introduce a "try" ZIR and AIR instruction

[andrewrk]

Implements #11772.

Merge Checklist:

• implement support in the backends
  • x86_64
  • wasm
  • aarch64
  • arm
  • riscv64 not covered by CI tests
  • sparc64 not covered by CI tests
  • C
• Sema: avoid emitting unused is_non_error AIR instruction

Analysis

Note: master branch at time of writing is e498fb1.

It did not close the bloat gap (#11498). File size stats after this change:

self-hosted binary size, LLVM backend, -Drelease, -Dstrip, LLVM disabled:
  master:
    stage3: 8370744 bytes
    stage2: 6963432 bytes
  this branch:
    stage3: 8343144 bytes
    stage2: 6939800 bytes

So it seems that this branch produces a slightly smaller executable (a delta of 25 KB) in both cases, but did not change the relative size difference.

However it did have a small improvement on how long it took to build in release mode with LLVM:

building self-hosted, LLVM backend, -Drelease, -Dstrip, LLVM disabled:
  master: 
    stage1 build stage2: 3m28.3s
    stage2 build stage3: 3m16.4s
    stage3 build stage3: 3m14.4s
  this branch:
    stage1 build stage2: 3m27.8s
    stage2 build stage3: 3m10.6s
    stage3 build stage3: 3m07.7s

Here's the measurement that should probably be used to decide whether this patch is overall an improvement:

building self-hosted, LLVM backend, debug mode, LLVM disabled:
  master: 
    stage1 build stage2: 0m34.3s
    stage2 build stage3: 0m33.1s
    stage3 build stage3: error: AccessDenied
  this branch:
    stage1 build stage2: 0m30.7s
    stage2 build stage3: 0m31.4s
    stage3 build stage3: error: AccessDenied

I keep getting error: AccessDenied intermittently when running stage3 and I have not diagnosed it yet. This is tracked in #11450. It's unfortunate here because the stage3 measurement is what we are most interested in.

Here's the same measurement but with -Denable-llvm which apparently does not trip the "access denied" error:

building self-hosted, LLVM backend, debug mode, -Denable-llvm
  master: 
    stage1 build stage2: 0m51.0s
    stage2 build stage3: 0m45.9s
    stage3 build stage3: 0m43.7s
  this branch:
    stage1 build stage2: 0m50.8s
    stage2 build stage3: 0m44.5s
    stage3 build stage3: 0m42.4s

Finally, let's look at perf of debug builds:

building self-hosted using debug mode, LLVM backend, debug mode, LLVM disabled:
  master:
    stage2 build stage3: 1m16.9s
    stage3 build stage3: 1m15.0s
  this branch:
    stage2 build stage3: 1m16.3s
    stage3 build stage3: 1m13.8s

Looking at the optimized LLVM IR, did buildOutputType get the desired cleanup pattern?

❌ no

Conclusion

Overall, this seems to have had a small, but noticeable impact. Given that it produces better runtime code for a common zig syntax in debug builds, I am inclined to press forward with this change.

[Luukdegram]

Implemented for Wasm backend in 7fb9650

[andrewrk]

This was a pretty big win for the C backend:

behavior test C backend line count:
       master: 72846 lines
  this branch: 60361 lines

[Luukdegram]

Some stats on those changes for the Wasm backend:

behavior test Wasm backend total instruction byte size (code section):
     master: 529749 bytes
this branch: 518855 bytes

This means an improvement of 10894 bytes worth of instructions.

[kubkon]

Some stats on those changes for the x64 backend:

behavior test total byte size (entire binary):
master: 2497585 bytes
this branch: 2484940 bytes

This means an improvement of 12645 bytes in total.

[joachimschmidt557]

Reduction in size for ARM (behavior test binary):

master: 1398205 bytes
stage2-try: 1344952 bytes

reduction: 53253 bytes = 0.0380866897200339 (about 3.8%)