Files
Nim/compiler/icprof.nim
araq 39b2f1830f IC: break down processTopLevel, and refute the obvious frontend win
The frontend's loading is 3.55s of Atlas's 9.2s and had no breakdown
finer than "TopLevel 1459ms". Now it does — `tTopReplay`, `tTopLogOps`,
`tTopOffers`, `tTopStmts` and a counter for the records the loader
skips:

    TopLevel 1459ms = Offers 569 + ExportBranch 312 + LogOps 137
                      + the bare cursor walk ~371 + Replay/Stmts ~11

That immediately suggests a target, and it is a trap. 80% of every
module header the loader walks is tooling-only records — `sig`, one per
signature-symbol occurrence, plus `expansion`/`modulesrc` — which it
skips on sight and which exist only for `idetools`: 3.36M of 4.19M nodes
on Atlas. Grouping them under one tag, or moving them past the
`(implementation)` marker where the loop stops, is an easy change and
buys nothing.

Probed before building any of it, by emitting none of them at all:

    nodes walked  4.19M -> 0.85M      .s.bif    44.7MB -> 44.0MB
    TopLevel      1459ms -> 1406ms    frontend  9.29s  -> 9.18s

Walking 3.3M records costs 53ms. `skip` on a `TagLit` is a jump, not a
scan — about 16ns a node — so the count was never the cost. A format
change for 0.5% would have been a bad trade discovered late.

What is left is the per-process re-load itself: 180 `nim m` each parsing
~20 modules' interfaces out of 44.7MB of `.s.bif`, with no dominant item
because there is no single item. It is the amortisation problem batching
already solved for the backend (`loadDepClosure` 10.2s -> 1.3s), and the
frontend is where it has not been solved.

Behind `-d:icBNodeProf`. ic 42/42, `koch boot -d:release` equal
executables.
2026-08-31 20:30:36 +02:00

115 lines
4.8 KiB
Nim

#
#
# The Nim Compiler
# (c) Copyright 2026 Andreas Rumpf
#
# See the file "copying.txt", included in this
# distribution, for details about the copyright.
#
## Opt-in instrumentation for the IC backend, enabled with `-d:icBNodeProf`.
## Off, every template below is `discard` and nothing is linked in.
##
## It lives in its own module with NO compiler imports so that any stage can
## use it without creating a cycle — `bnode` needs it for the accessors,
## `nifbackend` for the stage phases, `cgen` for what happens per routine.
##
## Each backend process appends ONE line to `$NIM_IC_BNODE_PROF` at exit (or to
## stderr when that is unset), because a `--ic:on` build fans out a process per
## module per stage and interleaved writes would tear. Use `-d:icNoParallel`
## when the numbers need to be attributable to a particular module.
##
## Counts are for volume, timings for cost, and the two answer different
## questions: the accessors turned out to be 700k calls worth 8ms, while `info`
## was 259k calls worth 1.36s. Neither number alone would have found that.
when defined(icBNodeProf):
import std / [envvars, exitprocs, syncio, monotimes]
from std / times import inNanoseconds
type
ProfSlot* = enum
pKind, pTagKindHit, pTagKindMiss, pAstChildren, pSkip, pSon, pLen,
pLastSon, pIterYield, pSym, pTyp, pTypTagLit, pOrigin, pNilType,
pGenBodyCalls, pInfo, pIfaceExported, pIfaceHidden, pIfaceModules,
pTopNodes, pExportSyms, pPeekKind, pPeekFallback, pPeekLoaded,
pTopToolingSkip
TimeSlot* = enum
tLoadClosure, tModuleId, tBifLoad, tPosIndex, tTopLevel, tInterfTables,
tTransform, tHandOff, tGenBody, tAnalyses,
tSym, tTyp, tInfo, tOrigin, tExportBranch, tResolveSym, tEnumFields,
# Coarse phases, added to find where a backend process spends the time
# that none of the slots above account for. `tStage` is the whole stage
# body, so `Process - tStage` is everything before it: exec, the Nim
# runtime, config replay, `registerNifSuffix`/graph setup.
tStage,
tLowerOwned, tLowerHooks, tLowerWrite,
tCgGen, tCgInit, tCgFinish, tCgWrite,
tMergeStage, tEmitRender, tLinkStage,
# `nim m` (the frontend): the sem pass as a whole, and writing the module's
# `.s.bif`. `Stage - WriteNif - <the loading slots>` is then sem proper.
tWriteNif,
# `processTopLevel`'s branches: which part of a module HEADER costs what.
tTopReplay, tTopLogOps, tTopOffers, tTopStmts
let procStart = getMonoTime()
## Set when this module initialises, i.e. essentially at process start, so
## the dump can report total process wall time and the startup share can be
## derived as `Process - Stage`.
var profStageName* = "frontend"
## Which invocation this is: the backend stage name, or "frontend" for a
## `nim m` process, which arms the profiler through ast2nif but never enters
## a backend stage. Without it the `Process - Stage` startup figure is
## meaningless — 204 frontend processes' whole runtime lands in it.
var profCounts: array[ProfSlot, int]
var profNanos: array[TimeSlot, int64]
var profStart: array[TimeSlot, MonoTime]
var profArmed = false
proc profDump() =
var line = "BNODEPROF stage=" & profStageName
for s in ProfSlot: line.add " " & ($s)[1..^1] & "=" & $profCounts[s]
for s in TimeSlot: line.add " " & ($s)[1..^1] & "ms=" & $(profNanos[s] div 1_000_000)
line.add " Processms=" & $((getMonoTime() - procStart).inNanoseconds div 1_000_000)
let f = getEnv("NIM_IC_BNODE_PROF")
if f.len > 0:
let h = open(f, fmAppend)
h.writeLine line
h.close()
else:
stderr.writeLine line
template armProf() =
if not profArmed:
profArmed = true
addExitProc profDump
template prof*(s: ProfSlot; n = 1) =
armProf()
inc profCounts[s], n
template icProfStart*(s: TimeSlot) =
armProf()
profStart[s] = getMonoTime()
template icProfStop*(s: TimeSlot) =
profNanos[s] += (getMonoTime() - profStart[s]).inNanoseconds
template timed*(s: TimeSlot; body: untyped) =
## Leaf timing. NOT re-entrant, and the phase slots are not disjoint —
## `tTransform` contains body materialization, `tTyp` reaches `tSym`. Read
## them as nested, not additive.
##
## Arms the dump like `prof`/`icProfStart` do. It did not, and so a process
## whose ONLY instrumentation is a `timed` never reported at all: the
## `merge`, `emit` and `link` stages were silently absent from every profile.
armProf()
let t0 = getMonoTime()
body
profNanos[s] += (getMonoTime() - t0).inNanoseconds
else:
template prof*(s: untyped; n = 1) = discard
template icProfStart*(s: untyped) = discard
template icProfStop*(s: untyped) = discard
template timed*(s: untyped; body: untyped) = body