]> Cypherpunks repositories - gostls13.git/log
gostls13.git
4 years agocmd/compile: fix ANDI/SRWI merge on ppc64
Paul E. Murphy [Thu, 15 Apr 2021 18:41:01 +0000 (13:41 -0500)]
cmd/compile: fix ANDI/SRWI merge on ppc64

The shift amount should be masked to avoid rotation values
beyond the numer of bits. In this case, if the shift amount
is 0, it should rotate 0, not 32.

Fixes #45589

Change-Id: I1e764497a39d0ec128e29af42352b70c70b2ecc5
Reviewed-on: https://go-review.googlesource.com/c/go/+/310569
Run-TryBot: Paul Murphy <murp@ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Carlos Eduardo Seo <carlos.seo@linaro.org>
Trust: Carlos Eduardo Seo <carlos.seo@linaro.org>

4 years agocmd/go/internal/modconv: involve GOPROXY in ConvertLegacyConfig
Aofei Sheng [Fri, 2 Apr 2021 06:45:26 +0000 (14:45 +0800)]
cmd/go/internal/modconv: involve GOPROXY in ConvertLegacyConfig

Fixes #33767

Change-Id: Ia2d6dd32bcb0e142f59346232b3b1aa3babb034a
Reviewed-on: https://go-review.googlesource.com/c/go/+/306809
Trust: Jay Conrod <jayconrod@google.com>
Trust: Bryan C. Mills <bcmills@google.com>
Run-TryBot: Jay Conrod <jayconrod@google.com>
Reviewed-by: Bryan C. Mills <bcmills@google.com>
Reviewed-by: Aofei Sheng <aofei@aofeisheng.com>
Reviewed-by: Jay Conrod <jayconrod@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/compile/internal/types2: add sizeof test
Robert Griesemer [Thu, 15 Apr 2021 19:18:08 +0000 (12:18 -0700)]
cmd/compile/internal/types2: add sizeof test

Change-Id: I75d3c8546b7ad51d8c7369e7289036b697d6b913
Reviewed-on: https://go-review.googlesource.com/c/go/+/310530
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Griesemer <gri@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
4 years agocmd/compile/internal/syntax: fix error message for ... without type
Robert Griesemer [Wed, 14 Apr 2021 20:10:34 +0000 (13:10 -0700)]
cmd/compile/internal/syntax: fix error message for ... without type

Only complain about missing type; leave it to type-checking
to decide whether "..." is permitted in the first place.

Fixes #43674.

Change-Id: Icbc8f084e364fe3ac16076406a134354219c08d0
Reviewed-on: https://go-review.googlesource.com/c/go/+/310209
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Griesemer <gri@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
4 years agocmd/internal/obj: consolidate AMOVW and AMOVWZ optab entries
Paul E. Murphy [Tue, 9 Mar 2021 22:55:11 +0000 (16:55 -0600)]
cmd/internal/obj: consolidate AMOVW and AMOVWZ optab entries

This requires consolidating the register move operations into a
single case entry in asmout.  These moves are also used to
sign/zero-extend register values.

Combine the three asmout cases for register moves.  This allows
AMOVWZ and AMOVW to be handled with the same optab entries.

Likewise, remove the diagnostic error for non-zero constant
loads into R0 using the register move operations, it is not
possible to match this asmout case with a non-zero constant.

Finally, fix the load constant 0 via "MOV[BHW]{,Z} $0, Rx".
These now generate "li Rx, $0" instead of a zero/sign-extend
of the contents of R0.

Change-Id: Ia4a263661582f10feda27ee21cb121e05ea931dc
Reviewed-on: https://go-review.googlesource.com/c/go/+/308190
Run-TryBot: Paul Murphy <murp@ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Carlos Eduardo Seo <carlos.seo@linaro.org>
Trust: Carlos Eduardo Seo <carlos.seo@linaro.org>

4 years agobytes: add asm implementation for index on ppc64x
Lynn Boger [Mon, 15 Mar 2021 21:36:47 +0000 (16:36 -0500)]
bytes: add asm implementation for index on ppc64x

This adds an asm implementation of index on ppc64le and
ppc64. It results in a significant improvement in some
of the benchmarks that use bytes.Index.

The implementation is based on a port of the s390x asm
implementation. Comments on the design are found
with the code.

The following improvements occurred on power8:

Index/10       70.7ns ± 0%    18.8ns ± 0%   -73.4
Index/32        165ns ± 0%      95ns ± 0%   -42.6
Index/4K       9.23µs ± 0%    4.91µs ± 0%   -46
Index/4M       9.52ms ± 0%    5.10ms ± 0%   -46.4
Index/64M       155ms ± 0%      85ms ± 0%   -45.1

Count/10       83.0ns ± 0%    32.1ns ± 0%   -61.3
Count/32        178ns ± 0%     109ns ± 0%   -38.8
Count/4K       9.24µs ± 0%    4.93µs ± 0%   -46
Count/4M       9.52ms ± 0%    5.10ms ± 0%   -46.4
Count/64M       155ms ± 0%      85ms ± 0%   -45.1

IndexHard1     2.36ms ± 0%    0.13ms ± 0%   -94.4
IndexHard2     2.36ms ± 0%    1.28ms ± 0%   -45.8
IndexHard3     2.36ms ± 0%    1.19ms ± 0%   -49.4
IndexHard4     2.36ms ± 0%    2.35ms ± 0%    -0.1

CountHard1     2.36ms ± 0%    0.13ms ± 0%   -94.4
CountHard2     2.36ms ± 0%    1.28ms ± 0%   -45.8
CountHard3     2.36ms ± 0%    1.19ms ± 0%   -49.4

IndexPeriodic/IndexPeriodic2  146µs ± 0%       8µs ± 0%   -94
IndexPeriodic/IndexPeriodic4  146µs ± 0%       8µs ± 0%   -94

Change-Id: I7dd2bb7e278726e27f51825ca8b2f8317d460e60
Reviewed-on: https://go-review.googlesource.com/c/go/+/309730
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Paul Murphy <murp@ibm.com>
Reviewed-by: Carlos Eduardo Seo <carlos.seo@linaro.org>
Trust: Carlos Eduardo Seo <carlos.seo@linaro.org>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>

4 years agonet/http: allow multiple dials in TestTransportMaxConnsPerHost
Michael Fraenkel [Thu, 15 Apr 2021 00:35:49 +0000 (18:35 -0600)]
net/http: allow multiple dials in TestTransportMaxConnsPerHost

If there is more than the expected single dial, the channel will block.
Allow at least one connection per client, and do the expected cleanup.

Updates #45570

Change-Id: Iaecd45298a7d7c591b7d7b1be13cea6e4a1e2e85
Reviewed-on: https://go-review.googlesource.com/c/go/+/310213
Reviewed-by: Bryan C. Mills <bcmills@google.com>
Reviewed-by: Damien Neil <dneil@google.com>
Trust: Damien Neil <dneil@google.com>
Trust: Emmanuel Odeke <emmanuel@orijtech.com>

4 years agomath: avoid assembly stubs
Austin Clements [Thu, 15 Apr 2021 01:57:24 +0000 (21:57 -0400)]
math: avoid assembly stubs

Currently almost all math functions have the following pattern:

func Sin(x float64) float64

func sin(x float64) float64 {
    // ... pure Go implementation ...
}

Architectures that implement a function in assembly provide the
assembly implementation directly as the exported function (e.g., Sin),
and architectures that don't implement it in assembly use a small stub
to jump back to the Go code, like:

TEXT ·Sin(SB), NOSPLIT, $0
JMP ·sin(SB)

However, most functions are not implemented in assembly on most
architectures, so this jump through assembly is a waste. It defeats
compiler optimizations like inlining. And, with regabi, it actually
adds a small but non-trivial overhead because the jump from assembly
back to Go must go through an ABI0->ABIInternal bridge function.

Hence, this CL reorganizes this structure across the entire package.
It now leans on inlining to achieve peak performance, but allows the
compiler to see all the way through the pure Go implementation.

Now, functions follow this pattern:

func Sin(x float64) float64 {
if haveArchSin {
return archSin(x)
}
return sin(x)
}

func sin(x float64) float64 {
    // ... pure Go implementation ...
}

Architectures that have assembly implementations use build-tagged
files to set haveArchX to true an provide an archX implementation.
That implementation can also still call back into the Go
implementation (some of them do this).

Prior to this change, enabling ABI wrappers results in a geomean
slowdown of the math benchmarks of 8.77% (full results:
https://perf.golang.org/search?q=upload:20210415.6) and of the Tile38
benchmarks by ~4%. After this change, enabling ABI wrappers is
completely performance-neutral on Tile38 and all but one math
benchmark (full results:
https://perf.golang.org/search?q=upload:20210415.7). ABI wrappers slow
down SqrtIndirectLatency-12 by 2.09%, which makes sense because that
call must still go through an ABI wrapper.

