Leave the code unchanged under -c, and comment every instruction.

[glampert]

-c writes each instruction's source line above it as a ; Line N: comment. It had two problems:

1. -c changed the program. The comments went into the same list as the code, ahead of each instruction's padding. padForBranchPreBubble looks at the previous line to decide whether a branch needs a bubble. Under -c it found the comment instead of the padding that was already there, and added another nop pair. Quake II's textured_triangles got 5 more nop pairs with -c than without, and each of those branches' comment sat above the extra nop:

 ; Line 40:     ibeq   iFormat, vi00, lStrideSet
                    nop                             nop
                    nop                             ibeq VI07, VI00, lStrideSet

2. Some instructions had no comment. Instructions emitted as an upper/lower pair, or with waitq/waitp, got none: 103 of the 1001 instruction lines in textured_triangles.

The comments now live in their own list, each with the index of the code line it goes above, and write() interleaves them. Code generation only ever sees code. Each instruction line gets the comments of the source instructions on it, after any padding: both halves of a pair (upper first), an instruction issued with waitq/waitp, and a branch and its delay-slot filler. Only openvcl's own nop/waitq/waitp padding has no comment:

 ; Line 22:     sub.x fMinusOne, vf00, vf00[w]
 ; Line 13:     lq fMVP0,      0(vi00)
                    subw.x VF08, VF00, VF00w        lq VF01, 0(VI00)

The two lines that the experimental OPENVCL_USE_GENERIC_KERNEL_REWRITE path pushes directly are left as they are, since that code is being reworked on the optimizations branch.

Testing

• New test/integration/test_emit_source.cpp, 3 new test cases. Each compiles a program with and without -c, then checks that the -c output without its comments equals the plain output, and that every instruction has a comment above it.
• The full suite passes, also under ASan/UBSan.
• Across 48 VU programs (the ps2gl fixtures, the examples, Quake II and Tyra), the plain output is byte-identical to master. With -c the code no longer differs from the plain output in any of them (16 differed before), and every instruction from the source has its comment.
• Merges cleanly with Check that instruction immediates fit their encoding; Add warnings-as-errors option. #3 and Clearer diagnostics, no output after an error, and two UBSan fixes. #4; with all three merged, the full test suite passes.

🤖 Co-authored by Claude Code

[fjtrujy]

Nice catch!
Cheers

[glampert]

No worries @fjtrujy.

By the way, I also have a local ASan/UBSan setup that found the uninitialized variable issue from #4. Is that something you think would be valuable to add to the project's CI? I could send a change just with that.

Cheers!