The NonZero types don't tell LLVM that they're non-zero on get

[scottmcm]

Repro: https://play.rust-lang.org/?gist=0981601dda9c5800e353e31b22682bb5&version=nightly&mode=release

pub fn foo(x: std::num::NonZeroU32) -> bool {
    x.get() != 0
}

Actual:

  %0 = icmp ne i32 %x, 0
  ret i1 %0

Expected:

  ret i1 true

cc #49137

[hanna-kruppe]

In writing #50156 I found out that LLVM supports range metadata on call instructions. So you can have %x.2 = call i32 @NonZeroU32_get(i32 %x.1), !range !33 and it will use the range metadata to reason about the return value. If I add the range metadata manually to the IR for the above test case, the optimization indeed fires. So we can probably fix some of this (maybe it falls apart in more complex cases) by emitting range metadata for return values, which we currently don't. cc @eddyb

[hanna-kruppe]

Filed #50157 for adding metadata to calls, but this issue is more general -- once the get() gets inlined (or if it never gets called to begin with, e.g., because the u32 is accessed by type punning), we'd lose the information from the metadata on the call. The code in the OP is optimized because instcombine turns the compare into true before inlining happens, but in more complex cases #50156 might also help (though that only helps if the NonZero* is passed in as argument, it doesn't help with locals or if foo is inlined).

[hanna-kruppe]

Scratch that, range metadata on return values doesn't help with this issue since get returns a plain old u32. (thanks @nox)

[cuviper]

With debuginfo, we emit an alloca for %x that does have range metadata when loaded from. But it appears that is still not sufficiently propagated through optimization. https://rust.godbolt.org/z/sYjptQ

define internal i32 @_ZN4core3num10NonZeroU323get17h3d02812f86f2e907E(i32) unnamed_addr #0 !dbg !4 {
  %self = alloca i32, align 4
  store i32 %0, i32* %self, align 4
  call void @llvm.dbg.declare(metadata i32* %self, metadata !19, metadata !DIExpression()), !dbg !21
  %1 = load i32, i32* %self, align 4, !dbg !22
  ret i32 %1, !dbg !23
}

define zeroext i1 @_ZN7example3foo17h7a9a6aa0879e3714E(i32) unnamed_addr #1 !dbg !24 {
  %x = alloca i32, align 4
  store i32 %0, i32* %x, align 4
  call void @llvm.dbg.declare(metadata i32* %x, metadata !29, metadata !DIExpression()), !dbg !30
  %1 = load i32, i32* %x, align 4, !dbg !31, !range !32
  %2 = call i32 @_ZN4core3num10NonZeroU323get17h3d02812f86f2e907E(i32 %1), !dbg !31
  br label %bb1, !dbg !31

bb1:                                              ; preds = %start
  %3 = icmp ne i32 %2, 0, !dbg !31
  ret i1 %3, !dbg !33
}

[...]
!32 = !{i32 1, i32 0}

optimized:

define zeroext i1 @_ZN7example3foo17h7a9a6aa0879e3714E(i32) unnamed_addr #0 !dbg !4 {
  call void @llvm.dbg.value(metadata i32 %0, metadata !21, metadata !DIExpression()), !dbg !22
  %1 = icmp ne i32 %0, 0, !dbg !23
  ret i1 %1, !dbg !24
}