llvm dev - Oct 2020 - Codegen bug in variable-sized alloca interaction with calling convention?

Hello,

It seems like if rbp register is not callee-saved (e.g. GHC calling
convention), LLVM still assumes that rbp is unchanged after a function
call, resulting in buggy assembly.

An example:
> define ghccc void @f(i64 %0) {
>   %2 = alloca i8, i64 %0, align 16
>   call ghccc void @g(i8* nonnull %2)
>   ret void
> }
> declare ghccc void @g(i8*)
>
The generated assembly is (llc test.bc -O3 -o test.o -filetype=obj):
> 0000000000000000 <f>:
>    0: 55                   push   %rbp
>    1: 48 89 e5             mov    %rsp,%rbp
>    4: 48 89 e0             mov    %rsp,%rax
>    7: 49 83 c5 0f          add    $0xf,%r13
>    b: 49 83 e5 f0          and    $0xfffffffffffffff0,%r13
>    f: 4c 29 e8             sub    %r13,%rax
>   12: 48 89 c4             mov    %rax,%rsp
>   15: 49 89 c5             mov    %rax,%r13
>   18: e8 00 00 00 00       callq  1d <f+0x1d>
> 19: R_X86_64_PLT32 g-0x4
>   1d: 48 89 ec             *mov    %rbp,%rsp*
>   20: 5d                   pop    %rbp
>   21: c3                   retq
>
In line '1d', rbp may have been garbaged by the call.

Is this a bug in LLVM?

Thanks,
Haoran
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201007/4b33573d/attachment.html>