Reuse one prepared frontend program for multiple emit outputs

[LunaStev]

Problem

Each Wave artifact emitter prepares the frontend again. Emitting AST, IR, bitcode, assembly and object for one source repeats tokenization five times, whereas object-only emission tokenizes once.

Starting points: src/cli.rs, src/runner.rs.

Reproduction and measurement

Tested at 6257ba3eb4b9f1ab6c247a920c6e0c4b5d699331 on Linux amd64 with LLVM 21.1.8 and a compiler built from that checkout. Build it with LLVM_SYS_211_PREFIX=/usr/lib64/llvm21 cargo build --locked --jobs 2.

Save fun main() -> i32 { return 0; } as repro.wave, then run:

target/debug/wavec build repro.wave --emit=obj --debug-wave=tokens --out-dir /tmp/wave-emit-one > /tmp/wave-emit-one.log
target/debug/wavec build repro.wave --emit=ast,ir,bc,asm,obj --debug-wave=tokens --out-dir /tmp/wave-emit-many > /tmp/wave-emit-many.log
rg -c '===== Tokens =====' /tmp/wave-emit-one.log /tmp/wave-emit-many.log

One run of each command showed 1 versus 5 token-stream dumps. Source inspection confirms each emitter calls frontend preparation again. These are invocation counts, not a measured speedup.

Results

Object-only: 1 frontend token stream. Combined five-artifact emit: 5 token streams. Output generation itself succeeds.

Scope

Share one prepared frontend result per source/target/options combination across emitters. Keep LLVM module/pass mutation concerns separate; reusing typed HIR must not accidentally share mutable backend state or cross configuration boundaries.

Difficulty: hard.

Completion criteria

• Prepare the frontend once for a combined emit request.
• Preserve artifacts and diagnostics compared with separate emits.
• Keep source identities/spans and target/options isolation intact.
• Measure preparation counts and, if claimed, compilation time with repeatable runs.