With ABI wrappers disabled (which won't be an option on amd64 much
longer), on linux/amd64, this change is largely performance-neutral
and slightly improves the performance of a few benchmarks:

(Because there are so many benchmarks, I've applied the Šidák
correction to the alpha threshold. It makes relatively little
difference in which benchmarks are statistically significant.)

name                    old time/op  new time/op  delta
Acos-12                 22.3ns ± 0%  18.8ns ± 1%  -15.44%  (p=0.000 n=18+16)
Acosh-12                28.2ns ± 0%  28.2ns ± 0%     ~     (p=0.404 n=18+20)
Asin-12                 18.1ns ± 0%  18.2ns ± 0%   +0.20%  (p=0.000 n=18+16)
Asinh-12                32.8ns ± 0%  32.9ns ± 1%     ~     (p=0.891 n=18+20)
Atan-12                 9.92ns ± 0%  9.90ns ± 1%   -0.24%  (p=0.000 n=17+16)
Atanh-12                27.7ns ± 0%  27.5ns ± 0%   -0.72%  (p=0.000 n=16+20)
Atan2-12                18.5ns ± 0%  18.4ns ± 0%   -0.59%  (p=0.000 n=19+19)
Cbrt-12                 22.1ns ± 0%  22.1ns ± 0%     ~     (p=0.804 n=16+17)
Ceil-12                 0.84ns ± 0%  0.84ns ± 0%     ~     (p=0.663 n=18+16)
Copysign-12             0.84ns ± 0%  0.84ns ± 0%     ~     (p=0.762 n=16+19)
Cos-12                  12.7ns ± 0%  12.7ns ± 1%     ~     (p=0.145 n=19+18)
Cosh-12                 22.2ns ± 0%  22.5ns ± 0%   +1.60%  (p=0.000 n=17+19)
Erf-12                  11.1ns ± 1%  11.1ns ± 1%     ~     (p=0.010 n=19+19)
Erfc-12                 12.6ns ± 1%  12.7ns ± 0%     ~     (p=0.066 n=19+15)
Erfinv-12               16.1ns ± 0%  16.1ns ± 0%     ~     (p=0.462 n=17+20)
Erfcinv-12              16.0ns ± 1%  16.0ns ± 1%     ~     (p=0.015 n=17+16)
Exp-12                  16.3ns ± 0%  16.5ns ± 1%   +1.25%  (p=0.000 n=19+16)
ExpGo-12                36.2ns ± 1%  36.1ns ± 1%     ~     (p=0.242 n=20+18)
Expm1-12                18.6ns ± 0%  18.7ns ± 0%   +0.25%  (p=0.000 n=16+19)
Exp2-12                 34.7ns ± 0%  34.6ns ± 1%     ~     (p=0.010 n=19+18)
Exp2Go-12               34.8ns ± 1%  34.8ns ± 1%     ~     (p=0.372 n=19+19)
Abs-12                  0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.766 n=18+16)
Dim-12                  0.84ns ± 1%  0.84ns ± 1%     ~     (p=0.167 n=17+19)
Floor-12                0.84ns ± 0%  0.84ns ± 0%     ~     (p=0.993 n=18+16)
Max-12                  3.35ns ± 0%  3.35ns ± 0%     ~     (p=0.894 n=17+19)
Min-12                  3.35ns ± 0%  3.36ns ± 1%     ~     (p=0.214 n=18+18)
Mod-12                  35.2ns ± 0%  34.7ns ± 0%   -1.45%  (p=0.000 n=18+17)
Frexp-12                5.31ns ± 0%  4.75ns ± 0%  -10.51%  (p=0.000 n=19+18)
Gamma-12                14.8ns ± 0%  16.2ns ± 1%   +9.21%  (p=0.000 n=20+19)
Hypot-12                6.16ns ± 0%  6.17ns ± 0%   +0.26%  (p=0.000 n=19+20)
HypotGo-12              7.79ns ± 1%  7.78ns ± 0%     ~     (p=0.497 n=18+17)
Ilogb-12                4.47ns ± 0%  4.47ns ± 0%     ~     (p=0.167 n=19+19)
J0-12                   76.0ns ± 0%  76.3ns ± 0%   +0.35%  (p=0.000 n=19+18)
J1-12                   76.8ns ± 1%  75.9ns ± 0%   -1.14%  (p=0.000 n=18+18)
Jn-12                    167ns ± 1%   168ns ± 1%     ~     (p=0.038 n=18+18)
Ldexp-12                6.98ns ± 0%  6.43ns ± 0%   -7.97%  (p=0.000 n=17+18)
Lgamma-12               15.9ns ± 0%  16.0ns ± 1%     ~     (p=0.011 n=20+17)
Log-12                  13.3ns ± 0%  13.4ns ± 1%   +0.37%  (p=0.000 n=15+18)
Logb-12                 4.75ns ± 0%  4.75ns ± 0%     ~     (p=0.831 n=16+18)
Log1p-12                19.5ns ± 0%  19.5ns ± 1%     ~     (p=0.851 n=18+17)
Log10-12                15.9ns ± 0%  14.0ns ± 0%  -11.92%  (p=0.000 n=17+16)
Log2-12                 7.88ns ± 1%  8.01ns ± 0%   +1.72%  (p=0.000 n=20+20)
Modf-12                 4.75ns ± 0%  4.34ns ± 0%   -8.66%  (p=0.000 n=19+17)
Nextafter32-12          5.31ns ± 0%  5.31ns ± 0%     ~     (p=0.389 n=17+18)
Nextafter64-12          5.03ns ± 1%  5.03ns ± 0%     ~     (p=0.774 n=17+18)
PowInt-12               29.9ns ± 0%  28.5ns ± 0%   -4.69%  (p=0.000 n=18+19)
PowFrac-12              91.0ns ± 0%  91.1ns ± 0%     ~     (p=0.029 n=19+19)
Pow10Pos-12             1.12ns ± 0%  1.12ns ± 0%     ~     (p=0.363 n=20+20)
Pow10Neg-12             3.90ns ± 0%  3.90ns ± 0%     ~     (p=0.921 n=17+18)
Round-12                2.31ns ± 0%  2.31ns ± 1%     ~     (p=0.390 n=18+18)
RoundToEven-12          0.84ns ± 0%  0.84ns ± 0%     ~     (p=0.280 n=18+19)
Remainder-12            31.6ns ± 0%  29.6ns ± 0%   -6.16%  (p=0.000 n=18+17)
Signbit-12              0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.385 n=19+18)
Sin-12                  12.5ns ± 0%  12.5ns ± 0%     ~     (p=0.080 n=18+18)
Sincos-12               16.4ns ± 2%  16.4ns ± 2%     ~     (p=0.253 n=20+19)
Sinh-12                 26.1ns ± 0%  26.1ns ± 0%   +0.18%  (p=0.000 n=17+19)
SqrtIndirect-12         3.91ns ± 0%  3.90ns ± 0%     ~     (p=0.133 n=19+19)
SqrtLatency-12          2.79ns ± 0%  2.79ns ± 0%     ~     (p=0.226 n=16+19)
SqrtIndirectLatency-12  6.68ns ± 0%  6.37ns ± 2%   -4.66%  (p=0.000 n=17+20)
SqrtGoLatency-12        49.4ns ± 0%  49.4ns ± 0%     ~     (p=0.289 n=18+16)
SqrtPrime-12            3.18µs ± 0%  3.18µs ± 0%     ~     (p=0.084 n=17+18)
Tan-12                  13.8ns ± 0%  13.9ns ± 2%     ~     (p=0.292 n=19+20)
Tanh-12                 25.4ns ± 0%  25.4ns ± 0%     ~     (p=0.101 n=17+17)
Trunc-12                0.84ns ± 0%  0.84ns ± 0%     ~     (p=0.765 n=18+16)
Y0-12                   75.8ns ± 0%  75.9ns ± 1%     ~     (p=0.805 n=16+18)
Y1-12                   76.3ns ± 0%  75.3ns ± 1%   -1.34%  (p=0.000 n=19+17)
Yn-12                    164ns ± 0%   164ns ± 2%     ~     (p=0.356 n=18+20)
Float64bits-12          0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.383 n=18+18)
Float64frombits-12      0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.066 n=18+19)
Float32bits-12          0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.889 n=16+19)
Float32frombits-12      0.56ns ± 0%  0.56ns ± 0%     ~     (p=0.007 n=18+19)
FMA-12                  23.9ns ± 0%  24.0ns ± 0%   +0.31%  (p=0.000 n=16+17)
[Geo mean]              9.86ns       9.77ns        -0.87%

(https://perf.golang.org/search?q=upload:20210415.5)

For #40724.

Change-Id: I44fbba2a17be930ec9daeb0a8222f55cd50555a0
Reviewed-on: https://go-review.googlesource.com/c/go/+/310331
Trust: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agocmd/link: issue error if elf header overruns
Paul E. Murphy [Mon, 5 Apr 2021 18:36:13 +0000 (13:36 -0500)]
cmd/link: issue error if elf header overruns

This is probably unlikely in practice, but when debugging alignment
related issues on ppc64 using very small text section splits, the elf
header could grow beyond the preallocated space and quietly stomp
on the first few text sections.

Change-Id: Ided58aa0b1e60f9da4b3cb277e4ebafcee4ec693
Reviewed-on: https://go-review.googlesource.com/c/go/+/307430
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>

4 years agoruntime: unify C->Go ABI transitions
Austin Clements [Wed, 14 Apr 2021 03:30:19 +0000 (23:30 -0400)]
runtime: unify C->Go ABI transitions

The previous CL introduced macros for transitions from the Windows ABI
to the Go ABI. This CL does the same for SysV and uses them in almost
all places where we transition from the C ABI to the Go ABI.

Compared to Windows, this transition is much simpler and I didn't find
any places that were getting it wrong. But this does let us unify a
lot of code nicely and introduces some degree of abstraction around
these ABI transitions.

Change-Id: Ib6bdecafce587ce18fca4c8300fcf401284a2bcd
Reviewed-on: https://go-review.googlesource.com/c/go/+/309930
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime,runtime/cgo: save all necessary registers on entry to Go on Windows
Austin Clements [Tue, 13 Apr 2021 13:06:50 +0000 (09:06 -0400)]
runtime,runtime/cgo: save all necessary registers on entry to Go on Windows

There are several assembly functions that transition from the Windows
ABI to the Go ABI. These all need to save all registers that are
callee-save in the Windows ABI and caller-save in the Go ABI and
prepare the register state for Go. However, they all do this slightly
differently and most of them don't save the necessary XMM registers
for this transition (which could corrupt them in the C caller).
Furthermore, now that we have a carefully specified Go ABI, it's clear
that none of these actually get all of the details 100% right.

So, unify this code into two macros in a shared header in
runtime/cgo/abi_amd64.h that handle all necessary registers and setup
and use these macros everywhere on Windows that handles transitions
from C to Go.

