chore(ssa): Additional unrolling and mem2reg passes

[vezenovm]

Description

Problem*

Resolves #7995 (comment)

From that PRs description:

I don't think the regressions are as bad as they look if we chose not to go with #8017 [this PR]. The sample program is very small and is only regressing due to no longer triggering a loop unroll. A larger loop already would not have had triggered an unroll, and perhaps with more aggressive loop unrolling in the future it will not be an issue.

We are getting some large regressions in #7995:

Looking at the SSA on master we have the following before and after unrolling:

Details

After Loop Invariant Code Motion:
brillig(inline) predicate_pure fn main f0 {
  b0():
    v3 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v3
    v4 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v4
    inc_rc v3
    inc_rc v4
    v5 = allocate -> &mut [Field; 2]
    store v3 at v5
    v6 = allocate -> &mut u32
    store u32 0 at v6
    v8 = allocate -> &mut [Field; 2]
    store v4 at v8
    v9 = allocate -> &mut u32
    store u32 0 at v9
    jmp b1(u32 0)
  b1(v0: u32):
    v11 = lt v0, u32 2
    jmpif v11 then: b2, else: b3
  b2():
    jmpif u1 1 then: b4, else: b5
  b3():
    v12 = load v5 -> [Field; 2]
    v13 = load v6 -> u32
    v14 = load v8 -> [Field; 2]
    v15 = load v9 -> u32
    v18 = array_set v14, index u32 1, value Field 27
    store v18 at v8
    store v15 at v9
    inc_rc v12
    v19 = array_get v12, index u32 0 -> Field
    v20 = array_get v12, index u32 1 -> Field
    v21 = array_get v12, index u32 0 -> Field
    v22 = array_set v12, index u32 1, value v21
    v23 = array_set v22, index u32 0, value v20
    store v23 at v5
    store v13 at v6
    jmp b6(u32 0)
  b4():
    v29 = cast v0 as Field
    v30 = load v5 -> [Field; 2]
    v31 = load v6 -> u32
    v32 = lt v31, u32 2
    constrain v32 == u1 1, "push out of bounds"
    v33 = array_set v30, index v31, value v29
    v34 = add v31, u32 1
    store v33 at v5
    store v34 at v6
    jmp b5()
  b5():
    v35 = unchecked_add v0, u32 1
    jmp b1(v35)
  b6(v1: u32):
    v24 = eq v1, u32 0
    jmpif v24 then: b7, else: b8
  b7():
    v25 = array_get v12, index v1 -> Field
    v26 = eq v25, v19
    constrain v25 == v19
    v27 = unchecked_add v1, u32 1
    jmp b6(v27)
  b8():
    return
}

After Unrolling:
brillig(inline) predicate_pure fn main f0 {
  b0():
    v1 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v1
    v2 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v2
    inc_rc v1
    inc_rc v2
    v3 = allocate -> &mut [Field; 2]
    store v1 at v3
    v4 = allocate -> &mut u32
    store u32 0 at v4
    v6 = allocate -> &mut [Field; 2]
    store v2 at v6
    v7 = allocate -> &mut u32
    store u32 0 at v7
    jmp b3()
  b1():
    v22 = load v3 -> [Field; 2]
    v23 = load v4 -> u32
    v24 = load v6 -> [Field; 2]
    v25 = load v7 -> u32
    v27 = array_set v24, index u32 1, value Field 27
    store v27 at v6
    store v25 at v7
    inc_rc v22
    v28 = array_get v22, index u32 0 -> Field
    v29 = array_get v22, index u32 1 -> Field
    v30 = array_get v22, index u32 0 -> Field
    v31 = array_set v22, index u32 1, value v30
    v32 = array_set v31, index u32 0, value v29
    store v32 at v3
    store v23 at v4
    v33 = array_get v22, index u32 0 -> Field
    v34 = eq v33, v28
    constrain v33 == v28
    jmp b2()
  b2():
    return
  b3():
    v8 = load v3 -> [Field; 2]
    v9 = load v4 -> u32
    v11 = lt v9, u32 2
    constrain v11 == u1 1, "push out of bounds"
    v13 = array_set v8, index v9, value Field 0
    v15 = add v9, u32 1
    store v13 at v3
    store v15 at v4
    jmp b4()
  b4():
    jmp b5()
  b5():
    v16 = load v3 -> [Field; 2]
    v17 = load v4 -> u32
    v18 = lt v17, u32 2
    constrain v18 == u1 1, "push out of bounds"
    v20 = array_set v16, index v17, value Field 1
    v21 = add v17, u32 1
    store v20 at v3
    store v21 at v4
    jmp b6()
  b6():
    jmp b1()
}

