 noundef) #2
+
+; Function Attrs: nocallback nofree nosync nounwind willreturn memory(write)
+declare void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float>, <4 x i32>, i32, i32, i32, i32 immarg) #3
+
+; Function Attrs: nocallback nofree nosync nounwind willreturn memory(read)
+declare float @llvm.amdgcn.struct.buffer.load.format.f32(<4 x i32>, i32, i32, i32, i32 immarg) #4
+
+attributes #0 = { memory(readwrite) "amdgpu-flat-work-group-size"="64,64" "amdgpu-memory-bound"="false" "amdgpu-num-sgpr"="4294967295" "amdgpu-num-vgpr"="4294967295" "amdgpu-prealloc-sgpr-spill-vgprs" "amdgpu-unroll-threshold"="1200" "amdgpu-wave-limiter"="false" "amdgpu-work-group-info-arg-no"="3" "denormal-fp-math"="ieee" "denormal-fp-math-f32"="preserve-sign" "target-features"=",+wavefrontsize64,+cumode,+enable-flat-scratch" }
+attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
+attributes #2 = { nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: write) }
+attributes #3 = { nocallback nofree nosync nounwind willreturn memory(write) }
+attributes #4 = { nocallback nofree nosync nounwind willreturn memory(read) }
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!12, !13}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "dxcoob 1.7.2308.16 (52da17e29)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, globals: !3)
+!1 = !DIFile(filename: "tests\\basic_var.hlsl", directory: "")
+!2 = !{}
+!3 = !{!4, !10}
+!4 = distinct !DIGlobalVariableExpression(var: !5, expr: !DIExpression())
+!5 = !DIGlobalVariable(name: "u0", linkageName: "\01?u0@@3V?$RWBuffer at M@@A", scope: !0, file: !1, line: 2, type: !6, isLocal: false, isDefinition: true)
+!6 = !DICompositeType(tag: DW_TAG_class_type, name: "RWBuffer<float>", file: !1, line: 2, size: 32, align: 32, elements: !2, templateParams: !7)
+!7 = !{!8}
+!8 = !DITemplateTypeParameter(name: "element", type: !9)
+!9 = !DIBasicType(name: "float", size: 32, align: 32, encoding: DW_ATE_float)
+!10 = distinct !DIGlobalVariableExpression(var: !11, expr: !DIExpression())
+!11 = !DIGlobalVariable(name: "u1", linkageName: "\01?u1@@3V?$RWBuffer at M@@A", scope: !0, file: !1, line: 3, type: !6, isLocal: false, isDefinition: true)
+!12 = !{i32 2, !"Dwarf Version", i32 5}
+!13 = !{i32 2, !"Debug Info Version", i32 3}
+!14 = distinct !DISubprogram(name: "main", scope: !1, file: !1, line: 7, type: !15, scopeLine: 7, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0)
+!15 = !DISubroutineType(types: !16)
+!16 = !{null, !17}
+!17 = !DIDerivedType(tag: DW_TAG_typedef, name: "uint3", file: !1, baseType: !18)
+!18 = !DICompositeType(tag: DW_TAG_class_type, name: "vector<unsigned int, 3>", file: !1, size: 96, align: 32, elements: !19, templateParams: !24)
+!19 = !{!20, !22, !23}
+!20 = !DIDerivedType(tag: DW_TAG_member, name: "x", scope: !18, file: !1, baseType: !21, size: 32, align: 32, flags: DIFlagPublic)
+!21 = !DIBasicType(name: "unsigned int", size: 32, align: 32, encoding: DW_ATE_unsigned)
+!22 = !DIDerivedType(tag: DW_TAG_member, name: "y", scope: !18, file: !1, baseType: !21, size: 32, align: 32, offset: 32, flags: DIFlagPublic)
+!23 = !DIDerivedType(tag: DW_TAG_member, name: "z", scope: !18, file: !1, baseType: !21, size: 32, align: 32, offset: 64, flags: DIFlagPublic)
+!24 = !{!25, !26}
+!25 = !DITemplateTypeParameter(name: "element", type: !21)
+!26 = !DITemplateValueParameter(name: "element_count", type: !27, value: i32 3)
+!27 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed)
+!28 = !DILocation(line: 7, column: 17, scope: !14)
+!29 = !DILocalVariable(name: "dtid", arg: 1, scope: !14, file: !1, line: 7, type: !17)
+!30 = !DILocation(line: 11, column: 18, scope: !14)
+!31 = !DILocalVariable(name: "my_var", scope: !14, file: !1, line: 11, type: !9)
+!32 = !DILocation(line: 11, column: 9, scope: !14)
+!33 = !DILocation(line: 14, column: 26, scope: !14)
+!34 = !DILocalVariable(name: "my_var2", scope: !14, file: !1, line: 14, type: !9)
+!35 = !DILocation(line: 14, column: 9, scope: !14)
+!36 = !DILocation(line: 17, column: 14, scope: !14)
+!37 = !DILocation(line: 19, column: 1, scope: !14)
\ No newline at end of file



More information about the llvm-commits mailing list