cmd/compile: reduce allocations by reusing SSA scratch state

[siutsin]

Overall

Seven changes reuse temporary cmd/compile scratch state through the per-worker ssa.Cache or the generated SSA allocator pools.

The stack is split by compiler pass. CL 792805 covers prove scratch reuse, and CL 793640 covers SCCP scratch map reuse.

CL rows are commit-alone isolated allocs/op geomeans. CL 792800 through CL 792803 were measured on linux/amd64 at base fd6f414c65e6; CL 792804, CL 792805, CL 793640, and the full stack were measured on linux/amd64 at base 40f0f8125549.

CL	Area	Geomean allocs/op
CL 792800	deadcode inline-index set	-0.29%
CL 792801	slot canonicalizer	-0.19%
CL 792802	edgeState value slices	-1.39%
CL 792803	per-block defvars maps	-0.82%
CL 792804	regalloc scratch buffers	-0.73%
CL 792805	prove scratch	-0.50%
CL 793640	SCCP scratch maps	-0.34%
Full stack	combined	-4.28%

Full stack compilebench -alloc, linux/amd64, -count 20, -compileflags=-c=8, base 40f0f8125549:

Benchmark	allocs/op	vs master
Template	752.8k	-4.36%
Unicode	547.4k	-0.63%
GoTypes	4.522M	-6.37%
Compiler	686.7k	-2.54%
SSA	42.89M	-6.03%
Flate	754.5k	-4.37%
GoParser	723.7k	-4.59%
Reflect	2.139M	-4.62%
Tar	857.6k	-4.69%
XML	1.005M	-4.47%
geomean	1.494M	-4.28%

The same full-stack run reports B/op geomean -5.27%. A companion linux/amd64 wall-time and IPC study on the revised tail reports, with GOGC=off, instructions -0.52%, cycles -0.92%, and wall -0.83% for -c=1; the wall win strengthens to -2.27% at -c=16.

Validation:

• Generated code: toolstash -cmp over std reported no diffs.
• Full suite: cd src && ./all.bash passes.

CL 792800: cmd/compile: reuse deadcode liveInlIdx set across functions

Reuses liveValues' inline-index set in ssa.Cache. Keys are ints, so no *Value or *Block is retained across functions.

Metrics

Commit-alone compilebench -alloc, base fd6f414c65e6:

Benchmark	allocs/op	vs master
Template	784.9k	-0.29%
Unicode	550.8k	-0.01%
GoTypes	4.817M	-0.28%
Compiler	698.8k	-0.30%
SSA	45.53M	-0.21%
Flate	786.9k	-0.25%
GoParser	755.6k	-0.39%
Reflect	2.232M	-0.47%
Tar	896.6k	-0.35%
XML	1.049M	-0.33%
geomean	1.555M	-0.29%

CL 792801: cmd/compile: reuse slot canonicalizer across functions

Reuses PopulateABIInRegArgOps' slot canonicalizer through debugState on ssa.Cache. Reset clears LocalSlot values that can hold IR/type pointers.

Metrics

Commit-alone compilebench -alloc, base fd6f414c65e6:

Benchmark	allocs/op	vs master
Template	785.2k	-0.24%
Unicode	550.7k	-0.02%
GoTypes	4.819M	-0.22%
Compiler	700.1k	-0.11%
SSA	45.56M	-0.15%
Flate	787.3k	-0.20%
GoParser	757.1k	-0.19%
Reflect	2.237M	-0.28%
Tar	897.3k	-0.28%
XML	1.050M	-0.22%
geomean	1.557M	-0.19%

CL 792802: cmd/compile: reuse edgeState value-slice backing arrays across edges

Reuses edgeState per-vid []*Value backing arrays while shuffling registers. Arrays are returned from the map once, avoiding shared live backing arrays.

Metrics

Commit-alone compilebench -alloc, base fd6f414c65e6:

Benchmark	allocs/op	vs master
Template	777.6k	-1.21%
Unicode	549.1k	-0.31%
GoTypes	4.715M	-2.37%
Compiler	696.0k	-0.70%
SSA	44.66M	-2.13%
Flate	777.0k	-1.50%
GoParser	745.3k	-1.74%
Reflect	2.218M	-1.09%
Tar	889.1k	-1.18%
XML	1.035M	-1.64%
geomean	1.538M	-1.39%

CL 792803: cmd/compile: reuse per-block defvars maps across functions

Reuses ssagen startBlock defvars maps through a per-worker free list. Each map is cleared before reuse, so no ir.Node or *Value is retained.