And the following after #7995

Details

After Loop Invariant Code Motion:
brillig(inline) predicate_pure fn main f0 {
  b0():
    v3 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v3
    v4 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v4
    inc_rc v3
    inc_rc v4
    v5 = allocate -> &mut [Field; 2]
    store v3 at v5
    v6 = allocate -> &mut u32
    store u32 0 at v6
    v8 = allocate -> &mut [Field; 2]
    store v4 at v8
    v9 = allocate -> &mut u32
    store u32 0 at v9
    jmp b1(u32 0)
  b1(v0: u32):
    v11 = lt v0, u32 2
    jmpif v11 then: b2, else: b3
  b2():
    jmpif u1 1 then: b4, else: b5
  b3():
    v12 = load v5 -> [Field; 2]
    v13 = load v6 -> u32
    v14 = load v8 -> [Field; 2]
    v15 = load v9 -> u32
    v18 = array_set v14, index u32 1, value Field 27
    store v18 at v8
    store v15 at v9
    inc_rc v12
    v19 = array_get v12, index u32 0 -> Field
    v20 = array_get v12, index u32 1 -> Field
    v21 = array_get v12, index u32 0 -> Field
    v22 = array_set v12, index u32 1, value v21
    v23 = array_set v22, index u32 0, value v20
    store v23 at v5
    store v13 at v6
    jmp b6(u32 0)
  b4():
    v29 = cast v0 as Field
    v30 = load v5 -> [Field; 2]
    v31 = load v6 -> u32
    v32 = lt v31, u32 2
    constrain v32 == u1 1, "push out of bounds"
    v33 = lt v31, u32 2
    constrain v33 == u1 1, "Index out of bounds"
    v34 = array_set v30, index v31, value v29
    v35 = add v31, u32 1
    store v34 at v5
    store v35 at v6
    jmp b5()
  b5():
    v36 = unchecked_add v0, u32 1
    jmp b1(v36)
  b6(v1: u32):
    v24 = eq v1, u32 0
    jmpif v24 then: b7, else: b8
  b7():
    v25 = array_get v12, index v1 -> Field
    v26 = eq v25, v19
    constrain v25 == v19
    v27 = unchecked_add v1, u32 1
    jmp b6(v27)
  b8():
    return
}

After Unrolling:
brillig(inline) predicate_pure fn main f0 {
  b0():
    v2 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v2
    v3 = make_array [Field 0, Field 0] : [Field; 2]
    inc_rc v3
    inc_rc v2
    inc_rc v3
    v4 = allocate -> &mut [Field; 2]
    store v2 at v4
    v5 = allocate -> &mut u32
    store u32 0 at v5
    v7 = allocate -> &mut [Field; 2]
    store v3 at v7
    v8 = allocate -> &mut u32
    store u32 0 at v8
    jmp b1(u32 0)
  b1(v0: u32):
    v10 = lt v0, u32 2
    jmpif v10 then: b2, else: b3
  b2():
    jmpif u1 1 then: b4, else: b5
  b3():
    v11 = load v4 -> [Field; 2]
    v12 = load v5 -> u32
    v13 = load v7 -> [Field; 2]
    v14 = load v8 -> u32
    v17 = array_set v13, index u32 1, value Field 27
    store v17 at v7
    store v14 at v8
    inc_rc v11
    v18 = array_get v11, index u32 0 -> Field
    v19 = array_get v11, index u32 1 -> Field
    v20 = array_get v11, index u32 0 -> Field
    v21 = array_set v11, index u32 1, value v20
    v22 = array_set v21, index u32 0, value v19
    store v22 at v4
    store v12 at v5
    v23 = array_get v11, index u32 0 -> Field
    v24 = eq v23, v18
    constrain v23 == v18
    jmp b6()
  b4():
    v26 = cast v0 as Field
    v27 = load v4 -> [Field; 2]
    v28 = load v5 -> u32
    v29 = lt v28, u32 2
    constrain v29 == u1 1, "push out of bounds"
    v30 = lt v28, u32 2
    constrain v30 == u1 1, "Index out of bounds"
    v31 = array_set v27, index v28, value v26
    v32 = add v28, u32 1
    store v31 at v4
    store v32 at v5
    jmp b5()
  b5():
    v33 = unchecked_add v0, u32 1
    jmp b1(v33)
  b6():
    return
}

