]> Cypherpunks repositories - gostls13.git/commit
cmd/compile: convert constants to interfaces without allocating
authorJosh Bleecher Snyder <josharian@gmail.com>
Sat, 21 Jan 2017 21:41:06 +0000 (13:41 -0800)
committerJosh Bleecher Snyder <josharian@gmail.com>
Thu, 2 Feb 2017 21:02:23 +0000 (21:02 +0000)
commitc682d3239e5aa05a77ad21f2267efc4e2e60c05f
tree9c18ad9eeff69a4b999b53343da8723de0c21ef4
parentf395e878887df1a28966353f32f5ee70973c8f3f
cmd/compile: convert constants to interfaces without allocating

The order pass is responsible for ensuring that
values passed to runtime functions, including
convT2E/convT2I, are addressable.

Prior to this CL, this was always accomplished
by creating a temp, which frequently escaped to
the heap, causing allocations, perhaps most
notably in code like:

fmt.Println(1, 2, 3) // allocates three times

None of the runtime routines modify the contents
of the pointers they receive, so in the case of
constants, instead of creating a temp value,
we can create a static value.

(Marking the static value as read-only provides
protection against accidental attempts by the runtime
to modify the constant data.)

This improves code generation for code like:

panic("abc")
c <- 2 // c is a chan int

which can now simply refer to "abc" and 2,
rather than going by way of a temporary.

It also allows us to optimize convT2E/convT2I,
by recognizing static readonly values
and directly constructing the interface.

This CL adds ~0.5% to binary size, despite
decreasing the size of many functions,
because it also adds many static symbols.

This binary size regression could be recovered in
future (but currently unplanned) work.

There is a lot of content-duplication in these
symbols; this statement generates six new symbols,
three containing an int 1 and three containing
a pointer to the string "a":

fmt.Println(1, 1, 1, "a", "a", "a")

These symbols could be made content-addressable.

Furthermore, these symbols are small, so the
alignment and naming overhead is large.
As with the go.strings section, these symbols
could be hidden and have their alignment reduced.

The changes to test/live.go make it impossible
(at least with current optimization techniques)
to place the values being passed to the runtime
in static symbols, preserving autotmp creation.

Fixes #18704

Benchmarks from fmt and go-kit's logging package:

github.com/go-kit/kit/log

name                      old time/op    new time/op    delta
JSONLoggerSimple-8          1.91µs ± 2%    2.11µs ±22%     ~     (p=1.000 n=9+10)
JSONLoggerContextual-8      2.60µs ± 6%    2.43µs ± 2%   -6.29%  (p=0.000 n=9+10)
Discard-8                    101ns ± 2%      34ns ±14%  -66.33%  (p=0.000 n=10+9)
OneWith-8                    161ns ± 1%     102ns ±16%  -36.78%  (p=0.000 n=10+10)
TwoWith-8                    175ns ± 3%     106ns ± 7%  -39.36%  (p=0.000 n=10+9)
TenWith-8                    293ns ± 3%     227ns ±15%  -22.44%  (p=0.000 n=9+10)
LogfmtLoggerSimple-8         704ns ± 2%     608ns ± 2%  -13.65%  (p=0.000 n=10+9)
LogfmtLoggerContextual-8     962ns ± 1%     860ns ±17%  -10.57%  (p=0.003 n=9+10)
NopLoggerSimple-8            188ns ± 1%     120ns ± 1%  -36.39%  (p=0.000 n=9+10)
NopLoggerContextual-8        379ns ± 1%     243ns ± 0%  -35.77%  (p=0.000 n=9+10)
ValueBindingTimestamp-8      577ns ± 1%     499ns ± 1%  -13.51%  (p=0.000 n=10+10)
ValueBindingCaller-8         898ns ± 2%     844ns ± 2%   -6.00%  (p=0.000 n=10+10)

name                      old alloc/op   new alloc/op   delta
JSONLoggerSimple-8            904B ± 0%      872B ± 0%   -3.54%  (p=0.000 n=10+10)
JSONLoggerContextual-8      1.20kB ± 0%    1.14kB ± 0%   -5.33%  (p=0.000 n=10+10)
Discard-8                    64.0B ± 0%     32.0B ± 0%  -50.00%  (p=0.000 n=10+10)
OneWith-8                    96.0B ± 0%     64.0B ± 0%  -33.33%  (p=0.000 n=10+10)
TwoWith-8                     160B ± 0%      128B ± 0%  -20.00%  (p=0.000 n=10+10)
TenWith-8                     672B ± 0%      640B ± 0%   -4.76%  (p=0.000 n=10+10)
LogfmtLoggerSimple-8          128B ± 0%       96B ± 0%  -25.00%  (p=0.000 n=10+10)
LogfmtLoggerContextual-8      304B ± 0%      240B ± 0%  -21.05%  (p=0.000 n=10+10)
NopLoggerSimple-8             128B ± 0%       96B ± 0%  -25.00%  (p=0.000 n=10+10)
NopLoggerContextual-8         304B ± 0%      240B ± 0%  -21.05%  (p=0.000 n=10+10)
ValueBindingTimestamp-8       159B ± 0%      127B ± 0%  -20.13%  (p=0.000 n=10+10)
ValueBindingCaller-8          112B ± 0%       80B ± 0%  -28.57%  (p=0.000 n=10+10)

