[llvm] Spill/restore FP/BP around instructions in which they are clobbered (PR #81048)

[via llvm-commits]

weiguozhi wrote:

> Still, the CFI info looks wrong once the frame pointer is clobbered (updates to rsp _and_ rbp should be accompanied by matching CFI directives) and I have concerns that this will break unwinding (not just asynchronous unwinding but also exceptions thrown from the called function). Could you test that unwinding through such a function call works as expected?
> 
I'm very ignorant in DWARF. After some study, I think we need the following additional CFI instructions

```
foo:                                    # @foo
        .cfi_startproc
# %bb.0:
        pushq   %rbp
        .cfi_def_cfa_offset 16
        .cfi_offset %rbp, -16
        movq    %rsp, %rbp
        .cfi_def_cfa_register %rbp
        pushq   %r15
        pushq   %r14
        pushq   %r13
        pushq   %r12
        pushq   %rbx
        andq    $-16, %rsp
        subq    $16, %rsp
        .cfi_offset %rbx, -56
        .cfi_offset %r12, -48
        .cfi_offset %r13, -40
        .cfi_offset %r14, -32
        .cfi_offset %r15, -24
                                        # kill: def $edi killed $edi def $rdi
        pushq   %rbp
        pushq   %rax
+        .cfi_def_cfa_expression    [%rsp + 8]   + 16              //  %rbp is stored in memory [%rsp + 8]
        movl    %esi, %ebp
        movq    %rdi, %r13
        callq   external at PLT
        addq    $8, %rsp
        popq    %rbp
+        .cfi_def_cfa   %rbp, 16                                   // %rbp is restored
        leaq    -40(%rbp), %rsp
        popq    %rbx
        popq    %r12
        popq    %r13
        popq    %r14
        popq    %r15
        popq    %rbp
        .cfi_def_cfa %rsp, 8
        retq
```
And the cfi_def_cfa_expression should be constructed similar to function X86FrameLowering::emitCalleeSavedFrameMoves. Is my understanding right?

> (fwiw, I'm not a maintainer and I'd personally like to see support for this use case, but I have concerns that this will break other things.)

Because the frame pointer is clobbered, so the unwinding information is already broken. The current patch doesn't make it worse.

https://github.com/llvm/llvm-project/pull/81048