Tweak LEB128 reading some more.

[nnethercote]

PR #69050 changed LEB128 reading and writing. After it landed I did some
double-checking and found that the writing changes were universally a
speed-up, but the reading changes were not. I'm not exactly sure why,
perhaps there was a quirk of inlining in the particular revision I was
originally working from.

This commit reverts some of the reading changes, while still avoiding
unsafe code. I have checked it on multiple revisions and the speed-ups
seem to be robust.

r? @michaelwoerister

[nnethercote]

Local check results:

clap-rs-check
        avg: -0.9%      min: -1.9%      max: 0.0%
packed-simd-check
        avg: -0.3%      min: -0.8%      max: 0.0%
issue-46449-check
        avg: 0.4%       min: 0.3%       max: 0.6%
tuple-stress-check
        avg: -0.1%      min: -0.5%      max: 0.0%
wg-grammar-check
        avg: -0.2%      min: -0.5%      max: 0.0%
helloworld-check
        avg: 0.2%       min: -0.2%      max: 0.5%
keccak-check
        avg: -0.1%      min: -0.4%      max: 0.0%
webrender-check
        avg: -0.2%      min: -0.4%      max: 0.0%
regex-check
        avg: -0.2%      min: -0.4%      max: 0.1%
ripgrep-check
        avg: -0.1%      min: -0.4%      max: 0.1%
piston-image-check
        avg: -0.2%      min: -0.4%      max: 0.1%
unify-linearly-check
        avg: 0.1%       min: -0.3%      max: 0.4%
serde-check
        avg: -0.1%      min: -0.4%      max: 0.0%
cranelift-codegen-check
        avg: -0.1%      min: -0.4%      max: 0.0%
script-servo-check
        avg: -0.2%      min: -0.4%      max: 0.0%
style-servo-check
        avg: -0.1%      min: -0.3%      max: 0.0%
wf-projection-stress-65510-che...
        avg: -0.1%      min: -0.3%      max: 0.0%
coercions-check
        avg: 0.1%?      min: 0.0%?      max: 0.3%?
cargo-check
        avg: -0.1%      min: -0.3%      max: 0.0%
trait-stress-check
        avg: 0.1%       min: -0.0%      max: 0.3%
unused-warnings-check
        avg: -0.1%      min: -0.3%      max: 0.0%
deeply-nested-check
        avg: 0.1%       min: -0.3%      max: 0.3%
futures-check
        avg: -0.1%      min: -0.3%      max: 0.1%
tokio-webpush-simple-check
        avg: 0.1%       min: -0.3%      max: 0.3%
syn-check
        avg: -0.1%      min: -0.3%      max: 0.1%
hyper-2-check
        avg: -0.1%      min: -0.3%      max: 0.1%
webrender-wrench-check
        avg: -0.1%      min: -0.3%      max: 0.2%
await-call-tree-check
        avg: 0.1%       min: -0.3%      max: 0.2%
serde-serde_derive-check
        avg: -0.1%      min: -0.2%      max: 0.0%
encoding-check
        avg: -0.1%      min: -0.2%      max: 0.1%
unicode_normalization-check
        avg: -0.1%      min: -0.2%      max: 0.0%
html5ever-check
        avg: -0.1%      min: -0.2%      max: 0.1%
ucd-check
        avg: -0.1%      min: -0.2%      max: 0.0%
inflate-check
        avg: -0.0%      min: -0.2%      max: 0.0%
regression-31157-check
        avg: 0.0%       min: -0.1%      max: 0.2%
deep-vector-check
        avg: -0.0%      min: -0.1%      max: 0.0%
ctfe-stress-4-check
        avg: -0.0%?     min: -0.1%?     max: 0.1%?
token-stream-stress-check
        avg: -0.0%      min: -0.1%      max: 0.0%

[nnethercote]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit 6cc131c with merge 3c36579...

[bors]

☀️ Try build successful - checks-azure
Build commit: 3c36579 (3c3657919929d11ff9535e70167778577db99f0a)

[rust-timer]

Queued 3c36579 with parent 21ed505, future comparison URL.

[nnethercote]

The CI results show a clear regression, in contrast to my local results, hmm. I have rebased against a more recent revision. Let's try doing another perf CI run, just for interest's sake.

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit e25bd1f with merge 921540b...

[bors]

☀️ Try build successful - checks-azure
Build commit: 921540b (921540bfb9cb9af5713af43fc21417559eb5d218)

[rust-timer]

Queued 921540b with parent 5e7af46, future comparison URL.

[nnethercote]

So, we have two quite different sets of results when the same change is measured on top of different revisions. I'm going to abandon this PR for the following reasons.

• The instruction counts regression from the first run is a bit worse than the instruction count improvement from the second run.
• Both runs look like regressions if you look at the cycle counts.
• This PR makes the code uglier.