Change-Id: I62f41345a507aad1ca383814ac8b7e2a9ffb821e
Reviewed-on: https://go-review.googlesource.com/c/go/+/309769
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime: minor refactoring of _rt0_amd64_lib
Austin Clements [Wed, 14 Apr 2021 03:46:18 +0000 (23:46 -0400)]
runtime: minor refactoring of _rt0_amd64_lib

This function bounces between the C and Go ABIs a few times. This CL
narrows the scope of the Go -> C transition to just around the branch
that calls C. This lets us take advantage of C callee-save registers
to simplify the code a little.

Change-Id: I1ffa0b9e50325425c5ec66596978aeb6450a6b57
Reviewed-on: https://go-review.googlesource.com/c/go/+/309929
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agomime: keep builtinTypesLower sorted alphabetically
Tobias Klauser [Thu, 15 Apr 2021 08:48:12 +0000 (10:48 +0200)]
mime: keep builtinTypesLower sorted alphabetically

Updates #44602

Change-Id: I2c32e388143e56928850821587f57d9729434220
Reviewed-on: https://go-review.googlesource.com/c/go/+/310034
Trust: Tobias Klauser <tobias.klauser@gmail.com>
Run-TryBot: Tobias Klauser <tobias.klauser@gmail.com>
Reviewed-by: Emmanuel Odeke <emmanuel@orijtech.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agostrconv: Implement Ryū algorithm for ftoa shortest mode
Rémy Oudompheng [Wed, 27 Mar 2019 05:55:52 +0000 (06:55 +0100)]
strconv: Implement Ryū algorithm for ftoa shortest mode

This patch implements the algorithm from Ulf Adams,
"Ryū: Fast Float-to-String Conversion" (doi:10.1145/3192366.3192369)
for formatting floating-point numbers with a fixed number of decimal
digits.

It is not a direct translation of the reference C implementation
but still follows the original paper. In particular, it uses full
128-bit powers of 10, which allows for more precision in the other
modes (fixed ftoa, atof).

name                              old time/op  new time/op   delta
AppendFloat/Decimal-4             49.6ns ± 3%   59.3ns ± 0%  +19.59%  (p=0.008 n=5+5)
AppendFloat/Float-4                122ns ± 1%     91ns ± 1%  -25.92%  (p=0.008 n=5+5)
AppendFloat/Exp-4                 89.3ns ± 1%  100.0ns ± 1%  +11.98%  (p=0.008 n=5+5)
AppendFloat/NegExp-4              88.3ns ± 2%   97.1ns ± 1%   +9.87%  (p=0.008 n=5+5)
AppendFloat/LongExp-4              143ns ± 2%    103ns ± 0%  -28.17%  (p=0.016 n=5+4)
AppendFloat/Big-4                  144ns ± 1%    110ns ± 1%  -23.26%  (p=0.008 n=5+5)
AppendFloat/BinaryExp-4           46.2ns ± 2%   46.0ns ± 1%     ~     (p=0.603 n=5+5)
AppendFloat/32Integer-4           49.1ns ± 1%   58.7ns ± 1%  +19.57%  (p=0.008 n=5+5)
AppendFloat/32ExactFraction-4     95.6ns ± 1%   88.6ns ± 1%   -7.30%  (p=0.008 n=5+5)
AppendFloat/32Point-4              122ns ± 1%     87ns ± 1%  -28.63%  (p=0.008 n=5+5)
AppendFloat/32Exp-4               88.6ns ± 2%   95.0ns ± 1%   +7.29%  (p=0.008 n=5+5)
AppendFloat/32NegExp-4            87.2ns ± 1%   91.3ns ± 1%   +4.63%  (p=0.008 n=5+5)
AppendFloat/32Shortest-4           107ns ± 1%     82ns ± 0%  -24.08%  (p=0.008 n=5+5)
AppendFloat/Slowpath64-4          1.00µs ± 1%   0.10µs ± 0%  -89.92%  (p=0.016 n=5+4)
AppendFloat/SlowpathDenormal64-4  34.1µs ± 3%    0.1µs ± 1%  -99.72%  (p=0.008 n=5+5)

Fixes #15672

Change-Id: Ib90dfa245f62490a6666671896013cf3f9a1fb22
Reviewed-on: https://go-review.googlesource.com/c/go/+/170080
Trust: Emmanuel Odeke <emmanuel@orijtech.com>
Trust: Nigel Tao <nigeltao@golang.org>
Run-TryBot: Emmanuel Odeke <emmanuel@orijtech.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Nigel Tao <nigeltao@golang.org>
4 years agostrconv: implement Ryū-like algorithm for fixed precision ftoa
Rémy Oudompheng [Sun, 24 Mar 2019 22:21:38 +0000 (23:21 +0100)]
strconv: implement Ryū-like algorithm for fixed precision ftoa

This patch implements a simplified version of Ulf Adams,
"Ryū: Fast Float-to-String Conversion" (doi:10.1145/3192366.3192369)
for formatting floating-point numbers with a fixed number of decimal
digits.

It uses the same principles but does not need to handle
the complex task of finding a shortest representation.
This allows to handle a few more cases than Grisu3, notably
formatting with up to 18 significant digits.

name                         old time/op  new time/op  delta
AppendFloat/32Fixed8Hard-4   72.0ns ± 2%  56.0ns ± 2%  -22.28%  (p=0.000 n=10+10)
AppendFloat/32Fixed9Hard-4   74.8ns ± 0%  64.2ns ± 2%  -14.16%  (p=0.000 n=8+10)
AppendFloat/64Fixed1-4       60.4ns ± 1%  54.2ns ± 1%  -10.31%  (p=0.000 n=10+9)
AppendFloat/64Fixed2-4       66.3ns ± 1%  53.3ns ± 1%  -19.54%  (p=0.000 n=10+9)
AppendFloat/64Fixed3-4       61.0ns ± 1%  55.0ns ± 2%   -9.80%  (p=0.000 n=9+10)
AppendFloat/64Fixed4-4       66.9ns ± 0%  52.0ns ± 2%  -22.20%  (p=0.000 n=8+10)
AppendFloat/64Fixed12-4      95.5ns ± 1%  76.2ns ± 3%  -20.19%  (p=0.000 n=10+9)
AppendFloat/64Fixed16-4      1.62µs ± 0%  0.07µs ± 2%  -95.69%  (p=0.000 n=10+10)
AppendFloat/64Fixed12Hard-4  1.27µs ± 1%  0.07µs ± 1%  -94.83%  (p=0.000 n=9+9)
AppendFloat/64Fixed17Hard-4  3.68µs ± 1%  0.08µs ± 2%  -97.86%  (p=0.000 n=10+9)
AppendFloat/64Fixed18Hard-4  3.67µs ± 0%  3.72µs ± 1%   +1.44%  (p=0.000 n=9+10)

Updates #15672

Change-Id: I160963e141dd48287ad8cf57bcc3c686277788e8
Reviewed-on: https://go-review.googlesource.com/c/go/+/170079
Reviewed-by: Emmanuel Odeke <emmanuel@orijtech.com>
Trust: Emmanuel Odeke <emmanuel@orijtech.com>
Trust: Nigel Tao <nigeltao@golang.org>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Emmanuel Odeke <emmanuel@orijtech.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agointernal/bytealg: port more performance-critical functions to ABIInternal
Austin Clements [Wed, 14 Apr 2021 23:15:42 +0000 (19:15 -0400)]
internal/bytealg: port more performance-critical functions to ABIInternal

CL 308931 ported several runtime assembly functions to ABIInternal so
that compiler-generated ABIInternal calls don't go through ABI
wrappers, but it missed the runtime assembly functions that are
actually defined in internal/bytealg.

This eliminates the cost of wrappers for the BleveQuery and
GopherLuaKNucleotide benchmarks, but there's still more to do for
Tile38.

                                      0-base                1-wrappers
                                     sec/op        sec/op            vs base
BleveQuery                          6.507 ± 0%    6.477 ± 0%  -0.46% (p=0.004 n=20)
GopherLuaKNucleotide                30.39 ± 1%    30.34 ± 0%       ~ (p=0.301 n=20)
Tile38IntersectsCircle100kmRequest 1.038m ± 1%   1.080m ± 2%  +4.03% (p=0.000 n=20)

For #40724.

Change-Id: I0b722443f684fcb997b1d70802c5ed4b8d8f9829
Reviewed-on: https://go-review.googlesource.com/c/go/+/310184
Trust: Austin Clements <austin@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agocmd/internal/obj/arm64: fix the wrong sp dst register of ADDS/SUBS instructions
eric fang [Tue, 13 Apr 2021 02:48:32 +0000 (02:48 +0000)]
cmd/internal/obj/arm64: fix the wrong sp dst register of ADDS/SUBS instructions

According the armv8-a specification, the destination register of the ADDS/ADDSW/
SUBS/SUBSW instructions can not be RSP, the current implementation does not
check this and encodes this wrong instruction format as a CMN instruction. This
CL adds a check and test cases for this situation.

Change-Id: I92cc2f8e17dbda70f0dce8fddf1ca6d5d7730589
Reviewed-on: https://go-review.googlesource.com/c/go/+/309989
Reviewed-by: eric fang <eric.fang@arm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Trust: eric fang <eric.fang@arm.com>
Run-TryBot: eric fang <eric.fang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agotime: add missing "os" import to zoneinfo_test.go
Damien Neil [Thu, 15 Apr 2021 00:31:09 +0000 (17:31 -0700)]
time: add missing "os" import to zoneinfo_test.go

Updates #45448

Change-Id: I2e79ae6b9cf43a481aa703578712619ea344e421
Reviewed-on: https://go-review.googlesource.com/c/go/+/310212
Trust: Damien Neil <dneil@google.com>
Run-TryBot: Damien Neil <dneil@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agocmd/compile: propagate pragmas from generic function to stenciled implementation
Keith Randall [Wed, 14 Apr 2021 23:18:58 +0000 (16:18 -0700)]
cmd/compile: propagate pragmas from generic function to stenciled implementation

Change-Id: I28a1910890659aaa449ffd2a847cd4ced5a8600d
Reviewed-on: https://go-review.googlesource.com/c/go/+/310211
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Dan Scales <danscales@google.com>
4 years agocmd/compile: look for function in instantiations in all global assignments
Dan Scales [Wed, 14 Apr 2021 21:54:14 +0000 (14:54 -0700)]
cmd/compile: look for function in instantiations in all global assignments

Add in some missing global assignment ops to the list of globals ops
that should be traversed to look for generic function instantiations.
The most common other one for global assigments (and the relevant one
for this bug) is OAS2FUNC, but also look at global assigments with
OAS2DOTTYPE, OAS2MAPR, OAS2RECV, and OASOP.

