[dwarfdump][AMDGPU] Support EF_AMDGPU_MACH_NONE

[scott.linder]

Summary

A quirk of the AMDGPU backend is EF_AMDGPU_MACH_NONE, which is not
specific to the r600 or amdgcn architecture, but can be combined with
either.

AMDGPU ELF code objects with this mach value cannot be mapped back to a
Triple architecture, as it could be either r600 or amdgcn. For
llvm-dwarfdump this means the normal method of inspecting
ObjectFile::getArch to determine how to handle relocations is
insufficient.

This patch adds an extra check for ELF code objects which would
otherwise be categorized as UnknownArch, making it possible to use
llvm-dwarfdump with them. For completeness it also adds support for
known r600 machines.

This also adds specific tests for llvm-dwarfdump for amdgcn and r600,
both with known and unknown mach values.

[jhenderson]

The patch title and code are to do with llvm-dwarfdump, but the commit message refers to llvm-objdump. Is this just a mistake?

llvm/lib/Object/RelocationResolver.cpp

260

Nit

850–851

No need for else after return.

[scott.linder]

In D144301#4138183, @jhenderson wrote:

The patch title and code are to do with llvm-dwarfdump, but the commit message refers to llvm-objdump. Is this just a mistake?

Yes, just a typo, thank you!

[scott.linder]

Address feedback

[scott.linder]

Ping

[jhenderson]

It looks like this code is testing code in libObject, but the testing is in llvm-dwarfdump. Should the tests be moved to test/Object?

llvm/lib/Object/RelocationResolver.cpp

262

I think this function is a bit incorrectly named. Strictly-speaking, this will return true/false only based on the EM* value, so it doesn't say anything about whether the object is a "known" AMDGPU version. Perhaps should just be named isAMDGPUObject or similar?

llvm/test/tools/llvm-dwarfdump/AMDGPU/known-amdgcn-relocs.yaml

1 ↗

(On Diff #498908)

I don't know whether this is a pattern you'll see widely in llvm-dwarfdump tests, but in newer tests I've been involved with, we've often used ## for comments, to distinguish them from the actual test logic.

7 ↗

(On Diff #498908)

Have you looked into using the yaml2obj support for Dwarf? I haven't used it much, but you should see some examples in the yaml2obj testing on how to use it. It may or may not be sufficient for your testing purposes.

Alternatively, what does using YAML here give you above using the assembly directly in the test?

68 ↗

(On Diff #498908)

I might be mistaken, but I believe this line is implicit, so can be omitted.

73 ↗

(On Diff #498908)

Rather than just labelling these as "Test 1" etc, it might be worth stating why they are interesting cases.

130–141 ↗

(On Diff #498908)

Nit: your indentation in the last line here is a little off, but in that case, you might as well remove the excessive spacing between key and value, so that there's just one space between the longest key and its value (and all the other values line up with each other).

llvm/test/tools/llvm-dwarfdump/AMDGPU/known-r600-relocs.yaml

1 ↗

(On Diff #498908)

Same comments as above.

llvm/test/tools/llvm-dwarfdump/AMDGPU/unknown-amdgcn-relocs.yaml

1 ↗

(On Diff #498908)

This test seems to be practically identical to the first one. Could you fold them into a separate test file and use yaml2obj's -D functionality to parameterise the YAML, so that you don't have to repeat it?

llvm/test/tools/llvm-dwarfdump/AMDGPU/unknown-r600-relocs.yaml

1 ↗

(On Diff #498908)

Same comments as other unknown test.

[scott.linder]

In D144301#4160749, @jhenderson wrote:

It looks like this code is testing code in libObject, but the testing is in llvm-dwarfdump. Should the tests be moved to test/Object?

I can add more in test/Object, but the change was originally about providing a better experience in llvm-dwarfdump, so it seemed like the right place. There are no unittests of the RelocationResolver APIs that I can find, which would be the other appropriate place.

llvm/lib/Object/RelocationResolver.cpp

262

Agreed, I changed the name to isAMDGPU, but retained the exposition in the doc-comment.

llvm/test/tools/llvm-dwarfdump/AMDGPU/known-amdgcn-relocs.yaml

7 ↗

(On Diff #498908)

I didn't realize it supported DWARF directly :^) thank you for the pointer

I moved everything into the YAML, let me know what you think

73 ↗

(On Diff #498908)

The description would just repeat the contents of the test; there is nothing particularly interesting about any case, I just threw multiple variations at each relocation kind to get some coverage.

