https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114843
--- Comment #17 from Wilco <wilco at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #16) > Patch posted with all of the testcases included: > https://gcc.gnu.org/pipermail/gcc-patches/2024-April/650080.html Not nearly enough testcases... What about: void g(void); int f(long offset, void *handler) { g(); if (offset > 5) return arr[offset]; __builtin_eh_return (offset, handler); } With -O2 -fomit-frame-pointer: f: .LFB0: .cfi_startproc stp x30, x0, [sp, -64]! .cfi_def_cfa_offset 64 .cfi_offset 30, -64 .cfi_offset 0, -56 stp x1, x2, [sp, 16] stp x3, x19, [sp, 32] .cfi_offset 1, -48 .cfi_offset 2, -40 .cfi_offset 3, -32 .cfi_offset 19, -24 mov x19, x0 str x20, [sp, 48] .cfi_offset 20, -16 mov x20, x1 bl g cmp x19, 5 ble .L8 mov w0, w19 ldp x19, x20, [sp, 40] ldp x30, x0, [sp], 64 ****** oops .cfi_remember_state .cfi_restore 0 .cfi_restore 30 .cfi_restore 19 .cfi_restore 20 .cfi_def_cfa_offset 0 ret .L8: .cfi_restore_state mov x5, x19 ldp x1, x2, [sp, 16] mov x6, x20 ldp x3, x19, [sp, 32] ldr x20, [sp, 48] ldp x30, x0, [sp], 64 .cfi_restore 0 .cfi_restore 30 .cfi_restore 20 .cfi_restore 3 .cfi_restore 19 .cfi_restore 1 .cfi_restore 2 .cfi_def_cfa_offset 0 add sp, sp, x5 br x6 .cfi_endproc So I don't believe you should change aarch64_pop_regs at all - it's too late to change things and just adds unnecessary complexity and more bugs. The best option would be to handle eh_return explicitly and insert the extra push/pops rather than treating them like a generic callee-save (because clearly they are not anymore).