name                      old allocs/op  new allocs/op  delta
JSONLoggerSimple-8            19.0 ± 0%      17.0 ± 0%  -10.53%  (p=0.000 n=10+10)
JSONLoggerContextual-8        25.0 ± 0%      21.0 ± 0%  -16.00%  (p=0.000 n=10+10)
Discard-8                     3.00 ± 0%      1.00 ± 0%  -66.67%  (p=0.000 n=10+10)
OneWith-8                     3.00 ± 0%      1.00 ± 0%  -66.67%  (p=0.000 n=10+10)
TwoWith-8                     3.00 ± 0%      1.00 ± 0%  -66.67%  (p=0.000 n=10+10)
TenWith-8                     3.00 ± 0%      1.00 ± 0%  -66.67%  (p=0.000 n=10+10)
LogfmtLoggerSimple-8          4.00 ± 0%      2.00 ± 0%  -50.00%  (p=0.000 n=10+10)
LogfmtLoggerContextual-8      7.00 ± 0%      3.00 ± 0%  -57.14%  (p=0.000 n=10+10)
NopLoggerSimple-8             4.00 ± 0%      2.00 ± 0%  -50.00%  (p=0.000 n=10+10)
NopLoggerContextual-8         7.00 ± 0%      3.00 ± 0%  -57.14%  (p=0.000 n=10+10)
ValueBindingTimestamp-8       5.00 ± 0%      3.00 ± 0%  -40.00%  (p=0.000 n=10+10)
ValueBindingCaller-8          4.00 ± 0%      2.00 ± 0%  -50.00%  (p=0.000 n=10+10)

fmt

name                             old time/op    new time/op    delta
SprintfPadding-8                   88.9ns ± 3%    79.1ns ± 1%   -11.09%  (p=0.000 n=10+7)
SprintfEmpty-8                     12.6ns ± 3%    12.8ns ± 3%      ~     (p=0.136 n=10+10)
SprintfString-8                    38.7ns ± 5%    26.9ns ± 6%   -30.65%  (p=0.000 n=10+10)
SprintfTruncateString-8            56.7ns ± 2%    47.0ns ± 3%   -17.05%  (p=0.000 n=10+10)
SprintfQuoteString-8                164ns ± 2%     153ns ± 2%    -7.01%  (p=0.000 n=10+10)
SprintfInt-8                       38.9ns ±15%    26.5ns ± 2%   -31.93%  (p=0.000 n=10+9)
SprintfIntInt-8                    60.3ns ± 9%    38.2ns ± 1%   -36.67%  (p=0.000 n=10+8)
SprintfPrefixedInt-8               58.6ns ±13%    51.2ns ±11%   -12.66%  (p=0.001 n=10+10)
SprintfFloat-8                     71.4ns ± 3%    64.2ns ± 3%   -10.08%  (p=0.000 n=8+10)
SprintfComplex-8                    175ns ± 3%     159ns ± 2%    -9.03%  (p=0.000 n=10+10)
SprintfBoolean-8                   33.5ns ± 4%    25.7ns ± 5%   -23.28%  (p=0.000 n=10+10)
SprintfHexString-8                 65.3ns ± 3%    51.7ns ± 5%   -20.86%  (p=0.000 n=10+9)
SprintfHexBytes-8                  67.2ns ± 5%    67.9ns ± 4%      ~     (p=0.383 n=10+10)
SprintfBytes-8                      129ns ± 7%     124ns ± 7%      ~     (p=0.074 n=9+10)
SprintfStringer-8                   127ns ± 4%     126ns ± 8%      ~     (p=0.506 n=9+10)
SprintfStructure-8                  357ns ± 3%     359ns ± 3%      ~     (p=0.469 n=10+10)
ManyArgs-8                          203ns ± 6%     126ns ± 3%   -37.94%  (p=0.000 n=10+10)
FprintInt-8                         119ns ±10%      74ns ± 3%   -37.54%  (p=0.000 n=10+10)
FprintfBytes-8                      122ns ± 4%     120ns ± 3%      ~     (p=0.124 n=10+10)
FprintIntNoAlloc-8                 78.2ns ± 5%    74.1ns ± 3%    -5.28%  (p=0.000 n=10+10)
ScanInts-8                          349µs ± 1%     349µs ± 0%      ~     (p=0.606 n=9+8)
ScanRecursiveInt-8                 43.8ms ± 7%    40.1ms ± 2%    -8.42%  (p=0.000 n=10+10)
ScanRecursiveIntReaderWrapper-8    43.5ms ± 4%    40.4ms ± 2%    -7.16%  (p=0.000 n=10+9)