Bonus small fix: get rid of -G=3 case in ir.IsAddressable. Now that we
don't call the old typechecker from noder2, we don't need this -G-3
check anymore.

Fixes #45547.

Change-Id: I75fecec55ea0d6f62e1c2294d4d77447ed9be6ae
Reviewed-on: https://go-review.googlesource.com/c/go/+/310210
Trust: Dan Scales <danscales@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Dan Scales <danscales@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
4 years agomime: add mime type for avif image file format
Amit Kumar [Tue, 22 Sep 2020 13:22:06 +0000 (18:52 +0530)]
mime: add mime type for avif image file format

Fixes #44602

Change-Id: I28b6df5e2523fc2ece6fd8251fcabc83c7c38d89
Reviewed-on: https://go-review.googlesource.com/c/go/+/256478
Trust: Ian Lance Taylor <iant@golang.org>
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agogo/build: replace os.Setenv with T.Setenv
Manlio Perillo [Wed, 14 Apr 2021 10:45:42 +0000 (12:45 +0200)]
go/build: replace os.Setenv with T.Setenv

Updates #45448

Change-Id: I4d8e5d7e57818355ef2bc33b57ddf9c8b8da3e62
Reviewed-on: https://go-review.googlesource.com/c/go/+/310030
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Michael Knyszek <mknyszek@google.com>

4 years agotime: replace os.Setenv with T.Setenv
Manlio Perillo [Wed, 14 Apr 2021 13:01:56 +0000 (15:01 +0200)]
time: replace os.Setenv with T.Setenv

Updates #45448

Change-Id: Ic096fe1c58c124fb8d84ee15c9446e7ed060b24f
Reviewed-on: https://go-review.googlesource.com/c/go/+/310032
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Michael Knyszek <mknyszek@google.com>

4 years agonet/http/httptest: panic on non-3 digit (XXX) status code in Recorder.WriteHeader
ian woolf [Sat, 10 Apr 2021 07:41:13 +0000 (15:41 +0800)]
net/http/httptest: panic on non-3 digit (XXX) status code in Recorder.WriteHeader

This change conforms Recorder with net/http servers, to panic
when a handler writes a non-3 digit XXX status code.

Fixes #45353

Change-Id: Id5ed4af652e8c150ae86bf50402b800d935e2203
Reviewed-on: https://go-review.googlesource.com/c/go/+/308950
Reviewed-by: Emmanuel Odeke <emmanuel@orijtech.com>
Run-TryBot: Emmanuel Odeke <emmanuel@orijtech.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: Michael Knyszek <mknyszek@google.com>

4 years agocmd/go: add a Go source file in TestScript/mod_sumdb
Bryan C. Mills [Mon, 12 Apr 2021 14:41:56 +0000 (10:41 -0400)]
cmd/go: add a Go source file in TestScript/mod_sumdb

This test expects 'go mod tidy' to fail if the existing module graph
has a bad checksum. However, there is no intrinsic reason why 'go mod
tidy' should fail in that case: the module contains no packages, and
thus no imports, so 'go mod tidy' can justifiably remove all
requirements without regard to any errors that may have already been
present in the module graph.

Adding a source file that imports a package from the module with the
bad checksum should guarantee that 'go mod tidy' reports the checksum
eror.

For #36460

Change-Id: I59734ac524031288bc03a11f58eed5abe2db76b3
Reviewed-on: https://go-review.googlesource.com/c/go/+/309334
Trust: Bryan C. Mills <bcmills@google.com>
Run-TryBot: Bryan C. Mills <bcmills@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Jay Conrod <jayconrod@google.com>
4 years agocmd/compile: rework/reduce partially lived argument spilling
Cherry Zhang [Sun, 11 Apr 2021 16:42:49 +0000 (12:42 -0400)]
cmd/compile: rework/reduce partially lived argument spilling

In CL 307909 we generate code that spills pointer-typed argument
registers if it is part of an SSA-able aggregate. The current
code spill the register unconditionally. Sometimes it is
unnecessary, because it is already spilled, or it is never live.

This CL reworks the spill generation. We move it to the end of
compilation, after liveness analysis, so we have information about
if a spill is necessary, and only generate spills for the
necessary ones.

Change-Id: I8d60be9b2c47651aeda14f5e2d1bbd207c134b26
Reviewed-on: https://go-review.googlesource.com/c/go/+/309331
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agoruntime: fix data race in abi finalizer test
Than McIntosh [Wed, 14 Apr 2021 19:40:01 +0000 (15:40 -0400)]
runtime: fix data race in abi finalizer test

Fix a buglet in TestFinalizerRegisterABI that was causing problems
when run with GOEXPERIMENT=regabi.

Updates #40724.

Change-Id: I950c4aa3df69eada23e600f01e6804eb136ce806
Reviewed-on: https://go-review.googlesource.com/c/go/+/310077
Trust: Than McIntosh <thanm@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime: update debug call protocol for register ABI
Michael Anthony Knyszek [Sat, 10 Apr 2021 00:05:07 +0000 (00:05 +0000)]
runtime: update debug call protocol for register ABI

The debug call tests currently assume that the target Go function is
ABI0; this is clearly no longer true when we switch to the new ABI, so
make the tests set up argument register state in the debug call handler
and copy back results returned in registers.

A small snag in calling a Go function that follows the new ABI is that
the debug call protocol depends on the AX register being set to a
specific value as it bounces in and out of the handler, but this
register is part of the new register ABI, so results end up being
clobbered. Use R12 instead.

Next, the new desugaring behavior for "go" statements means that
newosproc1 must always call a function with no frame; if it takes any
arguments, it closes over them and they're passed in the context
register. Currently when debugCallWrap creates a new goroutine, it uses
newosproc1 directly and passes a non-zero-sized frame, so that needs to
be updated. To fix this, briefly use the g's param field which is
otherwise only used for channels to pass an explicitly allocated object
containing the "closed over" variables. While we could manually do the
desugaring ourselves (we cannot do so automatically because the Go
compiler prevents heap-allocated closures in the runtime), that bakes in
more ABI details in a place that really doesn't need to care about them.

Finally, there's an old bug here where the context register was set up
in CX, so technically closure calls never worked. Oops. It was otherwise
harmless for other types of calls before, but now CX is an argument
register, so now that interferes with regular calls, too.

For #40724.

Change-Id: I652c25ed56a25741bb04c24cfb603063c099edde
Reviewed-on: https://go-review.googlesource.com/c/go/+/309169
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Alessandro Arzilli <alessandro.arzilli@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agogo/internal/gccgoimporter: replace os.MkdirTemp with T.TempDir
Manlio Perillo [Tue, 13 Apr 2021 10:52:35 +0000 (12:52 +0200)]
go/internal/gccgoimporter: replace os.MkdirTemp with T.TempDir

Updates #45402

Change-Id: Ie84795ed456883c0558fa9b5e3f2186f5f2c0fd9
Reviewed-on: https://go-review.googlesource.com/c/go/+/309356
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agotime: move slim test tzdata to testdata directory
Martin Sucha [Mon, 12 Apr 2021 05:57:25 +0000 (07:57 +0200)]
time: move slim test tzdata to testdata directory

The lines with tzdata are quite long.

It is easier to just copy a file than to encode it as
a string literal when adding a new test case.

Change-Id: Ibcaf347c3101a0f05b094e582b3473c7c35b685a
Reviewed-on: https://go-review.googlesource.com/c/go/+/308998
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agogo/build: replace os.MkdirTemp with T.TempDir
Manlio Perillo [Tue, 13 Apr 2021 10:25:19 +0000 (12:25 +0200)]
go/build: replace os.MkdirTemp with T.TempDir

Updates #45402

Change-Id: Ic2f696837034de17333a6a53127a4bfd301e96a4
Reviewed-on: https://go-review.googlesource.com/c/go/+/309354
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agocmd/compile/internal/types2: add Named.SetTParams and Named.Orig methods
Robert Griesemer [Wed, 14 Apr 2021 00:48:45 +0000 (17:48 -0700)]
cmd/compile/internal/types2: add Named.SetTParams and Named.Orig methods

Named.SetTParams sets the type parameters for a named type.

Named.Orig returns the original generic type an instantiated
type is derived from. Added a new field orig for that purpose
and renamed the already existing orig field to fromRHS.

Finally, updated various comments.

Change-Id: Ic9d173e42740422d195713d8bdc62a54dc8c5f54
Reviewed-on: https://go-review.googlesource.com/c/go/+/309832
Trust: Robert Griesemer <gri@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
4 years agocmd/dist: add tests using the typeparams build tag
Rob Findley [Wed, 14 Apr 2021 14:52:30 +0000 (10:52 -0400)]
cmd/dist: add tests using the typeparams build tag

Now that go/ast changes have been guarded behind the typeparams build
tag, we no longer have coverage for tests involving generic code.

Add a new testing step to cmd/dist to run go/... and cmd/gofmt tests
using -tags=typeparams.

Comment out parser object resolution assertions that currently fail, and
which will be fixed by CL 304456.

Fixes #44933

Change-Id: I481dd4246a016f410307865b6c6c2bb3c8e6e3bc
Reviewed-on: https://go-review.googlesource.com/c/go/+/310071
Trust: Robert Findley <rfindley@google.com>
Run-TryBot: Robert Findley <rfindley@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
4 years agogo/ast: fix broken build with typeparams build constraint
Rob Findley [Wed, 14 Apr 2021 13:34:28 +0000 (09:34 -0400)]
go/ast: fix broken build with typeparams build constraint

My rebase of https://golang.org/cl/300649 before submitting broke the
build (and tests) when using the typeparams build constraint. In a
subsequent CL I add test coverage back to cmd/dist.

