Issues to report upstream

[nemequ]

I just synced wav with the recent changes in LLVM (from emscripten tot), and noticed several issues.

These issues are clearly in LLVM, but I don't want to bombard them with lots of issues they are already aware of. @tlively, do you want me to file these in LLVM's bug tracker?

• Missing __builtin_wasm_abs_i64x2
• Missing 64-bit shl and shr (there are no intrinsics, but shifting the vectors doesn't output the right instruction)
• any_true should be element-type agnostic
• Missing 64-bit all_true
• Pointers parameters to __builtin_wasm_load{32,64}_zero should be const, otherwise you may end up with an annoying -Wcast-qual diagnostic.

Some issues which may be in LLVM or WABT:

• wasm2wat says that code generated by __builtin_wasm_popcnt_i8x16 uses i16x8.extadd_pairwise_i8x16_s. wasm-dis says i8x16.popcnt. Reported as [SIMD] wasm2wat outputs incorrect instruction for i8x16.popcount WebAssembly/wabt#1653
• For 64-bit comparisons (using i64x2.eq for example), wasm2wat says cmpeq.wasm:000007a: error: type mismatch in function, expected [] but got [v128]
• Both wasm2wat and wasm-dis have problems with __bulitin_wasm_load8_lane (I don't think I tried other load*_lane functions):
  • wasm2wat: [parse exception: surprising value (at 0:4)]
  • wasm-dis: load_lane.wasm:0000076: error: alignment (0) must be a power of 2

I'm guessing the first two are WABT issues, the third one feels like LLVM.

I also have a lot places where I have to use the built-in functions because portable (within clang) versions don't generate the right instructions. I've kept the code around for now because I think they'll be helpful for improving the WASM codegen, but they're not really bugs so much as missing features. There are quite a few of these, just search for WAV_BUILTIN_MISSING_OPT in the current version of the header. If you want, I can go through and file a bunch of bugs similar to 49655 ("Masks on SIMD shift values should be elided")?

[tlively]

No need to file separate LLVM issues for the first chunk of problems, although this list is helpful.

Generally WABT should be more up-to-date than Binaryen; I haven't gotten around to updating Binaryen's numbering yet. Not sure what's up with the 64-bit comparision errors on wasm2wat, though. That one's probably worth filing a WABT bug on.

If you want, I can go through and file a bunch of bugs similar to 49655 ("Masks on SIMD shift values should be elided")?

That would be really great, thank you :)

[nemequ]

@tlively, I've finally started reporting the missed optimizations to LLVM, here is the first one: https://bugs.llvm.org/show_bug.cgi?id=50232

There are probably going to be a lot of these; maybe a couple dozen. Before I start going crazy filing the reports, is that one okay or is there something else I could do to help make things easier for you? Different phrasing, test cases as attachments instead of inline, maybe there is a format I could use which would make it easy to add test cases to LLVM (if you want to do that)…?

[tlively]

This is perfect, thanks @nemequ! The only thing I would add would be to assign the bugs to me directly.