name                             old alloc/op   new alloc/op   delta
SprintfPadding-8                    24.0B ± 0%     16.0B ± 0%   -33.33%  (p=0.000 n=10+10)
SprintfEmpty-8                      0.00B          0.00B           ~     (all equal)
SprintfString-8                     21.0B ± 0%      5.0B ± 0%   -76.19%  (p=0.000 n=10+10)
SprintfTruncateString-8             32.0B ± 0%     16.0B ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfQuoteString-8                48.0B ± 0%     32.0B ± 0%   -33.33%  (p=0.000 n=10+10)
SprintfInt-8                        16.0B ± 0%      1.0B ± 0%   -93.75%  (p=0.000 n=10+10)
SprintfIntInt-8                     24.0B ± 0%      3.0B ± 0%   -87.50%  (p=0.000 n=10+10)
SprintfPrefixedInt-8                72.0B ± 0%     64.0B ± 0%   -11.11%  (p=0.000 n=10+10)
SprintfFloat-8                      16.0B ± 0%      8.0B ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfComplex-8                    48.0B ± 0%     32.0B ± 0%   -33.33%  (p=0.000 n=10+10)
SprintfBoolean-8                    8.00B ± 0%     4.00B ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfHexString-8                  96.0B ± 0%     80.0B ± 0%   -16.67%  (p=0.000 n=10+10)
SprintfHexBytes-8                    112B ± 0%      112B ± 0%      ~     (all equal)
SprintfBytes-8                      96.0B ± 0%     96.0B ± 0%      ~     (all equal)
SprintfStringer-8                   32.0B ± 0%     32.0B ± 0%      ~     (all equal)
SprintfStructure-8                   256B ± 0%      256B ± 0%      ~     (all equal)
ManyArgs-8                          80.0B ± 0%      0.0B       -100.00%  (p=0.000 n=10+10)
FprintInt-8                         8.00B ± 0%     0.00B       -100.00%  (p=0.000 n=10+10)
FprintfBytes-8                      32.0B ± 0%     32.0B ± 0%      ~     (all equal)
FprintIntNoAlloc-8                  0.00B          0.00B           ~     (all equal)
ScanInts-8                         15.2kB ± 0%    15.2kB ± 0%      ~     (p=0.248 n=9+10)
ScanRecursiveInt-8                 21.6kB ± 0%    21.6kB ± 0%      ~     (all equal)
ScanRecursiveIntReaderWrapper-8    21.7kB ± 0%    21.7kB ± 0%      ~     (all equal)

name                             old allocs/op  new allocs/op  delta
SprintfPadding-8                     2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfEmpty-8                       0.00           0.00           ~     (all equal)
SprintfString-8                      2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfTruncateString-8              2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfQuoteString-8                 2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfInt-8                         2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfIntInt-8                      3.00 ± 0%      1.00 ± 0%   -66.67%  (p=0.000 n=10+10)
SprintfPrefixedInt-8                 2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfFloat-8                       2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfComplex-8                     2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfBoolean-8                     2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfHexString-8                   2.00 ± 0%      1.00 ± 0%   -50.00%  (p=0.000 n=10+10)
SprintfHexBytes-8                    2.00 ± 0%      2.00 ± 0%      ~     (all equal)
SprintfBytes-8                       2.00 ± 0%      2.00 ± 0%      ~     (all equal)
SprintfStringer-8                    4.00 ± 0%      4.00 ± 0%      ~     (all equal)
SprintfStructure-8                   7.00 ± 0%      7.00 ± 0%      ~     (all equal)
ManyArgs-8                           8.00 ± 0%      0.00       -100.00%  (p=0.000 n=10+10)
FprintInt-8                          1.00 ± 0%      0.00       -100.00%  (p=0.000 n=10+10)
FprintfBytes-8                       1.00 ± 0%      1.00 ± 0%      ~     (all equal)
FprintIntNoAlloc-8                   0.00           0.00           ~     (all equal)
ScanInts-8                          1.60k ± 0%     1.60k ± 0%      ~     (all equal)
ScanRecursiveInt-8                  1.71k ± 0%     1.71k ± 0%      ~     (all equal)
ScanRecursiveIntReaderWrapper-8     1.71k ± 0%     1.71k ± 0%      ~     (all equal)

Change-Id: I7ba72a25fea4140a0ba40a9f443103ed87cc69b5
Reviewed-on: https://go-review.googlesource.com/35554
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
src/cmd/compile/internal/gc/order.go
src/cmd/compile/internal/gc/walk.go
src/cmd/internal/obj/data.go
test/live.go