ELF/AMDGPU: Add support for R_AMDGPU_REL32 relocations

[ruiu]

ELF/Target.cpp

1435–1440 ↗

(On Diff #60540)

Then you can make this to

assert(Type == R_AMDGPU_REL32);
write32le(Loc, Val);

1444 ↗

(On Diff #60540)

This runs before relocateOne, so it is better to do error checking here rather than later.

if (Type == R_AMDGPU_REL32)
  return R_PC;
error("do not know how to handle relocation");

[ruiu]

Please update the commit message -- this handles not ABS32 but REL32.

[• rafael]

test/ELF/amdgpu-relocs.s

11 ↗

(On Diff #60540)

Do you really need all this just to generate a relocation?

[• tstellarAMD]

Add better error handling.

[• tstellarAMD]

ELF/Target.cpp

1440–1442 ↗

(On Diff #60787)

I made a slight modification to your suggestion to avoid "no return value" warnings for getRelExpr().

[ruiu]

LGTM. Please address Rafael's comment before committing.

lld/trunk/ELF/Target.cpp

if (Type == R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC) {
     // Generate MOVK.
     uint32_t RegNo = read32le(Loc) & 0x1f;
     write32le(Loc, (0xf2800000 | RegNo) | ((Val & 0xffff) << 5));
     return;
   }
   llvm_unreachable("invalid relocation for TLS IE to LE relaxation");
 }
 
-// Implementing relocations for AMDGPU is low priority since most
-// programs don't use relocations now. Thus, this function is not
-// actually called (relocateOne is called for each relocation).
-// That's why the AMDGPU port works without implementing this function.
 void AMDGPUTargetInfo::relocateOne(uint8_t *Loc, uint32_t Type,
                                    uint64_t Val) const {
-  llvm_unreachable("not implemented");
+  assert(Type == R_AMDGPU_REL32);
+  write32le(Loc, Val);
 }
 
 RelExpr AMDGPUTargetInfo::getRelExpr(uint32_t Type, const SymbolBody &S) const {
-  llvm_unreachable("not implemented");
+  if (Type != R_AMDGPU_REL32)
+    error("do not know how to handle relocation");
+  return R_PC;
 }
 
 ARMTargetInfo::ARMTargetInfo() {
   CopyRel = R_ARM_COPY;
   RelativeRel = R_ARM_RELATIVE;
   IRelativeRel = R_ARM_IRELATIVE;
   GotRel = R_ARM_GLOB_DAT;
   PltRel = R_ARM_JUMP_SLOT;

lld/trunk/test/ELF/amdgpu-relocs.s

+# RUN: llvm-mc -filetype=obj -triple=amdgcn--amdhsa -mcpu=fiji %s -o %t.o
+# RUN: ld.lld -shared %t.o -o %t.so
+# RUN: llvm-readobj -r %t.so | FileCheck %s
+
+# REQUIRES: amdgpu
+
+# Make sure that the reloc for local_var is resolved by lld.
+
+  .text
+
+kernel0:
+  s_mov_b32 s0, local_var+4
+  s_endpgm
+
+  .local local_var
+
+# CHECK: Relocations [
+# CHECK-NEXT: ]