[PATCH] D146862: [X86] Fix the incorrect displacement for prolog/epilog

Fri Mar 24 20:40:53 PDT 2023

LuoYuanke added inline comments.

================
Comment at: llvm/lib/Target/X86/X86FrameLowering.cpp:1590
     BuildStackAlignAND(MBB, MBBI, DL, StackPtr, MaxAlign);
-    int64_t Offset = Is64Bit ? -2 * (int64_t)SlotSize : -1 * (int64_t)SlotSize;
+    int64_t Offset = -1 * (int64_t)SlotSize;
     BuildMI(MBB, MBBI, DL, TII.get(Is64Bit ? X86::PUSH64rmm: X86::PUSH32rmm))
----------------
pengfei wrote:
> Curious why using 2 before. I was thinking it's related to stack aligement. AFAIK, Windows requires the stack always aligned to 16B on 64-bits.
> 
> Is better to directly `int64_t Offset = -SlotSize`?
Because my previous version is to push a callee saved register `push %rbx` as well, and then I realized that it is more complex on unwind support, and I turn to a scratch register which doesn't need to be saved and restore. But I forget to modify the offset. I'm developing a runtime test case to cover this issue.

================
Comment at: llvm/test/CodeGen/X86/i386-baseptr.ll:77
+; CHECK-NEXT:    leal -4(%ecx), %esp
 ; CHECK-NEXT:    .cfi_def_cfa %esp, 4
 ; CHECK-NEXT:    retl
----------------
pengfei wrote:
> The dwarf doesn't need to update together?
The dwarf seems correct. The ecx pointer to the base stack, and `esp = ecx - 4` pointer to the offset 4 which is the function return address.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D146862/new/

https://reviews.llvm.org/D146862