I just deleted the redundant "Test #" comments

[scott.linder]

Address feedback

[jhenderson]

No need to move the tests unless you want to. Your logic for where they are is as good as anything else.

llvm/test/tools/llvm-dwarfdump/AMDGPU/amdgcn-relocs.yaml

1

Does this test need renaming, since it covers R600 too?

17

Nit: I'd get rid of the extra blank line (but I would have one before the start of the YAML).

18–19

Up to you, but you could change these to be shared with the R600 cases too, since the message is the same.

That being said, why are you checking them?

138–141

Again, I'd get rid of the extra blank line, and then add one after the CHECK, but I don't feel strongly about it.

[scott.linder]

llvm/test/tools/llvm-dwarfdump/AMDGPU/amdgcn-relocs.yaml

18–19

AFAICT there are no other tests of llvm-dwarfdump with amdgpu code objects, so I wanted to capture as much of the current behavior as possible, including the warning about MCRegInfo that comes up when the object is an "Unknown" amdgpu arch. Ideally we would eventually have a more meaningful error message for the lay-user, or have some fallback behavior, but I didn't want to invest the time now.

[scott.linder]

Address feedback:

• Fix test name to reflect it testing both r600 and amdcn
• Common up some checks between r600 and amdgcn
• Adjust some whitespace to make test more readable

[jhenderson]

Looks good to me.

[scott.linder]

Rebase

llvm/lib/Object/RelocationResolver.cpp

static uint64_t resolveSparc64(uint64_t Type, uint64_t Offset, uint64_t S,
   case ELF::R_SPARC_UA32:
   case ELF::R_SPARC_UA64:
     return S + Addend;
   default:
     llvm_unreachable("Invalid relocation type");
   }
 }
 
+/// Returns true if \c Obj is an AMDGPU code object based solely on the value
+/// of e_machine.
+///
+/// AMDGPU code objects with an e_machine of EF_AMDGPU_MACH_NONE do not
+/// identify their arch as either r600 or amdgcn, but we can still handle
+/// their relocations. When we identify an ELF object with an UnknownArch,
+/// we use isAMDGPU to check for this case.
+static bool isAMDGPU(const ObjectFile &Obj) {
+  if (const auto *ELFObj = dyn_cast<ELFObjectFileBase>(&Obj))
+    return ELFObj->getEMachine() == ELF::EM_AMDGPU;
+  return false;
+}
+
 static bool supportsAmdgpu(uint64_t Type) {
   switch (Type) {
   case ELF::R_AMDGPU_ABS32:
   case ELF::R_AMDGPU_ABS64:
     return true;
   default:
     return false;
   }
if (Obj.getBytesInAddress() == 8) {
         return {supportsSystemZ, resolveSystemZ};
       case Triple::sparcv9:
         return {supportsSparc64, resolveSparc64};
       case Triple::amdgcn:
         return {supportsAmdgpu, resolveAmdgpu};
       case Triple::riscv64:
         return {supportsRISCV, resolveRISCV};
       default:
+        if (isAMDGPU(Obj))
+          return {supportsAmdgpu, resolveAmdgpu};
         return {nullptr, nullptr};
       }
     }
 
     // 32-bit object file
     assert(Obj.getBytesInAddress() == 4 &&
            "Invalid word size in object file");
 
