rustc generates badly optimized for wrapping_div

[eefriedman]

pub fn f(x: i32, y: i32) -> i32 {
  x.wrapping_div(y)
}

Gives:

_ZN8rust_out1f17h164f16efff59de66E:
    pushq   %rax
.Ltmp0:
    cmpl    $-2147483648, %edi
    jne .LBB0_2
    movl    $-2147483648, %eax
    cmpl    $-1, %esi
    je  .LBB0_5
.LBB0_2:
    cmpl    $-1, %esi
    je  .LBB0_6
    testl   %esi, %esi
    jne .LBB0_4
    leaq    panic_loc7436(%rip), %rdi
    callq   _ZN4core9panicking5panic17h907815f47e914305E@PLT
.LBB0_6:
    cmpl    $-2147483648, %edi
    je  .LBB0_7
.LBB0_4:
    movl    %edi, %eax
    cltd
    idivl   %esi
.LBB0_5:
    popq    %rcx
    retq
.LBB0_7:
    leaq    panic_loc7440(%rip), %rdi
    callq   _ZN4core9panicking5panic17h907815f47e914305E@PLT

Specifically, .LBB0_7 is unreachable, but LLVM can't quite figure that out. It's probably possible to write wrapping_div in a way which avoids this problem.

[Aatch]

Not sure why LLVM doesn't optimise this more (the duplicated check of %esi against -1 should be easy enough to handle), but the good news is that the MIR-driven code generator produces this ASM instead:

    cmpl    $-2147483648, %edi
    jne .LBB0_2
    movl    $-2147483648, %eax
    cmpl    $-1, %esi
    je  .LBB0_3
.LBB0_2:
    movl    %edi, %eax
    cltd
    idivl   %esi
.LBB0_3:
    retq

Which is what is expected.

[nagisa]

Currently generated code looks just fine to me. Fixed.

[oyvindln]

Still looks fine it seems: (latest stable)

	cmpl	$-2147483648, %edi
	jne	.LBB0_2
	movl	$-2147483648, %eax
	cmpl	$-1, %esi
	je	.LBB0_4
.LBB0_2:
	testl	%esi, %esi
	je	.LBB0_5
	movl	%edi, %eax
	cltd
	idivl	%esi
.LBB0_4:
	popq	%rcx
	retq
.LBB0_5:
	leaq	panic_loc.2(%rip), %rdi
	callq	_ZN4core9panicking5panic17hec1812dcc135e139E@PLT

[nox]

Given this just needs a test, this could be transformed into an easy issue for newcomers.

[steveklabnik]

Triage; not aware of any test added. Still gives the good code when compiled.