Loop breaking with a counter. Benchmarked (see comments),
eyeball checked for sanity on popular loops. This code
ought to handle loops in general, and properly inserts phi
functions in cases where the earlier version might not have.
Includes test, plus modifications to test/run.go to deal with
timeout and killing looping test. Tests broken by the addition
of extra code (branch frequency and live vars) for added
checks turn the check insertion off.
If GOEXPERIMENT=preemptibleloops, the compiler inserts reschedule
checks on every backedge of every reducible loop. Alternately,
specifying GO_GCFLAGS=-d=ssa/insert_resched_checks/on will
enable it for a single compilation, but because the core Go
libraries contain some loops that may run long, this is less
likely to have the desired effect.
This is intended as a tool to help in the study and diagnosis
of GC and other latency problems, now that goal STW GC latency
is on the order of 100 microseconds or less.
Updates #17831.
Updates #10958.
Change-Id: I6206c163a5b0248e3f21eb4fc65f73a179e1f639
Reviewed-on: https://go-review.googlesource.com/33910
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Static branch predictions (which guide block ordering) are
adjusted based on:
loop/not-loop (favor looping)
abnormal-exit/not (avoid panic)
call/not-call (avoid call)
ret/default (treat returns as rare)
This appears to make no difference in performance of real
code, meaning the compiler itself. The earlier version of
this has been stripped down to help make the cost of this
only-aesthetic-on-Intel phase be as cheap as possible (we
probably want information about inner loops for improving
register allocation, but because register allocation follows
close behind this pass, conceivably the information could be
reused -- so we might do this anyway just to normalize
output).
For a ./make.bash that takes 200 user seconds, about .75
second is reported in likelyadjust (summing nanoseconds
reported with -d=ssa/likelyadjust/time ).
Upstream predictions are respected.
Includes test, limited to build on amd64 only.
Did several iterations on the debugging output to allow
some rough checks on behavior.
Debug=1 logging notes agree/disagree with earlier passes,
allowing analysis like the following:
Run on make.bash:
GO_GCFLAGS=-d=ssa/likelyadjust/debug \
./make.bash >& lkly5.log
grep 'ranch prediction' lkly5.log | wc -l
78242 // 78k predictions
grep 'ranch predi' lkly5.log | egrep -v 'agrees with' | wc -l
29633 // 29k NEW predictions
grep 'disagrees' lkly5.log | wc -l
444 // contradicted 444 times
grep '< exit' lkly5.log | wc -l
10212 // 10k exit predictions
grep '< exit' lkly5.log | egrep 'disagrees' | wc -l
5 // 5 contradicted by previous prediction
grep '< exit' lkly5.log | egrep -v 'agrees' | wc -l
702 // 702-5 redundant with previous prediction
grep '< call' lkly5.log | egrep -v 'agrees' | wc -l
16699 // 16k new call predictions
grep 'stay in loop' lkly5.log | egrep -v 'agrees' | wc -l
3951 // 4k new "remain in loop" predictions
Fixes#11451.
Change-Id: Iafb0504f7030d304ef4b6dc1aba9a5789151a593
Reviewed-on: https://go-review.googlesource.com/19995
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