This time, I've tested by running:
 - go test -tags=typeparams go/...
 - go test -tags=typeparams cmd/gofmt

All tests pass except for the new TestResolution/typeparams.go2, which I
will fix in a follow-up CL.

For #44933

Change-Id: I439d387841604cf43a90e2ce41dbe6bbbdb0306d
Reviewed-on: https://go-review.googlesource.com/c/go/+/310070
Trust: Robert Findley <rfindley@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Findley <rfindley@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
4 years agoos/exec: replace os.Setenv with T.Setenv
Manlio Perillo [Wed, 14 Apr 2021 10:57:21 +0000 (12:57 +0200)]
os/exec: replace os.Setenv with T.Setenv

Updates #45448

Change-Id: I570e9894c4976d0a875388ef9ea4a2aa31b78013
Reviewed-on: https://go-review.googlesource.com/c/go/+/310031
Trust: Tobias Klauser <tobias.klauser@gmail.com>
Run-TryBot: Tobias Klauser <tobias.klauser@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agocmd/compile: rescue stmt boundaries from OpArgXXXReg and OpSelectN.
David Chase [Tue, 13 Apr 2021 00:53:03 +0000 (20:53 -0400)]
cmd/compile: rescue stmt boundaries from OpArgXXXReg and OpSelectN.

Fixes this failure:
go test cmd/compile/internal/ssa -run TestStmtLines -v
=== RUN   TestStmtLines
    stmtlines_test.go:115: Saw too many (amd64, > 1%) lines without
    statement marks, total=88263, nostmt=1930
    ('-run TestStmtLines -v' lists failing lines)

The failure has two causes.

One is that the first-line adjuster in code generation was relocating
"first lines" to instructions that would either not have any code generated,
or would have the statment marker removed by a different believed-good heuristic.

The other was that statement boundaries were getting attached to register
values (that with the old ABI were loads from the stack, hence real instructions).
The register values disappear at code generation.

The fixes are to (1) note that certain instructions are not good choices for
"first value" and skip them, and (2) in an expandCalls post-pass, look for
register valued instructions and under appropriate conditions move their
statement marker to a compatible use.

Also updates TestStmtLines to always log the score, for easier comparison of
minor compiler changes.

Updates #40724.

Change-Id: I485573ce900e292d7c44574adb7629cdb4695c3f
Reviewed-on: https://go-review.googlesource.com/c/go/+/309649
Trust: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agocmd/internal/obj: don't emit args_stackmap for ABIInternal asm funcs
Than McIntosh [Tue, 13 Apr 2021 17:49:55 +0000 (13:49 -0400)]
cmd/internal/obj: don't emit args_stackmap for ABIInternal asm funcs

The compiler currently emits an "*.args_stackmap" symbol for all
bodyless functions, so that asm functions will have the proper stack
map. At the moment the code in the compiler that emits args_stackmap
assumes ABI0; to avoid misleading stackmaps, turn off args_stackmap
generation for non-ABI0 asm functions.

Updates #40724.

Change-Id: Ia5e3528d56da5fb107e799bd658e52496ba4a331
Reviewed-on: https://go-review.googlesource.com/c/go/+/309790
Trust: Than McIntosh <thanm@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agocmd/asm: require NOSPLIT for ABIInternal asm functions
Than McIntosh [Tue, 13 Apr 2021 17:46:52 +0000 (13:46 -0400)]
cmd/asm: require NOSPLIT for ABIInternal asm functions

Change the assembler to enforce the requirement that ABIInternal
functions need to be NOSPLIT. At the moment all of the assembly
routines in the runtime that are defined as ABIInternal also
happen to be NOSPLIT, but this CL makes it mandatory.

Updates #40724.

Change-Id: Ief80d22de1782edb44b798fcde9aab8a93548722
Reviewed-on: https://go-review.googlesource.com/c/go/+/309789
Trust: Than McIntosh <thanm@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agoreflect: keep pointer register results alive in callMethod
Cherry Zhang [Wed, 14 Apr 2021 03:41:45 +0000 (23:41 -0400)]
reflect: keep pointer register results alive in callMethod

When callMethod calls the underlying method, after reflectcall
it gets the result registers in "Ints" slots but not in "Ptrs"
slots. If the GC runs at this point, it may lose track of those
pointers and free the memory they point to.

To make sure the GC sees the pointer results, copy "Ints" to
"Ptrs", and keep them alive until we return to the caller.

This fixes test/fixedbugs/issue27695.go with register ABI.

Change-Id: I4092c91bcbd6954683740a12d91d689900446875
Reviewed-on: https://go-review.googlesource.com/c/go/+/309909
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agocmd/go: clarify comment on HashSeed
Austin Clements [Wed, 14 Apr 2021 15:33:46 +0000 (11:33 -0400)]
cmd/go: clarify comment on HashSeed

Change-Id: I3c8769f52b86ca82611af373b0a493a937dda9d5
Reviewed-on: https://go-review.googlesource.com/c/go/+/310090
Trust: Austin Clements <austin@google.com>
Reviewed-by: Bryan C. Mills <bcmills@google.com>
4 years agocmd/go/internal/modload: fix truncated error message from goModDirtyError
Bryan C. Mills [Wed, 14 Apr 2021 14:16:51 +0000 (10:16 -0400)]
cmd/go/internal/modload: fix truncated error message from goModDirtyError

The 'go mod tidy' hint was truncated due to a typo in CL 293689,
and that particular case was not covered by any existing test.

Updates #36460
Updates #40775

Change-Id: Ib6fa872a9dfdafc4e9a112e8add2ff5aecd2dbd0
Reviewed-on: https://go-review.googlesource.com/c/go/+/310089
Trust: Bryan C. Mills <bcmills@google.com>
Run-TryBot: Bryan C. Mills <bcmills@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agoruntime: incorporate hbits advancement in scanobject into loop
Josh Bleecher Snyder [Mon, 15 Mar 2021 23:15:23 +0000 (16:15 -0700)]
runtime: incorporate hbits advancement in scanobject into loop

This makes it clearer that i and hbits advance together.
As a bonus, it generates slightly better code.

Change-Id: I24d51102535c39f962a59c1a4a7c5c894339aa18
Reviewed-on: https://go-review.googlesource.com/c/go/+/309569
Trust: Josh Bleecher Snyder <josharian@gmail.com>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
4 years agoruntime: make gcEffectiveGrowthRatio a method on gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 19:12:02 +0000 (19:12 +0000)]
runtime: make gcEffectiveGrowthRatio a method on gcControllerState

For #44167.

Change-Id: Ie3cf8d2960c843a782ec85426fa73c279adaed64
Reviewed-on: https://go-review.googlesource.com/c/go/+/306605
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: pass work.userForced to gcController.endCycle explicitly
Michael Anthony Knyszek [Thu, 1 Apr 2021 19:09:40 +0000 (19:09 +0000)]
runtime: pass work.userForced to gcController.endCycle explicitly

For #44167.

Change-Id: I15817006f1870d6237cd06dabad988da3f23a6d6
Reviewed-on: https://go-review.googlesource.com/c/go/+/306604
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: move next_gc and last_next_gc into gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 18:38:14 +0000 (18:38 +0000)]
runtime: move next_gc and last_next_gc into gcControllerState

This change moves next_gc and last_next_gc into gcControllerState under
the names heapGoal and lastHeapGoal respectively. These are
fundamentally GC pacer related values, and so it makes sense for them to
live here.

Partially generated by

rf '
    ex . {
memstats.next_gc -> gcController.heapGoal
memstats.last_next_gc -> gcController.lastHeapGoal
    }
'

except for updates to comments and gcControllerState methods, where
they're accessed through the receiver, and trace-related renames of
NextGC -> HeapGoal, while we're here.

For #44167.

Change-Id: I1e871ad78a57b01be8d9f71bd662530c84853bed
Reviewed-on: https://go-review.googlesource.com/c/go/+/306603
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: fix formatting of gcMark
Leonard Wang [Fri, 9 Apr 2021 15:56:44 +0000 (23:56 +0800)]
runtime: fix formatting of gcMark

Change-Id: I08aed75f3aab0da705544665e532f332adfb075e
Reviewed-on: https://go-review.googlesource.com/c/go/+/308949
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Trust: David Chase <drchase@google.com>

4 years agoruntime: move roots' bases calculation to gcMarkRootPrepare
Ruslan Andreev [Tue, 22 Dec 2020 11:22:14 +0000 (19:22 +0800)]
runtime: move roots' bases calculation to gcMarkRootPrepare

This patch provides changes according to Austin's TODO. It just moves
calculation of base indexes of each root type from markroot function
to gcMarkRootPrepare.

Change-Id: Ib231de34e7f81e922762fc3ee2b1830921c0c7cf
Reviewed-on: https://go-review.googlesource.com/c/go/+/279461
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Austin Clements <austin@google.com>
4 years agoruntime: increase maxargs to avoid syscall18 crash when called with more than 16...
El Mostafa Idrassi [Tue, 13 Apr 2021 20:08:46 +0000 (20:08 +0000)]
runtime: increase maxargs to avoid syscall18 crash when called with more than 16 args

Fixes #45524

Change-Id: Id867f45ea98689b73d5b1b141c19317bc7608b05
GitHub-Last-Rev: e9b09fb557dda291fb6cf27c185063c26832a15b
GitHub-Pull-Request: golang/go#45531
Reviewed-on: https://go-review.googlesource.com/c/go/+/309390
Reviewed-by: El Mostafa Idrassi <el.mostafa.idrassi@gmail.com>
Reviewed-by: Alex Brainman <alex.brainman@gmail.com>
Trust: Alex Brainman <alex.brainman@gmail.com>
Trust: Alberto Donizetti <alb.donizetti@gmail.com>
Run-TryBot: Alex Brainman <alex.brainman@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agosyscall: don't defer close raw Socketpair fds in tests
Tobias Klauser [Tue, 13 Apr 2021 09:01:13 +0000 (11:01 +0200)]
syscall: don't defer close raw Socketpair fds in tests

The raw fds are successively wrapped using os.NewFile and will be closed
by (*os.File).Close. Avoids a double close, in the worst case closing an
unrelated fd.

