nondeterministic LLVM-IR with @anon.HASH.number

[fangism]

I am observing that repeating the same command twice with --emit=llvm-ir gives non-deterministic .rlib and .ll outputs.

[I apologize in advance for not providing a full set of repro inputs and commands right now, but I will work with my team on this.]

LLVM-IR differences look like this:

--- /home/me/run-1/libfoo_rust.ll.local  2021-11-24 01:39:16.232960224 +0000
+++ /home/me/run-2/libfoo_rust.ll.local  2021-11-24 01:41:24.957404136 +0000
@@ -864,8 +864,8 @@
 @alloc4102 = private unnamed_addr constant <{ [32 x i8] }> <{ [32 x i8] c"assertion failed: idx < CAPACITY" }>, align 1
 @alloc4103 = private unnamed_addr constant <{ [62 x i8] }> <{ [62 x i8] c"/b/s/w/ir/x/w/rust/library/alloc/src/collections/btree/node.rs" }>, align 1
 @alloc4104 = private unnamed_addr constant <{ i8*, [16 x i8] }> <{ i8* getelementptr inbounds (<{ [62 x i8] }>, <{ [62 x i8] }>* @alloc4103, i32 0, i32 0, i32 0), [16 x i8] c">\00\00\00\00\00\00\00r\02\00\00\09\00\00\00" }>, align 8
-@anon.15741daeb2a46d45fc5cbefe41b02e6d.0 = private unnamed_addr constant <{ [16 x i8] }> <{ [16 x i8] c"\08\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00" }>, align 8
-@anon.15741daeb2a46d45fc5cbefe41b02e6d.1 = private unnamed_addr constant <{ [16 x i8] }> <{ [16 x i8] c"\01\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00" }>, align 8
+@anon.95861573866d15199b79d9f6e49b68ab.0 = private unnamed_addr constant <{ [16 x i8] }> <{ [16 x i8] c"\08\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00" }>, align 8
+@anon.95861573866d15199b79d9f6e49b68ab.1 = private unnamed_addr constant <{ [16 x i8] }> <{ [16 x i8] c"\01\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00" }>, align 8
 @alloc868 = private unnamed_addr constant <{ [0 x i8] }> zeroinitializer, align 1
 @alloc4131 = private unnamed_addr constant <{ [4 x i8] }> <{ [4 x i8] c"Some" }>, align 1
 @vtable.y = private unnamed_addr constant <{ i8*, [16 x i8], i8*, [0 x i8] }> <{ i8* bitcast (void (%CapabilityName**)* @"_ZN4core3ptr48drop_in_place$LT$$RF$cm_rust..CapabilityName$GT$17he39a5a06256363e9E" to i8*), [16 x i8] c"\08\00\00\00\00\00\00\00\08\00\00\00\00\00\00\00", i8* bitcast (i1 (%CapabilityName**, %"core::fmt::Formatter"*)* @"_ZN42_$LT$$RF$T$u20$as$u20$core..fmt..Debug$GT$3fmt17h6a74665eeacccadaE" to i8*), [0 x i8] zeroinitializer }>, align 8, !dbg !61

Do we know why the hash value in @anon.HASH.n generated by rustc is non-deterministic? and what language construct is producing such anonymous symbols?

Meta

rustc --version --verbose:

rustc 1.58.0-nightly (c9c4b5d72 2021-11-17)

[fangism]

This should be labeled with A-reproducibility.

[Urgau]

Some reproducibility issue with LLVM have been fixed very recently. Can you check with the latest nightly ?

[tmandry]

This is separate from #90301 which was fixed recently. That was an issue where the output produced by LLVM was nondeterministic for the same LLVM IR. This is an issue where the LLVM IR itself is changing.

[fangism]

I'm still working on reducing a test case but wanted to add one more observation:

There has been an observable chance that running the same command twice does produce the same result, so it may take multiple attempts to see a difference. What these means for automatic reduction tools like creduce/delta is that it may encounter false-negatives that discard a reduction candidate. For example, I have a test case reduced to a state where its output is repeatable about 50% of the time.

[fangism]

Update: We've been able to track this down to a procedural macro whose implementation involves iterating over a HashMap. It was shown that this produces different code with -Zunpretty=expanded from run-to-run, which explains the LLVM IR differences.

[fangism]

Conclusion: The fault for this lies in the proc macro definition, not the tool itself.

That said, it would be nice if there were some way to detect this, be it statically or otherwise, in the interest of deterministic (and security-auditable) builds.