if (Obj.isCOFF()) {
     case Triple::mips:
       return {supportsMips32, resolveMips32};
     case Triple::msp430:
       return {supportsMSP430, resolveMSP430};
     case Triple::sparc:
       return {supportsSparc32, resolveSparc32};
     case Triple::hexagon:
       return {supportsHexagon, resolveHexagon};
+    case Triple::r600:
+      return {supportsAmdgpu, resolveAmdgpu};
     case Triple::riscv32:
       return {supportsRISCV, resolveRISCV};
     case Triple::csky:
       return {supportsCSKY, resolveCSKY};
     default:
+      if (isAMDGPU(Obj))
+        return {supportsAmdgpu, resolveAmdgpu};
       return {nullptr, nullptr};
     }
   } else if (Obj.isMachO()) {
     if (Obj.getArch() == Triple::x86_64)
       return {supportsMachOX86_64, resolveMachOX86_64};
     return {nullptr, nullptr};
   } else if (Obj.isWasm()) {
     if (Obj.getArch() == Triple::wasm32)
       return {supportsWasm32, resolveWasm32};
     if (Obj.getArch() == Triple::wasm64)
       return {supportsWasm64, resolveWasm64};
     return {nullptr, nullptr};

llvm/test/tools/llvm-dwarfdump/AMDGPU/amdgcn-relocs.yaml

+## Tests llvm-dwarfdump handling of AMDGPU relocations. We provide a .debug_info
+## section with multiple DW_AT_high_pc entries (that's one of the attributes for
+## which relocations are resolved by llvm-dwarfdump) and we add a relocation for
+## each of them. The first YAML document represents an amdgcn code object, and
+## the second an r600 code object.
+
+# RUN: yaml2obj --docnum=1 -DMACH= %s \
+# RUN:   | llvm-dwarfdump - 2>&1 | FileCheck --check-prefixes=AMDGCN,AMDGCN-UNKNOWN %s
+# RUN: yaml2obj --docnum=1 -DMACH=EF_AMDGPU_MACH_AMDGCN_GFX803 %s \
+# RUN:   | llvm-dwarfdump - 2>&1 | FileCheck --check-prefixes=AMDGCN,AMDGCN-KNOWN %s
+
+# RUN: yaml2obj --docnum=2 -DMACH= %s \
+# RUN:   | llvm-dwarfdump - 2>&1 | FileCheck --check-prefixes=R600,R600-UNKNOWN %s
+# RUN: yaml2obj --docnum=2 -DMACH=EF_AMDGPU_MACH_R600_R600 %s \
+# RUN:   | llvm-dwarfdump - 2>&1 | FileCheck --check-prefixes=R600,R600-KNOWN %s
+
+
+# AMDGCN-UNKNOWN:   -: Error in creating MCRegInfo
+# AMDGCN-KNOWN-NOT: -: Error in creating MCRegInfo
+# AMDGCN: -:      file format elf64-amdgpu
+--- !ELF
+FileHeader:
+  Class:      ELFCLASS64
+  Data:       ELFDATA2LSB
+  OSABI:      ELFOSABI_AMDGPU_HSA
+  ABIVersion: 0x02
+  Type:       ET_REL
+  Machine:    EM_AMDGPU
+  Flags:      [[[MACH]]]
+DWARF:
+  debug_abbrev:
+    - Table:
+        - Code:     1
+          Tag:      DW_TAG_compile_unit
+          Children: DW_CHILDREN_no
+          Attributes:
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+  debug_info:
+    - Version:  4
+      AddrSize: 8
+      Entries:
+        - AbbrCode: 1
+          Values:
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+            - Value: 0x4242424242424242
+Sections:
+  - Name:         .rela.debug_info
+    Type:         SHT_RELA
+    Flags:        [ SHF_INFO_LINK ]
+    AddressAlign: 0x0000000000000008
+    Info:         .debug_info
+    Relocations:
+
+      # AMDGCN: DW_AT_high_pc (0x0000000000000001)
+      - Offset: 0x000000000000000C # 0xC + 8*0
+        Type:   R_AMDGPU_ABS64
+        Symbol: v1
+        Addend: 0x0
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0x0000000000000043)
+      - Offset: 0x0000000000000014 # 0xC + 8*1
+        Type:   R_AMDGPU_ABS64
+        Symbol: v1
+        Addend: 0x42
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0xffffffffffffffff)
+      - Offset: 0x000000000000001C # 0xC + 8*2
+        Type:   R_AMDGPU_ABS64
