[compiler-v2][optimization] Create and retain temps for each arg

[vineethk]

Description

This PR builds on the optimization made in #15445 where we optimized Assign stackless-bytecode instructions. In this PR, when generating stackless-bytecode, we create temporaries for each argument (even if they are simple variables), retain those assignments (through dead-store elimination) if they represent cross-block def-uses. In this way, the file-format bytecode generator knows when to load the cross-block definitions (fixes #15339, by essentially implementing eager loads optimization in a different way). The flush writes optimization needs additional logic to handle this pipeline.

With the optimization in this PR, on the aptos-framework and all its dependencies (std libs):

• compiler v2 produces 3.6% less instructions than compiler v1.
• along with [compiler-v2] Optimize stackless-bytecode assign instructions #15445, v2 produces 4.7% less instructions than without these optimizations.

Note that all changes introduced here should be hidden behind OPTIMIZE_WAITING_FOR_COMPARE_TESTS experiment (which is turned off by default), so landing this PR should not affect code generated.

Safety arguments:

• assigning a new temp for each arg in a function call should be safe (because we already do this for non-trivial arguments)
• not removing certain deadcode of the form x = x should be safe
• flush writes optimization is purely a hinting mechanism which advises whether a definition loaded onto the stack should be flushed to a local write-away: so if it hints wrongly, we would have extra flushes to a local, but would not affect the execution semantics

A couple of examples of this optimization working:

This code:

    fun test1(x: u64) {
        bar(x, one(), one(), one(), one(), one());
    }

produces before this PR:

	0: Call one(): u64
	1: Call one(): u64
	2: Call one(): u64
	3: Call one(): u64
	4: Call one(): u64
	5: StLoc[1](loc0: u64)
	6: StLoc[2](loc1: u64)
	7: StLoc[3](loc2: u64)
	8: StLoc[4](loc3: u64)
	9: StLoc[5](loc4: u64)
	10: MoveLoc[0](Arg0: u64)
	11: MoveLoc[5](loc4: u64)
	12: MoveLoc[4](loc3: u64)
	13: MoveLoc[3](loc2: u64)
	14: MoveLoc[2](loc1: u64)
	15: MoveLoc[1](loc0: u64)
	16: Call bar(u64, u64, u64, u64, u64, u64)
	17: Ret

after this PR:

	0: MoveLoc[0](Arg0: u64)
	1: Call one(): u64
	2: Call one(): u64
	3: Call one(): u64
	4: Call one(): u64
	5: Call one(): u64
	6: Call bar(u64, u64, u64, u64, u64, u64)
	7: Ret

This code:

    fun test2(x: u64, y: u64): u64 {
        x + (y * x)
    }

produces before this PR:

	0: MoveLoc[1](Arg1: u64)
	1: CopyLoc[0](Arg0: u64)
	2: Mul
	3: StLoc[1](Arg1: u64)
	4: MoveLoc[0](Arg0: u64)
	5: MoveLoc[1](Arg1: u64)
	6: Add
	7: Ret

after this PR:

	0: CopyLoc[0](Arg0: u64)
	1: MoveLoc[1](Arg1: u64)
	2: MoveLoc[0](Arg0: u64)
	3: Mul
	4: Add
	5: Ret

How Has This Been Tested?

• New tests (like the ones described above added) added
• We show improvements on [compiler-v2] Test case reduced from move-stdlib showing opportunity for optimization #15338 and many of the tests added in [compiler-v2] Test cases reduced from the framework showcasing need for stack optimizations #14800
• Added file format generation for a lot of tests that did not have it, and checked that we created fewer or equal instructions in all those cases

I've gone through all the test output changes:

• In all existing tests, we produce fewer or equal instructions with the exception of minor regression in two tests (move-compiler-v2/tests/flush-writes/{def_use_03.on.exp, out_of_order_use_03.on.exp}. Filing an issue to follow up on these regressions.
• Some error messages have become more precise, some have become more abstract (due to the introduction of the additional temporaries).

Key Areas to Review

• Safety argument, correctness
• Code retains existing behavior when experiment is off

Type of Change

• Performance improvement

Which Components or Systems Does This Change Impact?

• Move Compiler

[trunk-io]

⏱️ 1h 39m total CI duration on this PR

Slowest 15 Jobs	Cumulative Duration	Recent Runs
forge-compat-test / forge	13m	🟩
rust-move-tests	13m	🟩
rust-move-tests	13m	🟩
rust-move-tests	13m	🟩
rust-move-tests	12m	🟩
rust-move-tests	12m	🟩
rust-cargo-deny	9m	🟩 🟩 🟩 🟩 🟩 (+1 more)
check-dynamic-deps	6m	🟩 🟩 🟩 🟩 🟩 (+1 more)
check	3m	🟩
general-lints	2m	🟩 🟩 🟩 🟩 🟩 (+1 more)
semgrep/ci	2m	🟩 🟩 🟩 🟩 🟩 (+1 more)
file_change_determinator	1m	🟩 🟩 🟩 🟩 🟩 (+1 more)
permission-check	18s	🟩 🟩 🟩 🟩 🟩 (+1 more)
permission-check	14s	🟩 🟩 🟩 🟩 🟩 (+1 more)
check-branch-prefix	1s	🟩

_{settings ⋅ feedback ⋅ docs ⋅ learn more about trunk.io}

[vineethk]

• [compiler-v2] Enable recent stack-optimizations by default #15595
• [compiler-v2][optimization] Create and retain temps for each arg #15514 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[github-actions]

