[MC] Fix HIP buildbot errors after 68472a3

[jplehr]

This fixes the errors on our HIP-Kokkos buildbot after 68472a3.

I do not know if this is the actually right fix, but afaict it restores the behavior for AMDGPU before the mentioned patch.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Jan Patrick Lehr (jplehr)

Changes

This fixes the errors on our HIP-Kokkos buildbot after 68472a3.

I do not know if this is the actually right fix, but afaict it restores the behavior for AMDGPU before the mentioned patch.

---

Full diff: https://github.com/llvm/llvm-project/pull/141138.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp (+10)

diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
index 19e394e5bc38c..0c98c32c21c2c 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp
@@ -18,6 +18,7 @@
 #include "llvm/MC/MCFixupKindInfo.h"
 #include "llvm/MC/MCObjectWriter.h"
 #include "llvm/MC/MCSubtargetInfo.h"
+#include "llvm/MC/MCValue.h"
 #include "llvm/MC/TargetRegistry.h"
 #include "llvm/Support/EndianStream.h"
 #include "llvm/TargetParser/TargetParser.h"
@@ -51,6 +52,8 @@ class AMDGPUAsmBackend : public MCAsmBackend {
 
   std::optional<MCFixupKind> getFixupKind(StringRef Name) const override;
   MCFixupKindInfo getFixupKindInfo(MCFixupKind Kind) const override;
+  bool shouldForceRelocation(const MCAssembler &, const MCFixup &,
+                             const MCValue &, const MCSubtargetInfo *) override;
 };
 
 } //End anonymous namespace
@@ -189,6 +192,13 @@ MCFixupKindInfo AMDGPUAsmBackend::getFixupKindInfo(MCFixupKind Kind) const {
   return Infos[Kind - FirstTargetFixupKind];
 }
 
+bool AMDGPUAsmBackend::shouldForceRelocation(const MCAssembler &,
+                                             const MCFixup &,
+                                             const MCValue &Target,
+                                             const MCSubtargetInfo *) {
+  return Target.getSpecifier();
+}
+
 unsigned AMDGPUAsmBackend::getMinimumNopSize() const {
   return 4;
 }

[jhuber6]

This follows the other targets that were changed, and it presumably fixes the regressions, so I'm going to assume it's correct for now.

[arsenm]

What kind of errors? This should probably include a test?

[jplehr]

We had runtime errors on the HIP bots. If this fixes these issues I'm happy to dig more into it and add some smaller test.
(It did fix the tests locally for me)

[MaskRay]

You may identify which relocatable file is different (llvm-objdump -dr) with this shouldForceRelocation change and then add a test to llvm/test/MC/AMDGPU. llvm/test/MC/Sparc/Relocations/relocation.s, which I recently cleaned up, has a good example

[MaskRay]

It is likely a PC-relative fixup. Could it be AMDGPU::fixup_si_sopp_br ?

AMDGPU also has GOTPCREL relocations similar to x86: s_mov_b32 s3, global_var1@gotpcrel32@lo. You may want a test that local@gotpcrel32@lo (where local is a local symbol defined in the current section) also forces a relocation. I improved the X86 GOT tests in a3d3931