Check RUSTC_CTFE_BACKTRACE much less by generating fewer errors

[wesleywiser]

Before this change, get_size_and_align() calls get_fn_alloc() a
lot in CTFE heavy code. This previously returned an Error which would
check if RUSTC_CTFE_BACKTRACE was set on construction. Doing this
turned out to be a performance hotspot as @nnethercote discovered in
#68792.

This is an alternate take on that PR which resolves the performance
issue by generating many fewer errors. Previously, ctfe-stress-4
would generate over 5,000,000 errors each of which would check for the
presence of the environment variable. With these changes, that number is
reduced to 30.

r? @RalfJung

[wesleywiser]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit e9d7e4c5637d1584551727e34e6fc7ffb98bdad8 with merge 86fcc9b22b448b73a89589ccc348771cdd426a09...

[RalfJung]

Looks good, r=me when perf agrees

[RalfJung]

(But I think there's a typo in the PR description where it fails to link to where @nnethercote discovered the problem.)

[wesleywiser]

Heh. Vim word-wrapped the line as I was typing the issue number and since git-commit ignores lines that start with #, it removed the line.

[RalfJung]

Also, adding the flag to the session as you proposed would still probably be better than the current env var scheme. Even if this solves the main perf problem, there is no good reason to make error creation that expensive.

[bors]

☀️ Try build successful - checks-azure
Build commit: 86fcc9b22b448b73a89589ccc348771cdd426a09 (86fcc9b22b448b73a89589ccc348771cdd426a09)

[rust-timer]

Queued 86fcc9b22b448b73a89589ccc348771cdd426a09 with parent ae54678, future comparison URL.

[rust-timer]

Finished benchmarking try commit 86fcc9b22b448b73a89589ccc348771cdd426a09, comparison URL.

[RalfJung]

8% speedup, nice :)

@bors r+ rollup=never
(@nnethercote I assume as this has perf impact it should not be rolled up?)

[bors]

📌 Commit e9d7e4c5637d1584551727e34e6fc7ffb98bdad8 has been approved by RalfJung

[wesleywiser]

@RalfJung Yeah, that's correct. PRs that impact performance shouldn't be rolled up.

[nnethercote]

Nice perf results! I didn't realize this would help benchmarks other than ctfe-stress-4, that's good to see.

[nnethercote]

BTW, before this lands can you reword the commit message so that the # doesn't appear at the start of a line?

[RalfJung]

@bors r-
for commit msg ediitng

[wesleywiser]

@bors r=RalfJung

[bors]

📌 Commit 9f3bc82 has been approved by RalfJung

[bors]

⌛ Testing commit 9f3bc82 with merge 01a8b5f...

[bors]

☀️ Test successful - checks-azure
Approved by: RalfJung
Pushing 01a8b5f to master...