Dead code/instructions generated to detect divide by zero

[vitalyd]

This is born out of https://users.rust-lang.org/t/division-by-nonzero/12822.

The gist is the following function:

pub fn d1(x: u64, y: u64) -> u64 { 
  x / y.saturating_add(1) 
}

generates the following x86_64 assembly:

example::d1:
  addq $1, %rsi
  movq $-1, %rcx
  cmovaeq %rsi, %rcx
  testq %rcx, %rcx
  je .LBB0_2
  xorl %edx, %edx
  movq %rdi, %rax
  divq %rcx
  retq
.LBB0_2:
  pushq %rbp
  movq %rsp, %rbp
  leaq panic_loc.2(%rip), %rdi
  callq core::panicking::panic@PLT

str.0:
  .ascii "/tmp/compiler-explorer-compiler117811-5-86yrz4.tvuxewb3xr/example.rs"

str.1:
  .ascii "attempt to divide by zero"

https://godbolt.org/g/8eh61N

As far as I can tell, that division cannot panic. If that's true, then the following inefficiences are present:

1. The panic basic block that's generated.
2. The testq and je instructions are essentially dead.

I don't know the innards of LLVM, but perhaps the non-zero/non-wrapping aspect isn't conveyed to LLVM here? Or perhaps it's not able to figure it out on its own based on whatever IR is given to it.

[Thiez]

It seems LLVM fails to optimize this in C++ as well, see godbolt. GCC appears to be smarter.

[nikic]

I've submitted https://reviews.llvm.org/D55735 to fix this, conditional on us switching to use saturating add intrinsics (#55286).

[scottmcm]

Confirmed that the new add.sat would fix this: https://play.rust-lang.org/?version=nightly&mode=release&edition=2018&gist=2aae4b5b1f9d38d5a81d71bdd3425372