[vchuravy]

Requested by Valentin

Enzyme writes the printed Julia type next to types it records in the IR:

• the enzymejl_parmtype_str parameter/return attribute (set_module_types!, lower_convention, specsig_function!),
• one enzymejl_source_type_<T> metadata kind per type (compile_unhooked_impl, memcpy_alloca_to_loadstore), which also registers a new metadata kind in the context for every distinct type,
• enzymejl_allocart_name.

Nothing reads these; they only help when reading IR dumps. The C++ side only copies enzymejl_allocart_name along with enzymejl_allocart when it replaces an alloca, and does nothing if it is missing.

Printing a type is slow on Julia 1.12 (show looks through module bindings for an alias of the type: make_typealias, jl_get_binding_partition), and Enzyme records a type for most loads, allocas and calls of every function in the module. In a profile of a TensorKit.jl test_reverse call this was the largest single item inside compile_unhooked_impl.

This PR emits the names only when Enzyme.Compiler.EmitTypeNames[] is set, like the existing Dump* debug flags.

Measurement

The Enzyme tests of TensorKit.jl (QuantumKitHub/TensorKit.jl#464, svd_trunc_no_error/lq_compact/qr_compact, Trivial and Z2 sectors, Float64 and ComplexF64, two tensor shapes; 72 test_reverse calls, each compiling one new thunk), Julia 1.12.7, both configurations run at the same time:

	total
main	472 s
this PR	427 s

That's 5-15% per call, and all tests passed in both runs.

Tests

Added a test that turns EmitTypeNames[] on and checks a gradient, so the debug path keeps compiling. optimize, basic, callconv and rooted_args pass locally on 1.12.

🤖 Generated with Claude Code

[github-actions]

Benchmark Results

	main	e42fd02...	main / e42fd02...
basics/make_zero/namedtuple	0.0436 ± 0.0038 μs	0.0433 ± 0.0037 μs	1.01 ± 0.12
basics/make_zero/struct	0.198 ± 0.0077 μs	0.198 ± 0.0071 μs	1 ± 0.053
basics/overhead	3.79 ± 0.01 ns	3.79 ± 0.01 ns	1 ± 0.0037
basics/remake_zero!/namedtuple	0.18 ± 0.0065 μs	0.183 ± 0.0066 μs	0.984 ± 0.05
basics/remake_zero!/struct	0.179 ± 0.0085 μs	0.18 ± 0.0077 μs	0.993 ± 0.064
fold_broadcast/multidim_sum_bcast/1D	0.436 ± 0.0081 μs	0.435 ± 0.0065 μs	1 ± 0.024
fold_broadcast/multidim_sum_bcast/2D	0.316 ± 0.0052 μs	0.316 ± 0.005 μs	1 ± 0.023
inline_abi/call	12.7 ± 0.019 ns	12.7 ± 0.01 ns	1 ± 0.0017
inline_abi/call_alloc	0.0368 ± 0.0011 μs	0.0366 ± 0.001 μs	1.01 ± 0.041
inline_abi/loop	3.52 ± 0.0036 μs	3.52 ± 0.0036 μs	1 ± 0.0015
time_to_load	1.28 ± 0.021 s	1.28 ± 0.0088 s	0.996 ± 0.017

Benchmark Plots

A plot of the benchmark results has been uploaded as an artifact at https://github.com/EnzymeAD/Enzyme.jl/actions/runs/37088582797/artifacts/11261497460.

[vchuravy]

Integration Tests - Flux - Julia 1.11 fails here with an EnzymeInternalError in the GRU, LSTM and Chain(LSTM, LSTM) testsets (inside broadcast_unalias → mightalias). That isn't this PR's doing: the same check fails the same way on #3709, which is also based on today's main and doesn't touch this code, and Flux passes on 1.10, 1.12 and 1.13 here. I don't know of a fix for it yet, so there's nothing to port.

---

Generated by Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 93.75000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 80.10%. Comparing base (8d7355b) to head (e42fd02).

Files with missing lines	Patch %	Lines
src/compiler/callconv.jl	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3711      +/-   ##
==========================================
+ Coverage   75.88%   80.10%   +4.21%     
==========================================
  Files          71       71              
  Lines       24486    24180     -306     
==========================================
+ Hits        18581    19369     +788     
+ Misses       5905     4811    -1094     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[vchuravy]

codecov/project is red for a -0.03% change (80.02% vs 126f916). The same check is red on #3709 against the same base, and the lines this PR gates are covered by the new EmitTypeNames test, so I'm leaving it as is.

---

Generated by Claude Code