In a previous post, I wrote about linker relaxation: the linker’s ability to replace a slower, larger instruction with a faster, smaller one when it has enough information at link time. For instance, an indirect call through the GOT can be relaxed into a direct call plus a nop. This is a well-known technique to optimize the instructions for performance.

Farid Zakaria’s Blog

When linking massive binaries that include Intel's prebuilt MKL static archive, relocation overflows occur because MKL's object files use LEA instructions with large negative addends that exceed the ±2 GiB limit of 32-bit PC-relative relocations. Since MKL can't be recompiled, a technique called 'linker pessimization' is introduced: the reverse of linker relaxation. Instead of shrinking instructions, a single opcode byte is changed (0x8D → 0x8B) to convert a LEA into a MOV, replacing a direct arithmetic computation with an indirect memory load through a nearby pointer slot. The pointer slot uses a 64-bit absolute R_X86_64_64 relocation that can address the full 64-bit address space without overflow. The fix requires three surgical changes to the object file: patching the opcode, creating a new .data.fixup section with the 64-bit pointer, and retargeting the original relocation. The trade-off is an extra memory load where pure arithmetic existed before.

Linker Pessimization