Serialize TypeTree offsets as signed constants

[wsmoses]

Fixes #3060 — Assertion \llvm::isUIntN(BitWidth, val) && "Value is not an N-bit unsigned value"'` on LLVM 23.

Root cause

TypeTree::toMD serializes offsets into !enzyme_type metadata:

llvm::ConstantInt::get(llvm::IntegerType::get(ctx, 32), pair.first)

pair.first is an int offset, and TypeTree uses -1 to mean "any offset". That converts to uint64_t 0xFFFFFFFFFFFFFFFF, which is not a 32-bit unsigned value.

LLVM changed the behavior under us:

	ConstantInt::get(IntegerType*, uint64_t V, bool IsSigned)
LLVM ≤ 21	APInt(BitWidth, V, isSigned, /*implicitTrunc=*/true) — silently truncates
LLVM ≥ 23	APInt(Ty->getBitWidth(), V, IsSigned, ImplicitTrunc=false) — asserts

Why it started firing now

This is a latent bug that only became reachable on this path in e4afe54 ("Speed up large constant ta"), which caches constant-global type analysis:

GV->setMetadata("enzyme_type", constTT.toMD(GV->getContext()));

Rust hits it on essentially every autodiff kernel, since its panic-location globals get a TypeTree containing -1 offsets (!{!"Unknown", i32 -1, ...}).

Backtrace from a local LLVM 23 build:

llvm::APInt::APInt(unsigned, unsigned long, bool, bool)
llvm::ConstantInt::get(llvm::IntegerType*, unsigned long, bool, bool)
TypeTree::toMD(llvm::LLVMContext&)
getConstantAnalysis(...)
TypeAnalyzer::getAnalysis / getCallInfo / visitIPOCall ...
EnzymeBase::HandleAutoDiff

Note this is not LLVM-23-only — LLVM 24 asserts identically. The opt21 link in the issue passes only because of the implicit truncation.

Fix

Pass IsSigned=true. This is bit-identical to the old truncating behavior (-1 → 0xFFFFFFFF) and matches the reader side, insertFromMD, which already uses getSExtValue().

Two other ConstantInt::get(Ty, -1) sites found by audit are fixed the same way (MustExitScalarEvolution.cpp, Utils.cpp:nextPowerOfTwo). Both would assert identically on LLVM ≥ 23 for sub-64-bit types; I could not construct an input that reaches either, so they carry no dedicated test.

Testing

New test test/Enzyme/ForwardMode/globalconsttypemd.ll is reduced from the issue reproducer and pins the emitted !enzyme_type offsets. Verified it fails (asserts) without the fix and passes with it, on LLVM 21, 23, and 24.

check-enzyme on LLVM 24, before vs. after:

	Passed	Failed	isUIntN aborts
baseline	287	752	30
with fix	290	749	0

The 30 aborts were spread across existing tests (ForwardMode/globalfn.ll, ReverseMode/storeconstexpr.ll, the ProbProg and omp* tests, …). The remaining failures are pre-existing CHECK-line drift for LLVM 23/24 IR printing, unrelated to this change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P65yn8LELKWU1dq6AQvA5f