dead-code optimize `if const { expr }` even in opt-level=0

[scottmcm]

inline_const tracking issue: #76001

Inspired by @oli-obk's comment in #85828 (comment)

For things like if const { expr } { 0 } else { do something complex } it'd be nice to not have to even give LLVM the code for the else branch to have to emit. Ought to save time and binary size.

Tentatively marked mir-opt, since we presumably still need to make the MIR for it, in order to borrow-check and such the unreachable code. But might need to be codegen-time instead to pick up on monomorphized inline constants.

Hopefully should be doable relatively easily when the const is an inline const, since that forces const folding already, and thus there should just be a const true in the MIR.

[scottmcm]

Well, turns out this already works for simple constants in MIR! 🎉
https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=4d1fe08eec8b705525535ec9ec9d1c89

However, the place I actually needed it was more like if const { size_of::<T>() < 100 }, which doesn't work with inline_const today. So maybe come back to this once that's fixed.
https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=80c275a75dc280ac3fcdfb24b6484adc

[nagisa]

This should be done by SimplifyBranches which AFAIK is run unconditionally. I wonder if its match needs to be expanded?

Is there mcve?

[oli-obk]

The problem is that this can't be done on generic constants, as we don't mir-optimize post-monomorphization. We could start doing some benchmarks to see how bad running all mir opts (or at least a select few) would be if we duplicated, substituted and optimized MIR bodies for concrete generic parameters.

[oli-obk]

I would say

trait SizeBiggerFour: Sized {
    const VALUE: bool = std::mem::size_of::<Self>() > 4;
}

impl<T> SizeBiggerFour for T {}

pub fn foo<T>() -> i32 { if <T as SizeBiggerFour>::VALUE { 3 } else { 5 } }
pub fn bar() -> i32 { foo::<isize>() }

qualifies as an MCVE as the monomorphized foo::<isize> gives the following llvm ir

; playground::foo
; Function Attrs: nonlazybind uwtable
define i32 @_ZN10playground3foo17hb9992b01cb0c258eE() unnamed_addr #0 !dbg !6 {
start:
  %0 = alloca i32, align 4
  br i1 true, label %bb1, label %bb2, !dbg !15

bb2:                                              ; preds = %start
  store i32 5, i32* %0, align 4, !dbg !16
  br label %bb3, !dbg !15

bb1:                                              ; preds = %start
  store i32 3, i32* %0, align 4, !dbg !17
  br label %bb3, !dbg !15

bb3:                                              ; preds = %bb2, %bb1
  %1 = load i32, i32* %0, align 4, !dbg !18
  ret i32 %1, !dbg !18
}

[the8472]

We could start doing some benchmarks to see how bad running all mir opts (or at least a select few) would be if we duplicated, substituted and optimized MIR bodies for concrete generic parameters.

Maybe it should also be evaluated together with polymorphization or const_evaluatable_checked since those can cut down on the duplication before applying further optimizations.

[scottmcm]

Bonus points if this can avoid monomorphizing functions called in the dead code too.

Demo extracted from core: https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=40cc11066ca1fa8462d0f0f403ee1d8f (cc #52051)

Would be great to not codegen all the downstream stuff that's only called in the else block of a br i1 true.

[nbdd0121]

However, the place I actually needed it was more like if const { size_of::<T>() < 100 }, which doesn't work with inline_const today. So maybe come back to this once that's fixed.

FYI this should work with #96557

[DemiMarie]

Totally wild (and probably bad) idea: could we avoid borrow-checking unreachable code? After all, the purpose of borrow-checking is to prevent memory unsafety at runtime, but unreachable code cannot be memory unsafe, because it never executes!