✅ Forge suite realistic_env_max_load success on 311694401230605775aa0dc163e53621f2670cc9

two traffics test: inner traffic : committed: 14616.33 txn/s, latency: 2719.18 ms, (p50: 2700 ms, p70: 2700, p90: 3000 ms, p99: 3300 ms), latency samples: 5557440
two traffics test : committed: 100.10 txn/s, latency: 1371.94 ms, (p50: 1300 ms, p70: 1400, p90: 1500 ms, p99: 2500 ms), latency samples: 1740
Latency breakdown for phase 0: ["MempoolToBlockCreation: max: 1.605, avg: 1.561", "ConsensusProposalToOrdered: max: 0.324, avg: 0.296", "ConsensusOrderedToCommit: max: 0.318, avg: 0.309", "ConsensusProposalToCommit: max: 0.612, avg: 0.605"]
Max non-epoch-change gap was: 0 rounds at version 0 (avg 0.00) [limit 4], 0.62s no progress at version 35432 (avg 0.20s) [limit 15].
Max epoch-change gap was: 0 rounds at version 0 (avg 0.00) [limit 4], 0.56s no progress at version 2229840 (avg 0.56s) [limit 16].
Test Ok

• Grafana dashboard
• Humio Logs
• Axiom Logs
• Validator CPU Profile
• Fullnode CPU Profile
• Test runner output
• Test run is land-blocking

[github-actions]

✅ Forge suite framework_upgrade success on 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9

Compatibility test results for 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9 (PR)
Upgrade the nodes to version: 311694401230605775aa0dc163e53621f2670cc9
framework_upgrade::framework-upgrade::full-framework-upgrade : committed: 1681.59 txn/s, submitted: 1683.33 txn/s, failed submission: 1.73 txn/s, expired: 1.73 txn/s, latency: 1990.15 ms, (p50: 2000 ms, p70: 2100, p90: 2700 ms, p99: 3800 ms), latency samples: 135740
framework_upgrade::framework-upgrade::full-framework-upgrade : committed: 1423.02 txn/s, submitted: 1426.82 txn/s, failed submission: 3.80 txn/s, expired: 3.80 txn/s, latency: 2080.53 ms, (p50: 2100 ms, p70: 2200, p90: 3000 ms, p99: 4100 ms), latency samples: 127200
5. check swarm health
Compatibility test for 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9 passed
Upgrade the remaining nodes to version: 311694401230605775aa0dc163e53621f2670cc9
framework_upgrade::framework-upgrade::full-framework-upgrade : committed: 1428.33 txn/s, submitted: 1433.21 txn/s, failed submission: 4.66 txn/s, expired: 4.87 txn/s, latency: 2106.93 ms, (p50: 2100 ms, p70: 2300, p90: 3000 ms, p99: 4300 ms), latency samples: 128661
Test Ok

• Grafana dashboard
• Humio Logs
• Axiom Logs
• Validator CPU Profile
• Fullnode CPU Profile
• Test runner output
• Test run is land-blocking

[github-actions]

✅ Forge suite compat success on 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9

Compatibility test results for 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9 (PR)
1. Check liveness of validators at old version: 3c6e693a27339e73520f41030dce8fc9cd504967
compatibility::simple-validator-upgrade::liveness-check : committed: 16414.22 txn/s, latency: 2080.76 ms, (p50: 2100 ms, p70: 2200, p90: 2400 ms, p99: 2800 ms), latency samples: 530400
2. Upgrading first Validator to new version: 311694401230605775aa0dc163e53621f2670cc9
compatibility::simple-validator-upgrade::single-validator-upgrading : committed: 6391.73 txn/s, latency: 4408.13 ms, (p50: 4900 ms, p70: 5100, p90: 6200 ms, p99: 6300 ms), latency samples: 120780
compatibility::simple-validator-upgrade::single-validator-upgrade : committed: 6524.84 txn/s, latency: 4960.76 ms, (p50: 5300 ms, p70: 5400, p90: 5500 ms, p99: 5600 ms), latency samples: 222940
3. Upgrading rest of first batch to new version: 311694401230605775aa0dc163e53621f2670cc9
compatibility::simple-validator-upgrade::half-validator-upgrading : committed: 6634.80 txn/s, latency: 4303.67 ms, (p50: 4800 ms, p70: 5100, p90: 5600 ms, p99: 5700 ms), latency samples: 126600
compatibility::simple-validator-upgrade::half-validator-upgrade : committed: 6930.88 txn/s, latency: 4800.15 ms, (p50: 5100 ms, p70: 5200, p90: 5400 ms, p99: 5800 ms), latency samples: 233260
4. upgrading second batch to new version: 311694401230605775aa0dc163e53621f2670cc9
compatibility::simple-validator-upgrade::rest-validator-upgrading : committed: 11082.84 txn/s, latency: 2512.21 ms, (p50: 2800 ms, p70: 2900, p90: 3000 ms, p99: 3100 ms), latency samples: 194660
compatibility::simple-validator-upgrade::rest-validator-upgrade : committed: 11389.07 txn/s, latency: 2809.47 ms, (p50: 2900 ms, p70: 3000, p90: 3100 ms, p99: 3200 ms), latency samples: 372260
5. check swarm health
Compatibility test for 3c6e693a27339e73520f41030dce8fc9cd504967 ==> 311694401230605775aa0dc163e53621f2670cc9 passed
Test Ok

• Grafana dashboard
• Humio Logs
• Axiom Logs
• Validator CPU Profile
• Fullnode CPU Profile
• Test runner output
• Test run is land-blocking