Missed optimization: trunc not de-duplicated

[Mark-Simulacrum]

This Rust program:

use std::cell::Cell;

thread_local! {
    pub static FOO: Cell<u32> = Cell::new(0xffff);
}

pub fn bar() -> u32 {
    FOO.get() + FOO.get()
}

results in (at least) a duplicate load + trunc:

First version:

  %0 = load i32, ptr @"_ZN7example3FOO29_$u7b$$u7b$constant$u7d$$u7d$28_$u7b$$u7b$closure$u7d$$u7d$3VAL17h49f9a1072bb59151E", align 4, !range !3, !noalias !4, !noundef !11
  %trunc.i.i.i.i.i = trunc nuw i32 %0 to i1

Second version:

  %self3.val.pre.i.i = load i32, ptr getelementptr inbounds nuw (i8, ptr @"_ZN7example3FOO29_$u7b$$u7b$constant$u7d$$u7d$28_$u7b$$u7b$closure$u7d$$u7d$3VAL17h49f9a1072bb59151E", i64 4), align 4
  %1 = trunc nuw i32 %0 to i1

At minimum it'd be great to eliminate the double trunc, but combining the two loads also seems worthwhile (it should be possible AFAICT to deduplicate those). That would avoid two TLS address lookups in the resulting IR and should allow for further simplification (e.g., dropping some branches).

define noundef i32 @_ZN7example3bar17h6628f733c3c7a49fE() unnamed_addr #0 personality ptr @rust_eh_personality {
start:
  %0 = load i32, ptr @"_ZN7example3FOO29_$u7b$$u7b$constant$u7d$$u7d$28_$u7b$$u7b$closure$u7d$$u7d$3VAL17h49f9a1072bb59151E", align 4, !range !3, !noalias !4, !noundef !11
  %trunc.i.i.i.i.i = trunc nuw i32 %0 to i1
  br i1 %trunc.i.i.i.i.i, label %"_ZN3std6thread5local17LocalKey$LT$T$GT$4with17h35c29327e034454aE.exit", label %"_ZN3std6thread5local17LocalKey$LT$T$GT$4with17h35c29327e034454aE.exit6.sink.split", !prof !12

"_ZN3std6thread5local17LocalKey$LT$T$GT$4with17h35c29327e034454aE.exit": ; preds = %start
  %self3.val.pre.i.i = load i32, ptr getelementptr inbounds nuw (i8, ptr @"_ZN7example3FOO29_$u7b$$u7b$constant$u7d$$u7d$28_$u7b$$u7b$closure$u7d$$u7d$3VAL17h49f9a1072bb59151E", i64 4), align 4
  %1 = trunc nuw i32 %0 to i1
...

Godbolt: https://rust.godbolt.org/z/dPb4bPTTe

[nikic]

I'm somewhat confused. Aren't these truncs of two different loads, at two different offsets? How can we combine them?

Deduplication of the __tls_get_addr calls needs a Rust-side fix, see rust-lang/rust#136044. And I just noticed that it has a PR waiting on my review ... duh.

[Mark-Simulacrum]

Aren't these truncs of two different loads, at two different offsets?

I don't think so? Both are trunc'ing the %0 SSA value unless I'm missing something?

  %trunc.i.i.i.i.i = trunc nuw i32 %0 to i1
...
  %1 = trunc nuw i32 %0 to i1

The second is in a different basic block but it has a single predecessor which contains the first trunc, and that basic block (start) has only the start of the function as its entry point.

[nikic]

Uh, you're right, sorry the confusion.

[nikic]

I believe this is already fixed on main due to some improvements to KnownBits analysis for trunc nuw. InstCombine is now sufficient to eliminate the second trunc, without requiring an EarlyCSE/GVN run: https://llvm.godbolt.org/z/rE1fKGxhf