Inline assembly register reuse weirdness

[canndrew]

I'm not sure whether this is a bug or whether it's just my own ignorance when it comes to writing inline assembly but I'm having issues passing rust values into assembly code. I'm using Linux x86_64 and I've written a simple program that calls the exit syscall with the value 23

#![feature(asm)]

fn main() {
  unsafe {
    asm!(
      "
      mov $$60, %rax
      mov $$23, %rdi
      syscall
      "
      :
      :
      : "%rax", "%rdi"
    );
  }
}

This program works fine, I can rustc exit.rs; ./exit; echo $? and get 23. Now I want to make the asm code interface with rust. So I pass the 23 in from rust rather than hardcoding it in the assembly.

#![feature(asm)]

fn main() {
  unsafe {
    asm!(
      "
      mov $$60, %rax
      mov $0, %rdi
      syscall
      "
      :
      : "r"(23)
      : "%rax", "%rdi"
    );
  }
}

Once again, this works. Now I try to pass 23 in using a variable.

#![feature(asm)]

fn main() {
  let x = 23i;
  unsafe {
    asm!(
      "
      mov $$60, %rax
      mov $0, %rdi
      syscall
      "
      :
      : "r"(x)
      : "%rax", "%rdi"
    );
  }
}

And this breaks. It exits with 60 rather than 23. Looking at the generated assembly shows why:

movq    $23, (%rsp)
movq    (%rsp), %rax
#APP

movq    $60, %rax
movq    %rax, %rdi
syscall

The compiler tries to save 23 into %rax but then %rax gets overwritten by 60 in the inline assembly. So is this a compiler bug? Or I am supposed to specifically tell rustc not to use %rax for for x? If it's useful, the LLVM IR for the failing code is:

; ModuleID = 'exit.0.rs'
target datalayout = "e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

; Function Attrs: uwtable
define internal void @_ZN4main20h2dd3717eeddb6da6eaaE() unnamed_addr #0 {
entry-block:
  %x = alloca i64
  store i64 23, i64* %x
  %0 = load i64* %x
  call void asm "\0A      mov $$60, %rax\0A      mov $0, %rdi\0A      syscall\0A      ", "r,~{%rax},~{%rdi},~{dirflag},~{fpsr},~{flags}"(i64 %0), !srcloc !0
  ret void
}

define i64 @main(i64, i8**) unnamed_addr #1 {
top:
  %2 = call i64 @_ZN2rt10lang_start20h6ebacfb5a732c9b9PfyE(i8* bitcast (void ()* @_ZN4main20h2dd3717eeddb6da6eaaE to i8*), i64 %0, i8** %1)
  ret i64 %2
}

declare i64 @_ZN2rt10lang_start20h6ebacfb5a732c9b9PfyE(i8*, i64, i8**) unnamed_addr #1

attributes #0 = { uwtable "split-stack" }
attributes #1 = { "split-stack" }

!0 = metadata !{i32 21}

[pczarn]

Use plain register names (rax rather than %rax) in the list of clobbers. LLVM doesn't warn about misspelled clobbers.

This code

#![feature(asm)]

fn main() {
  let x = 23i;
  unsafe {
    asm!(
      "
      mov $$60, %rax
      mov $0, %rdi
      syscall
      "
      :
      : "r"(x)
      : "rax", "rdi"
    );
  }
}

is translated to

  call void asm "\0A      mov $$60, %rax\0A      mov $0, %rdi\0A      syscall\0A      ", "r,~{rax},~{rdi},~{dirflag},~{fpsr},~{flags}"(i64 %0), !srcloc !0

[canndrew]

Use plain register names (rax rather than %rax) in the list of clobbers. LLVM doesn't warn about misspelled clobbers.

We have a winner! Thank you. Closing this non-bug now.