Invalid inline assembly is silently discarded

[devyn]

Minimal testcase:

#![feature(asm)]

fn main() {
    let byte = 0;
    let port = 0x80;

    unsafe { asm!("out %al, %dx" :: "a" (byte), "d" (port) :: "volatile"); }
}

rustc --version:

rustc 1.0.0-nightly (4db0b3246 2015-02-25) (built 2015-02-26)

rustc output:

$ rustc mintest.rs
$

(i.e., nothing)

Resulting objdump on x86_64-unknown-linux-gnu:

0000000000005070 <_ZN4main20h953519ae653eac04eaaE>:
    5070:       64 48 3b 24 25 70 00    cmp    %fs:0x70,%rsp
    5077:       00 00 
    5079:       77 1a                   ja     5095 <_ZN4main20h953519ae653eac04eaaE+0x25>
    507b:       49 ba 10 00 00 00 00    movabs $0x10,%r10
    5082:       00 00 00 
    5085:       49 bb 00 00 00 00 00    movabs $0x0,%r11
    508c:       00 00 00 
    508f:       e8 80 00 00 00          callq  5114 <__morestack>
    5094:       c3                      retq   
    5095:       48 83 ec 10             sub    $0x10,%rsp
    5099:       c7 44 24 0c 00 00 00    movl   $0x0,0xc(%rsp)
    50a0:       00 
    50a1:       c7 44 24 08 80 00 00    movl   $0x80,0x8(%rsp)
    50a8:       00 
    50a9:       8b 44 24 0c             mov    0xc(%rsp),%eax
    50ad:       8b 4c 24 08             mov    0x8(%rsp),%ecx
    50b1:       89 44 24 04             mov    %eax,0x4(%rsp)
    50b5:       89 0c 24                mov    %ecx,(%rsp)
    50b8:       48 83 c4 10             add    $0x10,%rsp
    50bc:       c3                      retq   
    50bd:       0f 1f 00                nopl   (%rax)

Unless I'm missing something, I should either get an error, or I should see "out %al, %dx" somewhere in there, but instead, absolutely nothing happens.

[devyn]

LLVM IR appears to include the inline asm, so I'm guessing this is an LLVM bug:

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

@const10 = internal unnamed_addr constant i32 0
@const14 = internal unnamed_addr constant i32 128

; Function Attrs: uwtable
define internal void @_ZN4main20h4fb6a247161ac946eaaE() unnamed_addr #0 {
entry-block:
  %byte = alloca i32
  %port = alloca i32
  store i32 0, i32* %byte
  store i32 128, i32* %port
  %0 = load i32* %byte
  %1 = load i32* %port
  call void asm sideeffect "out %al, %dx", "a,d,~{dirflag},~{fpsr},~{flags}"(i32 %0, i32 %1), !srcloc !0
  ret void
}

define i64 @main(i64, i8**) unnamed_addr #1 {
top:
  %2 = call i64 @_ZN2rt10lang_start20hb1b8686306685c6ftRLE(i8* bitcast (void ()* @_ZN4main20h4fb6a247161ac946eaaE to i8*), i64 %0, i8** %1)
  ret i64 %2
}

declare i64 @_ZN2rt10lang_start20hb1b8686306685c6ftRLE(i8*, i64, i8**) unnamed_addr #1

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

!0 = !{i32 1}

I'll check to see what clang does.

[devyn]

C source:

int main(int argc, char **argv)
{
    unsigned char  byte = 0;
    unsigned short port = 0x80;

    asm volatile("out %%al, %%dx" :: "a" (byte), "d" (port));
}

clang LLVM IR (clang -emit-llvm -S mintest.c):

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

; Function Attrs: nounwind uwtable
define i32 @main(i32 %argc, i8** %argv) #0 {
  %1 = alloca i32, align 4
  %2 = alloca i8**, align 8
  %byte = alloca i8, align 1
  %port = alloca i16, align 2
  store i32 %argc, i32* %1, align 4
  store i8** %argv, i8*** %2, align 8
  store i8 0, i8* %byte, align 1
  store i16 128, i16* %port, align 2
  %3 = load i8* %byte, align 1
  %4 = load i16* %port, align 2
  call void asm sideeffect "out %al, %dx", "{ax},{dx},~{dirflag},~{fpsr},~{flags}"(i8 %3, i16 %4) #1, !srcloc !1
  ret i32 0
}

attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
attributes #1 = { nounwind }

!llvm.ident = !{!0}

!0 = metadata !{metadata !"clang version 3.5.1 (tags/RELEASE_351/final)"}
!1 = metadata !{i32 115}

So it looks like it might be a Rust bug after all; Rust should be generating {ax} and {dx} constraints in the IR?

[devyn]

Okay, so I looked at this again today:

asm!("out %al, %dx" :: "{ax}" (byte), "{dx}" (port) :: "volatile");

seems to work just fine. The relevant LLVM documentation is for struct llvm::InlineAsm::SubConstraintInfo, field Codes where it specifies that the register name may come in braces, but doesn't really say anything more specific than that.

If Rust isn't receiving an error from LLVM about the invalid constraint, that's an LLVM bug, right? Still, I would like Rust documentation to be a little more specific about what constraints are allowed in asm! and perhaps for rustc to verify them, if LLVM won't.

[thepowersgang]

This issue appears to have returned with the recent LLVM update. If the data type passed to the inline assembly does not match the register size, the assembly is not emitted into assembly.

#![feature(asm)]

fn main() {
    unsafe {
        // - Errors, as expected
        //asm!("lttr %cx" : : "{ecx}" (7*8));
        // No error, no emitted instruction.
        asm!("lttr %cx" : : "{rcx}" (7*8));
    }
}

[jethrogb]

This is still a problem. Title of this bug should be: Invalid inline assembly is silently discarded.

[devyn]

Updated.

[Mark-Simulacrum]

Minimal test case from top comment produces, so presumably this is fixed. However, output files (rust_out.0.o and rust_out.crate.metadata.o) are still produced despite the error, so I'm not sure if this should be closed quite yet. If yes, please do so.

error: couldn't allocate input reg for constraint 'a'
 --> <anon>:7:14
  |
7 |     unsafe { asm!("out %al, %dx" :: "a" (byte), "d" (port) :: "volatile"); }
  |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error

Edit: This probably means #26648 is also fixed, but I'm not sure...

[Mark-Simulacrum]

If anyone can reproduce, please reopen. I haven't been able to reproduce since the original source doesn't quite compile--we should open another issue about incomplete compiles leaving files behind.

[jethrogb]

Might be good to add a regression test?

[Mark-Simulacrum]

Hmm, yeah. I'll reopen and mark as E-needstest.

[varkor]

There seem to still be issues with this test case on armhf-gnu: #59046 (comment).

[Amanieu]

This issue does not apply to the new asm! (RFC 2850) which validates the operand register classes.

The legacy llvm_asm! is deprecated and is no longer maintained.