We have constrain the BoundedVec len field when we push, but we now have an additional index OOB check on the array set. This is preventing us from unrolling this small Brillig loop.

Summary*

This regression can be solved with an additional unrolling pass following constant folding as we should be able to remove the repeat constrain instructions. We then need another mem2reg pass following simplify_cfg in case the blocks have been sufficiently simplified to remove more load/stores.

I was not sure if we wanted to include this optimization so I split it out from #7995. Two more passes for a very specific case seems a bit overblown. Loop unrolling should not be very expensive the second time as most loops that were not unrolled previously will be not be unrolled again. Most of the cost would be the additional mem2reg pass.

Additional Context

Documentation*

Check one:

• No documentation needed.
• Documentation included in this PR.
• [For Experimental Features] Documentation to be submitted in a separate PR.

PR Checklist*

• I have tested the changes locally.
• I have formatted the changes with Prettier and/or cargo fmt on default settings.

[github-actions]

Changes to Brillig bytecode sizes

Generated at commit: 7c633ab8786ce5178804adda008d24881c87032f, compared to commit: 511cb2fa0d9e910eac9375fc357dc7567ce7d92a

🧾 Summary (10% most significant diffs)

Program	Brillig opcodes (+/-)	%
missing_closure_env_inliner_zero	-16 ✅	-38.10%
brillig_rc_regression_6123_inliner_max	-121 ✅	-85.82%
brillig_rc_regression_6123_inliner_zero	-121 ✅	-85.82%

---

Full diff report 👇

Program	Brillig opcodes (+/-)	%
slice_regex_inliner_max	2,303 (-4)	-0.17%
hashmap_inliner_zero	6,842 (-12)	-0.18%
slice_regex_inliner_zero	1,657 (-16)	-0.96%
slice_regex_inliner_min	2,041 (-24)	-1.16%
slices_inliner_max	1,696 (-36)	-2.08%
references_inliner_zero	232 (-8)	-3.33%
array_to_slice_inliner_max	471 (-17)	-3.48%
fold_complex_outputs_inliner_max	385 (-17)	-4.23%
uhashmap_inliner_max	10,366 (-523)	-4.80%
hashmap_inliner_max	16,022 (-1,256)	-7.27%
derive_inliner_max	128 (-11)	-7.91%
side_effects_constrain_array_inliner_max	88 (-17)	-16.19%
tuples_inliner_max	32 (-9)	-21.95%
tuples_inliner_min	32 (-9)	-21.95%
tuples_inliner_zero	32 (-9)	-21.95%
references_inliner_max	99 (-29)	-22.66%
closures_mut_ref_inliner_zero	34 (-17)	-33.33%
missing_closure_env_inliner_zero	26 (-16)	-38.10%
brillig_rc_regression_6123_inliner_max	20 (-121)	-85.82%
brillig_rc_regression_6123_inliner_zero	20 (-121)	-85.82%

[github-actions]

Changes to number of Brillig opcodes executed

Generated at commit: 7c633ab8786ce5178804adda008d24881c87032f, compared to commit: 511cb2fa0d9e910eac9375fc357dc7567ce7d92a

