]> Cypherpunks.ru repositories - gostls13.git/commit
cmd/compile,runtime: stack maps only at calls, remove register maps
authorAustin Clements <austin@google.com>
Thu, 16 Apr 2020 12:13:58 +0000 (08:13 -0400)
committerAustin Clements <austin@google.com>
Wed, 29 Apr 2020 21:29:21 +0000 (21:29 +0000)
commit9d812cfa5cbb1f573d61c452c864072270526753
treebf2731ef44747066f0482780f2cb336cbb6b4f71
parentee7d9f1c37c768b859c27dec2a897b12fa928c4e
cmd/compile,runtime: stack maps only at calls, remove register maps

Currently, we emit stack maps and register maps at almost every
instruction. This was originally intended to support non-cooperative
preemption, but was only ever used for debug call injection. Now debug
call injection also uses conservative frame scanning. As a result,
stack maps are only needed at call sites and register maps aren't
needed at all except that we happen to also encode unsafe-point
information in the register map PCDATA stream.

This CL reduces stack maps to only appear at calls, and replace full
register maps with just safe/unsafe-point information.

This is all protected by the go115ReduceLiveness feature flag, which
is defined in both runtime and cmd/compile.

This CL significantly reduces binary sizes and also speeds up compiles
and links:

name                      old exe-bytes     new exe-bytes     delta
BinGoSize                      15.0MB ± 0%       14.1MB ± 0%   -5.72%

name                      old pcln-bytes    new pcln-bytes    delta
BinGoSize                      3.14MB ± 0%       2.48MB ± 0%  -21.08%

name                      old time/op       new time/op       delta
Template                        178ms ± 7%        172ms ±14%  -3.59%  (p=0.005 n=19+19)
Unicode                        71.0ms ±12%       69.8ms ±10%    ~     (p=0.126 n=18+18)
GoTypes                         655ms ± 8%        615ms ± 8%  -6.11%  (p=0.000 n=19+19)
Compiler                        3.27s ± 6%        3.15s ± 7%  -3.69%  (p=0.001 n=20+20)
SSA                             7.10s ± 5%        6.85s ± 8%  -3.53%  (p=0.001 n=19+20)
Flate                           124ms ±15%        116ms ±22%  -6.57%  (p=0.024 n=18+19)
GoParser                        156ms ±26%        147ms ±34%    ~     (p=0.070 n=19+19)
Reflect                         406ms ± 9%        387ms ±21%  -4.69%  (p=0.028 n=19+20)
Tar                             163ms ±15%        162ms ±27%    ~     (p=0.370 n=19+19)
XML                             223ms ±13%        218ms ±14%    ~     (p=0.157 n=20+20)
LinkCompiler                    503ms ±21%        484ms ±23%    ~     (p=0.072 n=20+20)
ExternalLinkCompiler            1.27s ± 7%        1.22s ± 8%  -3.85%  (p=0.005 n=20+19)
LinkWithoutDebugCompiler        294ms ±17%        273ms ±11%  -7.16%  (p=0.001 n=19+18)

(https://perf.golang.org/search?q=upload:20200428.8)

The binary size improvement is even slightly better when you include
the CLs leading up to this. Relative to the parent of "cmd/compile:
mark PanicBounds/Extend as calls":

name                      old exe-bytes     new exe-bytes     delta
BinGoSize                      15.0MB ± 0%       14.1MB ± 0%   -6.18%

name                      old pcln-bytes    new pcln-bytes    delta
BinGoSize                      3.22MB ± 0%       2.48MB ± 0%  -22.92%

(https://perf.golang.org/search?q=upload:20200428.9)

For #36365.

Change-Id: I69448e714f2a44430067ca97f6b78e08c0abed27
Reviewed-on: https://go-review.googlesource.com/c/go/+/230544
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
src/cmd/compile/internal/gc/gsubr.go
src/cmd/compile/internal/gc/plive.go
src/cmd/internal/obj/link.go
src/cmd/internal/objabi/funcdata.go
src/runtime/debugcall.go
src/runtime/preempt.go
src/runtime/symtab.go