llvm-project

400e4be5 - [JITLink][x86-64] Fix GOTPCRELX call/jmp relaxation to use PC-relative fixup (#190179)

[JITLink][x86-64] Fix GOTPCRELX call/jmp relaxation to use PC-relative fixup (#190179) The GOTPCRELX optimization in `optimizeGOTAndStubAccesses()` relaxes `call *foo@GOTPCREL(%rip)` → `addr32 call foo` and `jmp *foo@GOTPCREL(%rip)` → `jmp foo; nop`, but sets the edge kind to `Pointer32` (absolute). Since `e8`/`e9` are PC-relative instructions, `applyFixup` writes the absolute address instead of the displacement — producing a garbage target and SIGSEGV when JIT code is far from the callee (e.g., non-PIE executable with an arena allocator). **Fix:** - Guard: `TargetInRangeForImmU32` → `DisplacementInRangeForImmS32` (displacement must fit in signed 32-bit, not absolute address in unsigned 32-bit) - Edge kind: `Pointer32` → `BranchPCRel32` (so `applyFixup` writes `Target - (Fixup + 4) + Addend`) **Tests:** - Updated `ELF_got_plt_optimizations.s` to validate PC-relative displacement via `next_pc()` instead of raw absolute address. Also fixed the jmp test which was incorrectly checking `test_call_gotpcrelx` instead of `test_jmp_gotpcrelx`. - Added `ELF_gotpcrelx_no_relax.s`: regression test with slab at `0x7fff00000000` and extern at `0x00401000` — verifies relaxation is correctly skipped when displacement is out of int32 range. The `mov` → `lea` path (using `Delta32`) is unaffected and was already correct. cc: @tqchen Co-authored-by: Michael Buch <michaelbuch12@gmail.com>