Change-Id: I86aabe5ed865eff43d264ddae1fb07c935868e97
Reviewed-on: https://go-review.googlesource.com/c/go/+/309353
Trust: Tobias Klauser <tobias.klauser@gmail.com>
Run-TryBot: Tobias Klauser <tobias.klauser@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agocmd/compile: fix importing of method expressions
Keith Randall [Tue, 13 Apr 2021 23:31:56 +0000 (16:31 -0700)]
cmd/compile: fix importing of method expressions

For OMETHEXPR, the Name in the Selection needs to be properly
linked up to the method declaration. Use the same code we
already have for ODOTMETH and OCALLPART to do that.

Fixes #45503

Change-Id: I7d6f886d606bae6faad8c104f50c177f871d41c8
Reviewed-on: https://go-review.googlesource.com/c/go/+/309831
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Dan Scales <danscales@google.com>
4 years agoruntime: create setGCPercent method for gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 18:09:59 +0000 (18:09 +0000)]
runtime: create setGCPercent method for gcControllerState

This change breaks out the computations done by setGCPercent into
a method on gcControllerState for easier testing later. It leaves behind
the global implementation details.

For #44167.

Change-Id: I3b0cf1475b032fcd4ebbd01cf4e80de0b55ce7b0
Reviewed-on: https://go-review.googlesource.com/c/go/+/306602
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: create initializer for gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 18:01:46 +0000 (18:01 +0000)]
runtime: create initializer for gcControllerState

Now that gcControllerState contains almost all of the pacer state,
create an initializer for it instead of haphazardly setting some fields.

For #44167.

Change-Id: I4ce1d5dd82003cb7c263fa46697851bb22a32544
Reviewed-on: https://go-review.googlesource.com/c/go/+/306601
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: move gcPercent and heapMinimum into gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 17:56:32 +0000 (17:56 +0000)]
runtime: move gcPercent and heapMinimum into gcControllerState

These variables are core to the pacer, and will be need to be non-global
for testing later.

Partially generated via

rf '
    ex . {
gcPercent -> gcController.gcPercent
heapMinimum -> gcController.heapMinimum
    }
'

The only exception to this generation is usage of these variables
in gcControllerState methods.

For #44167.

Change-Id: I8b620b3061114f3a3c4b65006f715fd977b180a8
Reviewed-on: https://go-review.googlesource.com/c/go/+/306600
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: make gcSetTriggerRatio a method of gcControllerState
Michael Anthony Knyszek [Thu, 1 Apr 2021 16:31:29 +0000 (16:31 +0000)]
runtime: make gcSetTriggerRatio a method of gcControllerState

gcSetTriggerRatio's purpose is to set a bunch of downstream values when
we choose to commit to a new trigger ratio computed by the gcController.
Now that almost all the inputs it uses to compute the downstream values
are in gcControllerState anyway, make it a method of gcControllerState.

For #44167.

Change-Id: I1b7ea709e8378566f812ae3450ab169d7fb66aea
Reviewed-on: https://go-review.googlesource.com/c/go/+/306599
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agocmd/compile: set types properly for imported funcs with closures
Dan Scales [Sun, 11 Apr 2021 16:47:13 +0000 (09:47 -0700)]
cmd/compile: set types properly for imported funcs with closures

For the new export/import of node types, we were just missing setting
the types of the closure variables (which have the same types as the
captured variables) and the OCLOSURE node itself (which has the same
type as the Func node).

Re-enabled inlining of functions with closures.

Change-Id: I687149b061f3ffeec3244ff02dc6e946659077a9
Reviewed-on: https://go-review.googlesource.com/c/go/+/308974
Trust: Dan Scales <danscales@google.com>
Run-TryBot: Dan Scales <danscales@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
4 years agocmd/compile/internal/types2: use a global atomic counter for type parameter ids
Robert Griesemer [Tue, 13 Apr 2021 23:33:42 +0000 (16:33 -0700)]
cmd/compile/internal/types2: use a global atomic counter for type parameter ids

This avoids a dependency on a *Checker when we create type parameters
outside the type checker proper, e.g. in an importer. There may be
better solutions but this does the trick for now.

Change-Id: Icf22c934970cb04c88c2729555ae6a79ef5a2245
Reviewed-on: https://go-review.googlesource.com/c/go/+/309830
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Griesemer <gri@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
4 years agoruntime, cgo/test: improve debugging output
David Chase [Tue, 13 Apr 2021 18:12:43 +0000 (14:12 -0400)]
runtime, cgo/test: improve debugging output

tests that run commands should log their actions in a
shell-pasteable way.

Change-Id: Ifeee88397047ef5a76925c5f30c213e83e535038
Reviewed-on: https://go-review.googlesource.com/c/go/+/309770
Trust: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agoruntime: move internal GC statistics from memstats to gcController
Michael Anthony Knyszek [Wed, 31 Mar 2021 22:55:06 +0000 (22:55 +0000)]
runtime: move internal GC statistics from memstats to gcController

This change moves certain important but internal-only GC statistics from
memstats into gcController. These statistics are mainly used in pacing
the GC, so it makes sense to keep them in the pacer's state.

This CL was mostly generated via

rf '
    ex . {
memstats.gc_trigger -> gcController.trigger
memstats.triggerRatio -> gcController.triggerRatio
memstats.heap_marked -> gcController.heapMarked
memstats.heap_live -> gcController.heapLive
memstats.heap_scan -> gcController.heapScan
    }
'

except for a few special cases, like updating names in comments and when
these fields are used within gcControllerState methods (at which point
they're accessed through the reciever).

For #44167.

Change-Id: I6bd1602585aeeb80818ded24c07d8e6fec992b93
Reviewed-on: https://go-review.googlesource.com/c/go/+/306598
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agocmd/compile: always zero the temporary in mapKeyTemp
Cherry Zhang [Tue, 13 Apr 2021 22:01:59 +0000 (18:01 -0400)]
cmd/compile: always zero the temporary in mapKeyTemp

As we are taking its address, always zero it. In many cases the
temporary will be optimized out. But in case it does not (e.g. -N,
-race), this ensures it has the right liveness information.

May fix the noopt builder.

Change-Id: I3d5d617c276d2a1a1aaebff813b4cd60bc691592
Reviewed-on: https://go-review.googlesource.com/c/go/+/309771
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agocmd/compile: don't modify underlying type when creating bitmap for bodyless function
David Chase [Fri, 9 Apr 2021 22:05:57 +0000 (18:05 -0400)]
cmd/compile: don't modify underlying type when creating bitmap for bodyless function

This fixes the build crash for
    GOEXPERIMENT=regabi,regabiargs GOOS=windows go build syscall

Updates #40724.

Change-Id: I4400f6ff2e83e7e7e93ad5e58c6063b327532504
Reviewed-on: https://go-review.googlesource.com/c/go/+/309110
Trust: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
4 years agogo/*,cmd/gofmt: guard AST changes with the typeparams build tag
Rob Findley [Fri, 5 Mar 2021 15:57:48 +0000 (10:57 -0500)]
go/*,cmd/gofmt: guard AST changes with the typeparams build tag

This CL changes our approach to guarding type parameter functionality
and API. Previously, we guarded type parameter functionality with the
parser.parseTypeParams parser mode, and were in the process of hiding
the type parameter API behind the go1.18 build constraint.

These mechanisms had several limitations:
 + Requiring the parser.parseTypeParams mode to be set meant that
   existing tooling would have to opt-in to type parameters in all
   places where it parses Go files.
 + The parseTypeParams mode value had to be copied in several places.
 + go1.18 is not specific to typeparams, making it difficult to set up
   the builders to run typeparams tests.

This CL addresses the above limitations, and completes the task of
hiding the AST API, by switching to a new 'typeparams' build constraint
and adding a new go/internal/typeparams helper package.

The typeparams build constraint is used to conditionally compile the new
AST changes. The typeparams package provides utilities for accessing and
writing the new AST data, so that we don't have to fragment our parser
or type checker logic across build constraints. The typeparams.Enabled
const is used to guard tests that require type parameter support.

The parseTypeParams parser mode is gone, replaced by a new
typeparams.DisableParsing mode with the opposite sense. Now, type
parameters are only parsed if go/parser is compiled with the typeparams
build constraint set AND typeparams.DisableParsing not set. This new
parser mode allows opting out of type parameter parsing for tests.

How exactly to run tests on builders is left to a follow-up CL.

Updates #44933

Change-Id: I3091e42a2e5e2f23e8b2ae584f415a784b9fbd65
Reviewed-on: https://go-review.googlesource.com/c/go/+/300649
Trust: Robert Findley <rfindley@google.com>
Run-TryBot: Robert Findley <rfindley@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
4 years agoruntime: rename gcpercent, readgogc, and heapminimum to match Go style
Michael Anthony Knyszek [Wed, 31 Mar 2021 22:06:30 +0000 (22:06 +0000)]
runtime: rename gcpercent, readgogc, and heapminimum to match Go style

Generated with:

rf 'mv gcpercent gcPercent'
rf 'mv readgogc readGOGC'
rf 'mv heapminimum heapMinimum'

After this, comments referencing these symbols were updated via a simple
sed command.

For #44167.

Change-Id: I6bb01597c2130686c01f967d0f106b06860ad2db
Reviewed-on: https://go-review.googlesource.com/c/go/+/306597
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agoruntime: break out GC pacer into its own file
Michael Anthony Knyszek [Wed, 31 Mar 2021 21:47:41 +0000 (21:47 +0000)]
runtime: break out GC pacer into its own file

This change breaks out the GC pacer into its own file so that it'll be
easier to see the full implementation and change it. It also suggests an
obvious place to put tests (mgcpacer_test.go).

