[ARM] Incorrect calling convention when invoking __sync builtins

[Amanieu]

IR

define void @foo(ptr %a, i32 %x) {
  %val = trunc i32 %x to i8
  %1 = atomicrmw max ptr %a, i8 %val seq_cst
  ret void
}

Asm

foo:
        push    {r11, lr}
        bl      __sync_fetch_and_max_1
        pop     {r11, lr}
        mov     pc, lr

The problem

The ARM calling convention specified that integers smaller than the word size need to be sign/zero-extended to the word size. __sync_fetch_and_max_1 expects a signed i8 for its second argument, which should be in sign-extended form. However LLVM isn't performing a sign-extension for this parameter, which leads to issues such as rust-lang/rust#100650.

[llvmbot]

@llvm/issue-subscribers-backend-arm

[brad0]

@llvm/issue-subscribers-backend-arm

[brad0]

@Amanieu Is this still an issue?

[Amanieu]

Yes: https://godbolt.org/z/r7fbP7W8h

The correct code should have a sign extension of r1.

[brad0]

cc @davemgreen @stuij @AlfieRichardsArm

[brad0]

cc @TNorthover @asl