🧾 Summary (10% most significant diffs)

Program	Brillig opcodes (+/-)	%
side_effects_constrain_array_inliner_max	-35 ✅	-46.05%
brillig_rc_regression_6123_inliner_max	-235 ✅	-92.89%
brillig_rc_regression_6123_inliner_zero	-235 ✅	-92.89%

---

Full diff report 👇

Program	Brillig opcodes (+/-)	%
hashmap_inliner_zero	67,665 (-12)	-0.02%
slice_regex_inliner_min	8,438 (-24)	-0.28%
slice_regex_inliner_zero	3,997 (-16)	-0.40%
slices_inliner_max	2,570 (-57)	-2.17%
uhashmap_inliner_max	123,665 (-2,814)	-2.22%
slice_regex_inliner_max	3,220 (-90)	-2.72%
array_to_slice_inliner_max	1,046 (-31)	-2.88%
hashmap_inliner_max	42,059 (-1,596)	-3.66%
references_inliner_zero	366 (-14)	-3.68%
derive_inliner_max	127 (-9)	-6.62%
fold_complex_outputs_inliner_max	599 (-45)	-6.99%
references_inliner_max	162 (-30)	-15.63%
tuples_inliner_max	26 (-5)	-16.13%
tuples_inliner_min	26 (-5)	-16.13%
tuples_inliner_zero	26 (-5)	-16.13%
closures_mut_ref_inliner_zero	34 (-13)	-27.66%
missing_closure_env_inliner_zero	22 (-12)	-35.29%
side_effects_constrain_array_inliner_max	41 (-35)	-46.05%
brillig_rc_regression_6123_inliner_max	18 (-235)	-92.89%
brillig_rc_regression_6123_inliner_zero	18 (-235)	-92.89%

[github-actions]

Changes to circuit sizes

Generated at commit: 7c633ab8786ce5178804adda008d24881c87032f, compared to commit: 511cb2fa0d9e910eac9375fc357dc7567ce7d92a

🧾 Summary (10% most significant diffs)

Program	ACIR opcodes (+/-)	%	Circuit size (+/-)	%
regression	-4 ✅	-1.36%	-5 ✅	-0.13%

---

Full diff report 👇

Program	ACIR opcodes (+/-)	%	Circuit size (+/-)	%
regression	290 (-4)	-1.36%	3,833 (-5)	-0.13%

[github-actions]

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Test Suite Duration'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.20.

Benchmark suite	Current: d9e5306	Previous: 511cb2f	Ratio
test_report_noir-lang_noir_bigcurve_	266 s	220 s	1.21

This comment was automatically generated by workflow using github-action-benchmark.

CC: @TomAFrench

[vezenovm]

regression_6674_1_inliner_zero 199 (+46) +30.07%
regression_6674_2_inliner_zero 199 (+46) +30.07%
regression_6674_1_inliner_max 221 (+24) +12.18%
regression_6674_2_inliner_max 221 (+24) +12.18%
regression_7062_inliner_max 95 (+8) +9.20%
regression_7062_inliner_min 95 (+8) +9.20%
regression_7062_inliner_zero 95 (+8) +9.20%

This report was run against master. It looks like we did not get rid of these regressions with the extra passes, but the largest regression from #7995 (comment) has been removed and we in fact got some improvements in various unreleated programs.

[jfecher]

Wonder if the 50s compilation time increase on bigcurve is spurious or related

[michaeljklein]

LGTM, though this might be something to enable with a flag if we see more compile duration regressions

[vezenovm]

Wonder if the 50s compilation time increase on bigcurve is spurious or related

I just updated the branch again so we shall see. If you look at the edit history for that alert it does seem spurious. In the first alert bigcurve did not trigger an alert, in the second alert is was 300s current vs. 219s previous for bigcurve, and the third alert was 260s vs 210s previous for bigcurve.