+        Symbol: v0
+        Addend: 0xffffffffffffffff
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0xffffffffffffffff)
+      - Offset: 0x0000000000000024 # 0xC + 8*3
+        Type:   R_AMDGPU_ABS64
+        Symbol: vffffffffffffffff
+        Addend: 0x0
+
+      # AMDGCN: DW_AT_high_pc (0x0000000000000001)
+      - Offset: 0x000000000000002C # 0xC + 8*4
+        Type:   R_AMDGPU_ABS32
+        Symbol: v1
+        Addend: 0x0
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0x0000000000000043)
+      - Offset: 0x0000000000000034 # 0xC + 8*5
+        Type:   R_AMDGPU_ABS32
+        Symbol: v1
+        Addend: 0x42
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0xffffffffffffffff)
+      - Offset: 0x000000000000003C # 0xC + 8*6
+        Type:   R_AMDGPU_ABS32
+        Symbol: v0
+        Addend: 0xffffffffffffffff
+
+      # AMDGCN-NEXT: DW_AT_high_pc (0xffffffffffffffff)
+      - Offset: 0x0000000000000044 # 0xC + 8*7
+        Type:   R_AMDGPU_ABS32
+        Symbol: vffffffffffffffff
+        Addend: 0x0
+
+Symbols:
+  - Name:    v0
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0x0
+  - Name:    v1
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0x1
+  - Name:    vffffffffffffffff
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0xffffffffffffffff
+...
+
+
+# R600-UNKNOWN:   -: Error in creating MCRegInfo
+# R600-KNOWN-NOT: -: Error in creating MCRegInfo
+# R600: -:      file format elf32-amdgpu
+--- !ELF
+FileHeader:
+  Class:   ELFCLASS32
+  Data:    ELFDATA2LSB
+  Type:    ET_REL
+  Machine: EM_AMDGPU
+  Flags:   [[[MACH]]]
+DWARF:
+  debug_abbrev:
+    - Table:
+        - Code:     0x00000001
+          Tag:      DW_TAG_compile_unit
+          Children: DW_CHILDREN_no
+          Attributes:
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+            - Attribute: DW_AT_high_pc
+              Form:      DW_FORM_addr
+  debug_info:
+    - Version: 4
+      AddrSize: 4
+      Entries:
+        - AbbrCode: 1
+          Values:
+            - Value: 0x42424242
+            - Value: 0x42424242
+            - Value: 0x42424242
+            - Value: 0x42424242
+Sections:
+  - Name:         .rela.debug_info
+    Type:         SHT_RELA
+    Flags:        [ SHF_INFO_LINK ]
+    AddressAlign: 0x0000000000000001
+    Info:         .debug_info
+    Relocations:
+
+      ## FIXME: Is R_AMDGPU_ABS64 meaningful here?
+
+      # R600: DW_AT_high_pc (0x00000001)
+      - Offset: 0x0000000C # 0xC + 4*0
+        Type:   R_AMDGPU_ABS32
+        Symbol: v1
+        Addend: 0x0
+
+      # R600-NEXT: DW_AT_high_pc (0x00000043)
+      - Offset: 0x00000010 # 0xC + 4*1
+        Type:   R_AMDGPU_ABS32
+        Symbol: v1
+        Addend: 0x42
+
+      ## FIXME: Why is this field printed as sign-extended 64-bit in a 32-bit executable?
+      # R600-NEXT: DW_AT_high_pc (0xffffffffffffffff)
+      - Offset: 0x00000014 # 0xC + 4*2
+        Type:   R_AMDGPU_ABS32
+        Symbol: v0
+        Addend: 0xffffffff
+
+      # R600-NEXT: DW_AT_high_pc (0xffffffff)
+      - Offset: 0x00000018 # 0xC + 4*3
+        Type:   R_AMDGPU_ABS32
+        Symbol: vffffffff
+        Addend: 0x0
+
+Symbols:
+  - Name:    v0
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0x0
+  - Name:    v1
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0x1
+  - Name:    vffffffff
+    Type:    STT_SECTION
+    Section: .debug_info
+    Value:   0xffffffff
+...

llvm/test/tools/llvm-dwarfdump/AMDGPU/lit.local.cfg

+if not 'AMDGPU' in config.root.targets:
+    config.unsupported = True