This includes all of gcControllerState, gcSetTriggerRatio, anything
related to GOGC, and all related globals and constants.

This is almost a clean move, except that globals and constants are
formatted into blocks instead of having a separate "var" declaration for
each one.

For #44167.

Change-Id: I85aa84ce85c6cfbe0b33e8a3c91cbe9dc41de8cb
Reviewed-on: https://go-review.googlesource.com/c/go/+/306596
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
4 years agocmd/compile: make map functions ABI insensitive
Cherry Zhang [Mon, 12 Apr 2021 20:12:20 +0000 (16:12 -0400)]
cmd/compile: make map functions ABI insensitive

Following CL 309029, this CL does the same thing for map
functions (mapaccess, mapassign, mapdelete).

For simplicity, always wrap "defer delete(m, k)". With
regabidefers enabled, this call is wrapped in a closure and the
rewriting happens automatically. Without regabidefers, it may not
be wrapped for certain key types, and then we'd need special
handling of the delete (because earlier the order pass does not
handle it). And we will turn on defer wrapping by default anyway.

Change-Id: I30663b1aa8e1d6f98e1fb81bf8c0c0ce607ab80b
Reviewed-on: https://go-review.googlesource.com/c/go/+/309510
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agoruntime: eliminate externalthreadhandler
Austin Clements [Tue, 13 Apr 2021 12:44:56 +0000 (08:44 -0400)]
runtime: eliminate externalthreadhandler

This function is no longer used.

Eliminating this actually fixes several problems:

- It made assumptions about what registers memclrNoHeapPointers would
  preserve. Besides being an abstraction violation and lurking
  maintenance issue, this actively became a problem for regabi because
  the call to memclrNoHeapPointers now happens through an ABI wrapper,
  which is generated by the compiler and hence we can't easily control
  what registers it clobbers.

- The amd64 implementation (at least), does not interact with the host
  ABI correctly. Notably, it doesn't save many of the registers that
  are callee-save in the host ABI but caller-save in the Go ABI.

- It interacts strangely with the NOSPLIT checker because it allocates
  an entire M and G on its stack. It worked around this on arm64, and
  happened to do things the NOSPLIT checker couldn't track on 386 and
  amd64, and happened to be *4 bytes* below the limit on arm (so any
  addition to the m or g structs would cause a NOSPLIT failure). See
  CL 309031 for a more complete explanation.

Fixes #45530.
Updates #40724.

Change-Id: Ic70d4d7e1c17f1d796575b3377b8529449e93576
Reviewed-on: https://go-review.googlesource.com/c/go/+/309634
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime: use newm for profileloop
Austin Clements [Tue, 13 Apr 2021 12:32:21 +0000 (08:32 -0400)]
runtime: use newm for profileloop

This replaces the externalthreadhandler-based implementation of
profileloop with one that uses newm to start a new thread. This is a
step toward eliminating externalthreadhandler.

For #45530.

Change-Id: Id8e5540423fe2d2004024b649afec6998f77b092
Reviewed-on: https://go-review.googlesource.com/c/go/+/309633
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime: use compileCallback for ctrlhandler
Austin Clements [Tue, 13 Apr 2021 12:20:08 +0000 (08:20 -0400)]
runtime: use compileCallback for ctrlhandler

This replaces the externalthreadhandler-based implementation of
ctrlhandler with one based on compileCallback. This is a step toward
eliminating externalthreadhandler.

For #45530.

Change-Id: I2de2f2f37777af292db67ccf8057b7566aab81f8
Reviewed-on: https://go-review.googlesource.com/c/go/+/309632
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoarchive/tar: replace os.MkdirTemp with T.TempDir
Manlio Perillo [Tue, 13 Apr 2021 10:35:28 +0000 (12:35 +0200)]
archive/tar: replace os.MkdirTemp with T.TempDir

Updates #45402

Change-Id: I296f8c676c68ed1e10b6ad1a17b5b23d2c395252
Reviewed-on: https://go-review.googlesource.com/c/go/+/309355
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Joe Tsai <thebrokentoaster@gmail.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Joe Tsai <thebrokentoaster@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/link: force external linking for DragonFly cgo programs
Ian Lance Taylor [Tue, 13 Apr 2021 00:30:36 +0000 (17:30 -0700)]
cmd/link: force external linking for DragonFly cgo programs

Change-Id: Idce4eec2a4698a94aa7b40590dffb6bf9bd45342
Reviewed-on: https://go-review.googlesource.com/c/go/+/309571
Trust: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Tobias Klauser <tobias.klauser@gmail.com>
4 years agocmd/compile,cmd/link: resolve cgo symbols to the correct Go ABI
Austin Clements [Sun, 11 Apr 2021 19:44:25 +0000 (15:44 -0400)]
cmd/compile,cmd/link: resolve cgo symbols to the correct Go ABI

Currently, Go functions exported to cgo have some confusion around
ABIs that leads to crashes. The cmd/cgo-generated C code references an
exported Go wrapper function (which calls the underlying exported user
function). The linker resolves this reference to the ABI0 entry-point
to that Go wrapper function because all host object references are
currently assumed to be to version 0 of a symbol. This gets passed via
crosscall2 and winds its way to cgocallbackg1, which puts this ABI0
entry-point into a function value and calls it. Unfortunately,
function values always use the ABIInternal calling convention, so
calling this ABI0 entry-point goes poorly.

Fix this by threading definition ABIs through the cgo export mechanism
so the linker can resolve host object references (which have no
concept of multiple ABIs) to the correct Go symbol. This involves a
few pieces:

- The compiler extends the cgo_export_{static,dynamic} directives that
  get passed on to the linker with symbol definition ABIs.

- The linker parses the ABIs in the cgo_export_{static,dynamic}
  directives to look up the right symbol to apply export attributes to
  and put in the dynexp list.

- For internal linking, the linker's Loader structure tracks the right
  symbol (in particular the right ABI) to resolve host object
  references to, and we use this in all of the host object loaders.

- For external linking, we mangle only the non-ABIInternal symbols
  now, so the external linker is able to resolve the correct reference
  from host objects to Go symbols.

Updates #40724.

Change-Id: I70a0b1610596768c3f473745fa1a3e630afbf1a8
Reviewed-on: https://go-review.googlesource.com/c/go/+/309341
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agocmd/link: build dynexp symbol list directly
Austin Clements [Mon, 12 Apr 2021 17:21:55 +0000 (13:21 -0400)]
cmd/link: build dynexp symbol list directly

Currently, setCgoAttr populates the cgo_export_{static,dynamic} maps
with symbol names of exported symbols, which are then re-looked-up by
deadcode and setupdynexp, which in turn puts the re-looked-up symbols
in ctxt.dynexp. setCgoAttr already looked up the Syms, so simplify all
of this by making setCgoAttr populate ctxt.dynexp directly and
eliminating the cgo_export_{static,dynamic} maps. Recording Syms
directly also sets us up to use correct symbol versions for these
exports, rather than just assuming version 0 for all lookups.

Since setupdynexp doesn't really do any "setting up" of dynexp any
more with this change, we fold the remaining logic from setupdynexp
directly into addexport, where it has better context anyway. This also
eliminates a sorting step, since we no longer do a non-deterministic
map iteration to build the dynexp slice.

For #40724.

Change-Id: I3e1a65165268da8c2bf50d7485f2624133433260
Reviewed-on: https://go-review.googlesource.com/c/go/+/309340
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agocmd/link: move cgo export map from loadcgo to setCgoAttr
Austin Clements [Sun, 11 Apr 2021 21:29:17 +0000 (17:29 -0400)]
cmd/link: move cgo export map from loadcgo to setCgoAttr

Currently, both loadcgo and setCgoAttr do some processing of
cgo_export_static and cgo_export_dynamic cgo directives, which means
they both have to parse them. There's no reason to do this in loadcgo,
so move all directive processing to setCgoAttr.

For #40724.

Change-Id: Icb3cdf7ef3517e866dd220e40a5f5dec7fd47e2b
Reviewed-on: https://go-review.googlesource.com/c/go/+/309339
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/link: refactor setCgoAttr
Austin Clements [Sun, 11 Apr 2021 20:30:36 +0000 (16:30 -0400)]
cmd/link: refactor setCgoAttr

setCgoAttr takes a lookup function, but there's only a single call and
setCgoAttr already has access to the lookup function passed at that
call. Simplify setCgoAttr by eliminating the lookup parameter and
calling the lookup function directly.

For #40724.

Change-Id: Ib27c0fa2b88c387e30423365f7757e3ba02cf7d5
Reviewed-on: https://go-review.googlesource.com/c/go/+/309338
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/cgo: document generated cgo directives
Austin Clements [Sun, 11 Apr 2021 21:19:59 +0000 (17:19 -0400)]
cmd/cgo: document generated cgo directives

This took me a while to figure out. Save the next person some trouble.

Change-Id: Ifab2d426f67c21b08ef225c79125805a9008e578
Reviewed-on: https://go-review.googlesource.com/c/go/+/309336
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agomime: in globs2 file only keep first time extension is seen
Ian Lance Taylor [Tue, 13 Apr 2021 05:26:37 +0000 (22:26 -0700)]
mime: in globs2 file only keep first time extension is seen

Fixes #45534

Change-Id: I9855607e845951f26ab85cb179ec6dea40d92156
Reviewed-on: https://go-review.googlesource.com/c/go/+/309574
Trust: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Tobias Klauser <tobias.klauser@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/compile/internal/types: add example test for type inference
Robert Griesemer [Sun, 11 Apr 2021 01:26:01 +0000 (18:26 -0700)]
cmd/compile/internal/types: add example test for type inference

Follow-up on https://golang.org/cl/308372.

For #44799.

