[llvm-dev] Mapping stack to IR on x86 backend

[Camille Mougey via llvm-dev]

Hi,

I'm currently working on a project in which I start from a semantic
description of an assembly basic block, and I want to output a semantically
similar basic block through LLVM.

For instance:

 - Semantic:
EAX = EBX + 2
ECX = EBX

 - Expected output:
lea eax, [ebx + 2]
mov ebx, ecx

For now, I have a working PoC using inline asm (for loading a register and
saving to it). I also load and save registers to avoid having them erased
with temporary values.

But I get an unexpected behavior using the ESP register. It seems that the
backend "ignores" the relation between ESP value and the stack.

For instance:

 - Semantic:

@32[ESP - 4] = EAX (that is to say, write at (ESP - 4) a 32 bits value, EAX)

- Output:

   0:    55                       push   %ebp
   1:    53                       push   %ebx
   2:    57                       push   %edi
   3:    56                       push   %esi
   4:    89 44 24 fc              mov    %eax,-0x4(%esp)
   8:    5e                       pop    %esi
   9:    5f                       pop    %edi
   a:    5b                       pop    %ebx
   b:    5d                       pop    %ebp

Which is obviously wrong.
The corresponding IR LLVM:

define void @"fc"()
{
entry:
  %".2" = call i32 asm  "", "={eax}"
()
  %".3" = call i32 asm  "", "={ecx}"
()
...
  %".6" = call i32 asm  "", "={esp}"
()
...
  %".18" = add i32 %".6", 4294967292
  %".19" = inttoptr i32 %".18" to i32*
  store i32 %".2", i32* %".19"
  call void asm sideeffect "", "{eax}"
(i32 %".2")
...
  call void asm sideeffect "", "{esp}"
(i32 %".6")
...
  ret void
}

I understand that LLVM is not intended to work with this kind of ASM mix,
but maybe I'm missing a feature or something.

A hack would be to use an `alloca` at the beginning of the first basic
block, assuming that it will be mapped relatively to ESP, and then
re-arranging my memory accesses based on it. But it seems to be a very
hacky way to do it...

If you have any idea, I will be happy to test it and provide you feedback.
Thanks in advance,

-- commial
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161219/b55abd08/attachment.html>