Metrics

Commit-alone compilebench -alloc, base fd6f414c65e6:

Benchmark	allocs/op	vs master
Template	780.0k	-0.90%
Unicode	550.5k	-0.06%
GoTypes	4.774M	-1.15%
Compiler	697.7k	-0.45%
SSA	44.98M	-1.42%
Flate	781.8k	-0.90%
GoParser	752.9k	-0.73%
Reflect	2.223M	-0.91%
Tar	891.2k	-0.95%
XML	1.045M	-0.74%
geomean	1.547M	-0.82%

CL 792804: cmd/compile: reuse regalloc scratch buffers across functions

Reuses regalloc register-state, block-state, spill-live, and live/desired scratch buffers. Pointer-bearing register state and outer block/spill slice headers use the generated SSA allocator pools; pointer-free live/desired scratch stays on ssa.Cache.

Metrics

Commit-alone compilebench -alloc, base 40f0f8125549:

Benchmark	allocs/op	vs master
Template	781.6k	-0.70%
Unicode	550.7k	-0.04%
GoTypes	4.760M	-1.45%
Compiler	702.6k	-0.27%
SSA	45.00M	-1.43%
Flate	785.0k	-0.50%
GoParser	753.7k	-0.64%
Reflect	2.226M	-0.77%
Tar	892.6k	-0.80%
XML	1.045M	-0.71%
geomean	1.549M	-0.73%

CL 792805: cmd/compile: reuse prove scratch across functions

Reuses prove's orderings map and undo stacks. The pointer-bearing map is cleared before reuse, and the undo stacks use generated SSA allocator pools and pooled append helpers.

Metrics

Commit-alone compilebench -alloc, base 40f0f8125549:

Benchmark	allocs/op	vs master
Template	782.3k	-0.62%
Unicode	550.3k	-0.11%
GoTypes	4.806M	-0.50%
Compiler	701.8k	-0.40%
SSA	45.47M	-0.40%
Flate	784.4k	-0.58%
GoParser	754.4k	-0.54%
Reflect	2.228M	-0.66%
Tar	893.8k	-0.66%
XML	1.047M	-0.49%
geomean	1.553M	-0.50%

CL 793640: cmd/compile: reuse sccp scratch maps across functions

Reuses SCCP's visited, def-use, def-block, and lattice maps through ssa.Cache. These maps are pointer-bearing and the cache reuses *Value and *Block addresses across functions, so SCCP clears all four maps on every return path before reuse.

Metrics

Commit-alone compilebench -alloc, base 40f0f8125549:

Benchmark	allocs/op	vs master
Template	783.9k	-0.41%
Unicode	550.6k	-0.06%
GoTypes	4.813M	-0.37%
Compiler	702.8k	-0.25%
SSA	45.52M	-0.28%
Flate	785.6k	-0.42%
GoParser	755.7k	-0.38%
Reflect	2.233M	-0.44%
Tar	895.9k	-0.43%
XML	1.049M	-0.34%
geomean	1.555M	-0.34%

[gopherbot]

Change https://go.dev/cl/792803 mentions this issue: cmd/compile: reuse per-block defvars maps across functions

[gopherbot]

Change https://go.dev/cl/792801 mentions this issue: cmd/compile: reuse slot canonicalizer across functions

[gopherbot]

Change https://go.dev/cl/792800 mentions this issue: cmd/compile: reuse deadcode liveInlIdx set across functions

[gopherbot]

Change https://go.dev/cl/792802 mentions this issue: cmd/compile: reuse edgeState value-slice backing arrays across edges

[gopherbot]

Change https://go.dev/cl/792804 mentions this issue: cmd/compile: reuse regalloc scratch buffers across functions

[gopherbot]

Change https://go.dev/cl/792805 mentions this issue: cmd/compile: reuse prove and sccp scratch maps across functions

[gabyhelp]

Related Code Changes

• cmd/compile: reuse liveness structures
• cmd/compile: reduce some SSA garbage
• cmd/compile: reuse prove and sccp scratch maps across functions
• cmd/compile: reuse per-block defvars maps across functions
• cmd/compile: reuse regalloc scratch buffers across functions
• cmd/compile: make deadcode pass cheaper
• cmd/compile: stack-allocate values worklist in schedule
• cmd/compile: allocate less in regalloc's liveValues
• cmd/compile/internal/ssa: reduce location list memory use
• cmd/compile: optimize regalloc for phi value

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[dmitshur]

CC @golang/compiler.