Missed optimization (extra branch) using slice::get with enum elements with spare niches

[jrose-signal]

The following code

#[derive(Clone, Copy)]
pub enum Outer {
    A([u8; 8]),
    B([u8; 8]),
}

pub struct Error(u8);

pub fn test(x: usize) -> Result<Outer, Error> {
    [Outer::A([10; 8]), Outer::B([20; 8])].get(x).copied().ok_or(Error(5))
}

generates the following assembly

playground::test:
	movq	%rdi, %rax
	cmpq	$1, %rsi
	ja	.LBB0_2
	leaq	(%rsi,%rsi,8), %rsi
	leaq	.L__unnamed_1(%rip), %rcx
	movzbl	(%rcx,%rsi), %edx
	cmpb	$2, %dl
	jne	.LBB0_3

.LBB0_2:
	movw	$1282, (%rax)
	retq

.LBB0_3:
	addq	%rsi, %rcx
	movq	1(%rcx), %rcx
	movb	%dl, (%rax)
	movq	%rcx, 1(%rax)
	retq

.L__unnamed_1:
	.ascii	"\000\n\n\n\n\n\n\n\n\001\024\024\024\024\024\024\024\024"

using stable (1.71.0).

But the second branch is completely superfluous; it's not possible for a load from the array to produce a None value.

[Jules-Bertholet]

@rustbot label A-codegen I-heavy I-slow

[krtab]

If we look at the version with only one element in the array:

#[derive(Clone, Copy)]
pub enum Outer {
    A([u8; 8]),
    B([u8; 8]),
}

pub struct Error(u8);

pub fn test(x: usize) -> Result<Outer, Error> {
    [Outer::A([10; 8])].get(x).copied().ok_or(Error(5))
}

we see that LLVM manages to optimize the second comparison icmpq comparison in one of the instcombine passes: https://godbolt.org/z/c7x1hjr1h

But that's about all I good figure out about this issue for now.

[XChy]

It seems to be subject to foldCmpLoadFromIndexedGlobal in InstCombine.
The final LLVM-IR includes:

  %5 = getelementptr inbounds %1, ptr @0, i64 %1
  %6 = load i8, ptr %5, align 1
  %7 = icmp eq i8 %6, 2

Here @0 is a constant, while the last %1 is not a constant, which InstCombine doesn't handle currently.
I found that rustc doesn't emit related IR about the global constant. And I'm new to rust compiler. Is there any way get the content of this constant?

[XChy]

Upstream issue: llvm/llvm-project#64238

[krtab]

The constant is defined line 5 of the optimized IR of your link:

@alloc_5234e787363bc08d7958dfb9bf9d8086 = private unnamed_addr constant <{ [18 x i8] }> <{ [18 x i8] c"\00\0A\0A\0A\0A\0A\0A\0A\0A\01\14\14\14\14\14\14\14\14" }>, align 1

The issue is that LLVM would need to figure that the two first lines you pasted can only lead to loading byte number 0 or 9, respectively equal to 0 and 1, and hence that the icmp is irrelevant.