asm_goto_with_outputs miscompilation #137867
Description
This should return 7, but returns a, with or without optimizations.
#![feature(asm_goto)] #![feature(asm_goto_with_outputs)] #[unsafe(no_mangle)] pub fn asm_goto_test( mut a: i16, ) -> i16 { unsafe { std::arch::asm!( "jmp {op}", inout("eax") a, op = label { a = 7; }, options(nostack,nomem) ); a } }https://godbolt.org/z/Efze66G84
As far as I can tell, the MIR looks correct:
fn asm_goto_test(_1: i16) -> i16 { debug a => _1; let mut _0: i16; bb0: { asm!("jmp {1}", inout("ax") copy _1 => _1, label 1, options(NOMEM | NOSTACK)) -> [return: bb1, label: bb2, unwind unreachable]; } bb1: { _0 = copy _1; return; } bb2: { _1 = const 7_i16; goto -> bb1; } }In particular, the value assigned to _1 in bb2 is returned in bb1.
The LLVM-IR is incorrect, as bb2 stores the value to ptr %a, but bb1 then overwrites that with the previous value in %2.
define i16 @asm_goto_test(i16 %0) unnamed_addr { start: %a = alloca [2 x i8], align 2 store i16 %0, ptr %a, align 2 %1 = load i16, ptr %a, align 2 %2 = callbr i16 asm sideeffect inteldialect "jmp ${2:l}", "=&{ax},0,!i,~{dirflag},~{fpsr},~{flags}"(i16 %1) #1 to label %bb1 [label %bb2] bb2: store i16 %2, ptr %a, align 2 store i16 7, ptr %a, align 2 br label %bb1 bb1: store i16 %2, ptr %a, align 2 %_0 = load i16, ptr %a, align 2 ret i16 %_0 }rustc --version --verbose:
rustc 1.86.0-nightly (f85c6de55 2025-01-26) binary: rustc commit-hash: f85c6de55206dbee5ffedfd821df1503a7b92346 commit-date: 2025-01-26 host: x86_64-unknown-linux-gnu release: 1.86.0-nightly LLVM version: 19.1.7 asm_goto tracking issue: #119364
@rustbot label +F-asm +A-inline-assembly +T-compiler