Avoid hashing the key twice in `get_query()`.

[nnethercote]

For a single-threaded parallel compiler, this reduces instruction counts
across several benchmarks, by up to 2.8%.

The commit also adds documentation about Sharded's use of FxHasher.

r? @Zoxc

[nnethercote]

Here are the biggest instruction count changes for a parallel compiler with one thread.

packed-simd-check
        avg: -1.9%      min: -2.8%      max: -0.7%
ctfe-stress-3-check
        avg: -1.0%?     min: -1.4%?     max: -0.5%?
piston-image-check
        avg: -0.3%      min: -0.5%      max: -0.0%
serde-check
        avg: -0.3%      min: -0.5%      max: -0.1%
syn-check
        avg: -0.3%      min: -0.5%      max: -0.1%
regex-check
        avg: -0.2%      min: -0.5%      max: -0.1%
futures-check
        avg: -0.3%      min: -0.5%      max: -0.1%
webrender-check
        avg: -0.3%      min: -0.5%      max: -0.1%
ripgrep-check
        avg: -0.2%      min: -0.4%      max: -0.0%
cargo-check
        avg: -0.2%      min: -0.4%      max: -0.0%

Wall times were super-noisy as usual but appear to be slightly improved overall.

deeply-nested-check
        avg: 3.9%       min: -0.7%      max: 9.6%
coercions-check
        avg: -1.4%?     min: -7.1%?     max: 9.4%?
regression-31157-check
        avg: -1.2%      min: -7.6%      max: 3.3%
issue-46449-check
        avg: -0.4%      min: -4.9%      max: 7.4%
regex-check
        avg: 0.6%       min: -3.6%      max: 7.0%
html5ever-check
        avg: -1.8%      min: -7.0%      max: 0.8%
await-call-tree-check
        avg: 1.4%       min: -2.0%      max: 7.0%
unused-warnings-check
        avg: -2.6%      min: -6.5%      max: 0.2%
helloworld-check
        avg: 1.5%       min: -2.2%      max: 5.8%
syn-check
        avg: -0.1%      min: -4.0%      max: 5.1%
ripgrep-check
        avg: -0.2%      min: -4.5%      max: 3.5%
encoding-check
        avg: -3.7%      min: -4.2%      max: -2.9%
piston-image-check
        avg: -0.4%      min: -4.2%      max: 1.8%
tuple-stress-check
        avg: -2.7%      min: -4.1%      max: -2.0%
unify-linearly-check
        avg: 1.2%       min: -3.4%      max: 3.8%
webrender-check
        avg: -1.5%      min: -3.3%      max: 0.8%
token-stream-stress-check
        avg: -2.3%      min: -2.8%      max: -1.8%
deep-vector-check
        avg: -0.4%      min: -2.6%      max: 1.5%
futures-check
        avg: -1.3%      min: -2.4%      max: -0.1%
unicode_normalization-check
        avg: -1.7%      min: -2.2%      max: -1.3%
ctfe-stress-3-check
        avg: -1.7%?     min: -2.1%?     max: -1.4%?
wg-grammar-check
        avg: 0.3%       min: -1.3%      max: 1.6%
inflate-check
        avg: 0.3%       min: -0.5%      max: 1.4%
packed-simd-check
        avg: -1.1%      min: -1.4%      max: -0.6%
keccak-check
        avg: -0.4%      min: -1.3%      max: 0.4%
clap-rs-check
        avg: 0.7%       min: -0.4%      max: 1.3%
webr_api-check
        avg: -0.7%      min: -1.0%      max: -0.2%
ucd-check
        avg: -0.3%      min: -0.9%      max: 1.0%
cargo-check
        avg: 0.6%       min: 0.4%       max: 0.9%
cranelift-codegen-check
        avg: -0.3%      min: -0.8%      max: -0.0%
serde-check
        avg: -0.2%      min: -0.6%      max: 0.1%

[Zoxc]

Did you check performance for the non-parallel compiler?

[nnethercote]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit 6fb6a7b with merge 9b89036...

[bors]

☀️ Try build successful - checks-azure
Build commit: 9b89036 (9b890360102b94f2008fbbfad2c9ea4122487735)

[rust-timer]

Queued 9b89036 with parent aa4e57c, future comparison URL.

[rust-timer]

Finished benchmarking try commit 9b89036, comparison URL.

[nnethercote]

Non-parallel performance changes are minor. ctfe-stress-3 is slightly worse, various others are tiny bit improved.

[nnethercote]

I have updated the comments as requested.

[JohnCSimon]

Ping from triage:
@Zoxc Can you please review/merge this PR?
cc: @nnethercote

[Zoxc]

@bors r+

[bors]

📌 Commit 1aceaaa has been approved by Zoxc

[andjo403]

@nnethercote was there some reason that the calculated hash was not used on line 107 also?
can the same trick be used in the complete
function?

[nnethercote]

@andjo403: I think it would work in both cases, but those code locations are much less hot, so the potential benefit is small.