Finish and benchmark inline/mem2reg improvements.

[eddyb]

Background for these inliner and mem2reg changes:

• [Migrated] rustc_codegen_spirv taking a long time processing my (large) shader #113
  in early 2022 (by @hatoo) was the first sign that our mem2reg didn't scale well
• [Migrated] Inliner->mem2reg can quadratically amplify already-exponential inlining (turning seconds into minutes). #63
  in early 2024 (prompted by @schell) showed top-down inlining to be O(2ⁿ)
  (with mem2reg being quadratic on top of that, further amplifying the cost)
• most of the changes in this PR (those tagged [2024]) were developed in early 2024,
  and almost all of them landed a year later as part of Linker: speedup and debug info preservation #21
  (with WIP descriptions, and without the compile time impact being measured)
• rustc hangs for a few minutes since recent Rust toolchain upgrade #546
  more recently, still points to mem2reg performance leaving a lot to be desired
  (I wasn't able to reproduce that instance, however, as you'll see below)

---

In order to properly assess the impact of both already-landed changes, and newer ones,
this PR reverts commits from #21, for a "current main minus any improvements" baseline,
on top of which all changes are reapplied, allowing each step to be measured.
(this also results in a tiny effective diff, since most commits form revert+reapply pairs)

Sadly, @schell's renderling couldn't be included (due to indirect dependencies on specific glam versions), but the other old samples were simple enough to allow automating compatibility (with both Rust-GPU 0.9 and current main), and the measurement commands and raw output are available in a gist.

---

Commit	Description	rene@73e827b inline/mem2reg	rust-gpu-shadertoys@81a56fd inline/mem2reg	ab-proof-of-space-gpu@fc99af15 inline/mem2reg
c2f98b6	Last Rust-GPU release (v0.9)	1.4s/317.7s	15.7s/2747.7s
5556a28	main w/o inline/mem2reg changes (i.e. after all 5 reverts)	1.6s/283.3s	15.6s/1949.2s	0.9s/287.4s
5ae5fc7	↑+ inliner debuginfo fixes	8.7s/307.3s	97.6s/2220.4s	2.7s/298.8s
99d34af	↑+ bottom-up inliner	0.09s/49.9s	0.4s/622.5s	0.9s/213.9s
ccc08a2	↑+ mem2reg label ID indexing	0.09s/42.4s	0.4s/312.5s	0.9s/64.7s
1e02928	↑+ use phis to inline returns, instead of var+store+load [commit equivalent to main]	0.06s/3.9s	0.3s/3.9s	0.9s/4.8s
03d77ca	↑+ apply rewrite rules less often in mem2reg	0.07s/1.8s	0.4s/2.4s	0.9s/2.8s
1f18857	↑+ remove_duplicate_debuginfo during inlining	0.04s/1.2s	0.2s/1.7s	1s/2.7s
096fdcb	↑+ mem2reg during inlining	0.07s	0.7s	2.5s

---

While the changes with the largest impact have already landed in Rust-GPU, the last 3 commits still result in a combined 2-6x reduction in inlining+mem2reg times for the tested shaders (except for rene, where it's more like a 50x reduction).

(and that's without being able to reproduce #546 - until then, it's unclear how much of #547 is subsumed by the last 3 changes in this PR)

EDIT: see #552 (comment) and #547 (comment)

[LegNeato]

I rebased this on current main. I also did a perf run with this and this PR + cherry-picked #547. It is generally neutral overall, so this PR ate most of the benefit of that.

Details

Setup

• Baseline: c665c62c5d (this PR)
• Comparison: b92b3359ce (this PR + cherry-picked #547)

Median End-to-End Summary

workload	baseline total median (s)	+#547 total median (s)	delta
rene/rene-shader	2.937	2.790	-5.0%
rust-gpu-shadertoys/shaders	19.124	19.303	+0.9%
abundance/crates/farmer/ab-proof-of-space-gpu	26.206	26.122	-0.3%

Raw Results

rene/rene-shader

metric	baseline median	+#547 median	median delta	mean delta
link_inline	0.070	0.066	-5.7%	-9.9%
link_block_ordering_pass_and_mem2reg-after-inlining	0.018	0.016	-11.1%	-16.0%
link	2.754	2.495	-9.4%	-3.9%
total	2.937	2.790	-5.0%	-3.3%

rust-gpu-shadertoys/shaders

metric	baseline median	+#547 median	median delta	mean delta
link_inline	0.686	0.430	-37.3%	-27.5%
link_block_ordering_pass_and_mem2reg-after-inlining	0.097	0.040	-58.8%	-51.2%
link	18.506	18.680	+0.9%	+4.0%
total	19.124	19.303	+0.9%	+3.7%

abundance/crates/farmer/ab-proof-of-space-gpu

metric	baseline median	+#547 median	median delta	mean delta
link_inline	2.342	2.176	-7.1%	-10.1%
link_block_ordering_pass_and_mem2reg-after-inlining	0.293	0.244	-16.7%	-18.0%
link	25.991	25.905	-0.3%	+8.9%
total	26.206	26.122	-0.3%	+8.8%

[Firestar99]

Looks good, one question:

In the last commit f842afe mem2reg during inlining, you're measurements only have one value as if moved mem2reg to inlining, but you just added mem2reg to inlining without removing it whereever it was called previously.
So it should still have a (potentially small) runtime, or can it be removed from where it's called previously?

[eddyb]

I rebased this on current main. I also did a perf run ... It is generally neutral overall, so this PR ate most of the benefit of that.

Thanks, that's good to know. When it comes to rebasing, I will need to force-push a version that locally has the right commit identity for jj, sorry in advance (but I won't do that before updating the numbers).

---

In the last commit f842afe mem2reg during inlining, you're measurements only have one value as if moved mem2reg to inlining, but you just added mem2reg to inlining without removing it whereever it was called previously. So it should still have a (potentially small) runtime, or can it be removed from where it's called previously?

In theory the separate invocations of mem2reg might have become redundant, but I didn't want to risk them being skipped in case e.g. the inliner only runs mem2reg if inlining actually happened.

I could modify the separate mem2reg invocation to assert that nothing changes (i.e. we're only paying for a redundant rescanning), and if that never gets hit in anything we have access to, we can remove it.
(I was mainly trying to get big wins from small diffs, with minimal/zero risk of breakage)

---

Also see #547 (comment) for an update from @nazar-pc on the #546 mystery.
(I might re-measure on top of the Rust-GPU commit they were on, now that I have the automation for it)

[LegNeato]

Sure, force-push away. I guess I should probably switch to jj.