[vezenovm]

The next alert gave us 283s current vs 224s previous for bigcurve. So it does look like we have a compilation regression here. I'm going to run this PR with bench-show

[github-actions]

ACVM Benchmarks

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
purely_sequential_opcodes	264431 ns/iter (± 1578)	267669 ns/iter (± 7119)	0.99
perfectly_parallel_opcodes	232558 ns/iter (± 7246)	235345 ns/iter (± 9465)	0.99

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Compilation Time

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
regression_4709	0.641 s	0.653 s	0.98
ram_blowup_regression	14.7 s	14.7 s	1
global_var_regression_entry_points	0.46 s	0.451 s	1.02
private-kernel-inner	2.294 s	2.456 s	0.93
private-kernel-reset	7.04 s	7.038 s	1.00
private-kernel-tail	1.072 s	1.072 s	1
rollup-base-private	18.34 s	18.56 s	0.99
rollup-base-public	13.92 s	13.88 s	1.00
rollup-block-root-empty	0.896 s	0.904 s	0.99
rollup-block-root-single-tx	128 s	128 s	1
rollup-block-root	130 s	130 s	1
rollup-merge	0.881 s	0.849 s	1.04
rollup-root	1.528 s	1.482 s	1.03

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Execution Time

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
private-kernel-inner	0.028 s	0.029 s	0.97
private-kernel-reset	0.164 s	0.163 s	1.01
private-kernel-tail	0.017 s	0.017 s	1
rollup-base-private	0.334 s	0.337 s	0.99
rollup-base-public	0.215 s	0.214 s	1.00
rollup-block-root	16.3 s	16.5 s	0.99
rollup-merge	0.004 s	0.004 s	1
rollup-root	0.013 s	0.013 s	1

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Test Suite Duration

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
test_report_AztecProtocol_aztec-packages_noir-projects_aztec-nr	51 s	50 s	1.02
test_report_AztecProtocol_aztec-packages_noir-projects_noir-contracts	88 s	89 s	0.99
test_report_AztecProtocol_aztec-packages_noir-projects_noir-protocol-circuits_crates_blob	48 s	50 s	0.96
test_report_AztecProtocol_aztec-packages_noir-projects_noir-protocol-circuits_crates_private-kernel-lib	171 s	171 s	1
test_report_AztecProtocol_aztec-packages_noir-projects_noir-protocol-circuits_crates_rollup-lib	173 s	168 s	1.03
test_report_AztecProtocol_aztec-packages_noir-projects_noir-protocol-circuits_crates_types	51 s	53 s	0.96
test_report_noir-lang_noir-bignum_	76 s	79 s	0.96
test_report_noir-lang_noir_bigcurve_	252 s	220 s	1.15
test_report_noir-lang_sha512_	32 s	31 s	1.03

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Compilation Memory

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
private-kernel-inner	270.85 MB	270.14 MB	1.00
private-kernel-reset	688.39 MB	603.74 MB	1.14
private-kernel-tail	199.53 MB	199.52 MB	1.00
rollup-base-private	1560 MB	1560 MB	1
rollup-base-public	1320 MB	1320 MB	1
rollup-block-root-empty	270.22 MB	270.2 MB	1.00
rollup-block-root-single-tx	7770 MB	7770 MB	1
rollup-block-root	7780 MB	7780 MB	1
rollup-merge	271.42 MB	271.41 MB	1.00
rollup-root	321.92 MB	321.93 MB	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[github-actions]

Execution Memory

Details

Benchmark suite	Current: fa26c5d	Previous: 511cb2f	Ratio
private-kernel-inner	209.31 MB	209.31 MB	1
private-kernel-reset	246.05 MB	246.07 MB	1.00
private-kernel-tail	183.15 MB	183.15 MB	1
rollup-base-private	486.79 MB	486.79 MB	1
rollup-base-public	462.77 MB	462.77 MB	1
rollup-block-root	1710 MB	1710 MB	1
rollup-merge	254.16 MB	254.16 MB	1
rollup-root	260.72 MB	260.72 MB	1

This comment was automatically generated by workflow using github-action-benchmark.