Change-Id: I27d149241d465ee31e9c21d5f0935f31efcb5b60
Reviewed-on: https://go-review.googlesource.com/c/go/+/308973
Trust: Robert Griesemer <gri@golang.org>
Reviewed-by: Robert Findley <rfindley@google.com>
4 years agocmd/compile: allow OpArgXXXReg comes before LoweredGetClosurePtr
Cherry Zhang [Mon, 12 Apr 2021 23:42:28 +0000 (19:42 -0400)]
cmd/compile: allow OpArgXXXReg comes before LoweredGetClosurePtr

Both OpArgXXXReg and LoweredGetClosurePtr must come very early,
because they carry registers that are technically live on entry.
But no need to impose ordering requirement between them.

Change-Id: Iee1db6239a75e5b381e0ad25ba5503169333217b
Reviewed-on: https://go-review.googlesource.com/c/go/+/309629
Trust: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agotest: make codegen/memops.go work with both ABIs
Cherry Zhang [Mon, 12 Apr 2021 22:56:25 +0000 (18:56 -0400)]
test: make codegen/memops.go work with both ABIs

Following CL 309335, this fixes memops.go.

Change-Id: Ia2458b5267deee9f906f76c50e70a021ea2fcb5b
Reviewed-on: https://go-review.googlesource.com/c/go/+/309552
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agoall: simplify the spelling of Linux
Brad Fitzpatrick [Tue, 13 Apr 2021 02:05:01 +0000 (19:05 -0700)]
all: simplify the spelling of Linux

The prefix didn't help clarify anything.

Change-Id: I897fd4022ce9df42a548b15714e4b592618ca547
Reviewed-on: https://go-review.googlesource.com/c/go/+/309573
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Filippo Valsorda <filippo@golang.org>
4 years agocmd/link: support 32b TLS_LE offsets on PPC64
Paul E. Murphy [Mon, 15 Mar 2021 20:21:57 +0000 (15:21 -0500)]
cmd/link: support 32b TLS_LE offsets on PPC64

When using the GCC thread sanitizer, it links in additional
code which uses TLS, which causes us to exceed the range of
the 16 bit TLS relocation used by statically compiled go
code.

Rewrite objabi.R_POWER_TLS_LE to handle 32b offsets when
linking internally or externally into an ELF binary. The
elf relocation translation is changed to generate a pair
of R_PPC64_TPREL16_HA/LO relocations instead of a single
R_PPC64_TPREL16.

Likewise, updating the above exposed some behavioral differences
in gnu ld which can rewrite TLS sequences. It expects the
sequence to generate a valid TLS address, not offset. This was
exposed when compiling PIC code. The proper fix is to generate
the full TLS address in the destination register of the
"MOVD tlsaddr, $Rx" pseudo-op. This removes the need to insert
special objabi.R_POWER_TLS relocations elsewhere.

Unfortunately, XCOFF (used by aix) doesn't appear to support 32
bit offsets, so we rewrite this back into a 16b relocation when
externally linking a static binary.

Fixes #45040

Change-Id: I1ee9afd0b427cd79888032aa1f60d3e265073e1d
Reviewed-on: https://go-review.googlesource.com/c/go/+/302209
Run-TryBot: Paul Murphy <murp@ibm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/go: fix 'go help mod edit' JSON documentation
Jay Conrod [Mon, 12 Apr 2021 18:50:57 +0000 (14:50 -0400)]
cmd/go: fix 'go help mod edit' JSON documentation

The object representing a module directive may have a "Deprecated"
field but not a "Version" field. Other objects representing module
versions have "Path" and "Version" fields but not "Deprecated".

For #40357

Change-Id: Iad8063dfa6f7ceea22981a8a8f99e65fa3b7ffa0
Reviewed-on: https://go-review.googlesource.com/c/go/+/309337
Trust: Jay Conrod <jayconrod@google.com>
Run-TryBot: Jay Conrod <jayconrod@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Bryan C. Mills <bcmills@google.com>
4 years agocmd/compile: make interface conversion function selection ABI insensitive
Cherry Zhang [Fri, 9 Apr 2021 18:40:28 +0000 (14:40 -0400)]
cmd/compile: make interface conversion function selection ABI insensitive

Before register ABI, we always pass argument in memory, and the
compiler chooses interface conversion functions solely based on
the memory layout. As long as the two types have identical memory
layout, it is fine to mix and match, e.g. convT64 takes a uint64
argument, but it can be used for things like float64 or
struct { x [4]struct{}; y int64 }.

With register ABI, those types may be passed differently, e.g.
uint64 is passed in an integer register, float64 is passed in a
floating point register, the struct above is passed in memory.
I made a few attempts in the previous CLs to try to choose the
right function based on the argument type, but none of them is
really correct.

Instead, this CL changes it to always pass the argument in the
same type the runtime expects, and do conversion before the call
in the compiler. The conversion can be no-op (e.g. a named type
to its underlying type), direct (e.g. int64 to uint64), or
through memory (e.g. *(*uint64)(unsafe.Pointer(&arg))). This way,
the front end does not need to know the ABI. (It only needs to
know how to convert types, and it already does.)

TODO: do something similar for map functions.

Change-Id: I33fc780a47c3f332b765e09b5e527f52ea1d6b5c
Reviewed-on: https://go-review.googlesource.com/c/go/+/309029
Trust: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
4 years agoos: restore testErrNotExist's working directory on os.Chdir success
Grace Han [Fri, 9 Apr 2021 05:49:22 +0000 (15:49 +1000)]
os: restore testErrNotExist's working directory on os.Chdir success

The existing implementation calls os.Chdir expecting the call not to
succeed. This change restores the original working directory in the
case that the call does succeed.

Fixes #45407

Change-Id: I61c57f6858b9a9058226e45e24276c7af8913048
Reviewed-on: https://go-review.googlesource.com/c/go/+/308849
Trust: Emmanuel Odeke <emmanuel@orijtech.com>
Run-TryBot: Emmanuel Odeke <emmanuel@orijtech.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agotest: make codegen tests work with both ABIs
Cherry Zhang [Mon, 12 Apr 2021 18:00:49 +0000 (14:00 -0400)]
test: make codegen tests work with both ABIs

Some codegen tests were written with the assumption that
arguments and results are in memory, and with a specific stack
layout. With the register ABI, the assumption is no longer true.
Adjust the tests to work with both cases.

- For tests expecting in memory arguments/results, change to use
  global variables or memory-assigned argument/results.

- Allow more registers. E.g. some tests expecting register names
  contain only letters (e.g. AX), but  it can also contain numbers
  (e.g. R10).

- Some instruction selection changes when operate on register vs.
  memory, e.g. ADDQ vs. LEAQ, MOVB vs. MOVL. Accept both.

TODO: mathbits.go and memops.go still need fix.
Change-Id: Ic5932b4b5dd3f5d30ed078d296476b641420c4c5
Reviewed-on: https://go-review.googlesource.com/c/go/+/309335
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
4 years agodebug/pe: replace os.MkdirTemp with T.TempDir
Manlio Perillo [Mon, 12 Apr 2021 20:49:07 +0000 (22:49 +0200)]
debug/pe: replace os.MkdirTemp with T.TempDir

Updates #45402

Change-Id: I3d83a66270ca38e82d6bb7f8a1367af3d5343a98
Reviewed-on: https://go-review.googlesource.com/c/go/+/309352
Trust: Tobias Klauser <tobias.klauser@gmail.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
4 years agotext/template: replace os.MkdirTemp with T.TempDir
Manlio Perillo [Mon, 12 Apr 2021 20:24:01 +0000 (22:24 +0200)]
text/template: replace os.MkdirTemp with T.TempDir

Updates #45402

Change-Id: I9d55191c4021387b771550b5c93c91806f694aa6
Reviewed-on: https://go-review.googlesource.com/c/go/+/309351
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agointernal/execabs: replace ioutil.WriteFile with os.WriteFile
Manlio Perillo [Mon, 12 Apr 2021 19:46:14 +0000 (21:46 +0200)]
internal/execabs: replace ioutil.WriteFile with os.WriteFile

Fixes #45532.

Change-Id: I844acd50d6fa1ce918969bbb52f79dd7412d289f
Reviewed-on: https://go-review.googlesource.com/c/go/+/309350
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agoruntime: non-strict InlTreeIndex lookup in expandFinalInlineFrame
Michael Pratt [Fri, 9 Apr 2021 21:01:45 +0000 (17:01 -0400)]
runtime: non-strict InlTreeIndex lookup in expandFinalInlineFrame

This is a follow-up to golang.org/cl/301369, which made the same change
in Frames.Next. The same logic applies here: a profile stack may have
been truncated at an invalid PC provided by cgoTraceback.
expandFinalInlineFrame will then try to lookup the inline tree and
crash.

The same fix applies as well: upon encountering a bad PC, simply leave
it as-is and move on.

Fixes #44971
Fixes #45480

Change-Id: I2823c67a1f3425466b05384cc6d30f5fc8ee6ddc
Reviewed-on: https://go-review.googlesource.com/c/go/+/309109
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Trust: Michael Pratt <mpratt@google.com>

4 years agonet: never probe IPv4 map support on DragonFly BSD, OpenBSD
Josh Rickmar [Fri, 26 Mar 2021 13:17:36 +0000 (13:17 +0000)]
net: never probe IPv4 map support on DragonFly BSD, OpenBSD

DragonFly BSD and OpenBSD do not implement mapping IPv4 addresses to
the IPv6 address space, and a runtime check can be avoided.

As the IP stack capabilities probe was only being called from
supportsIPv4map to check for this support, the OS-specific handling
can be added to this function rather than continuing to run the probe.

Change-Id: I5800c197b1be502a6efa79e3edd6356bde8637fb
GitHub-Last-Rev: 7eb67189cd955fb12382a6f2a05f324d2747dcc3
GitHub-Pull-Request: golang/go#45243
Reviewed-on: https://go-review.googlesource.com/c/go/+/304870
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agomime: support reading shared mime-info database on unix systems
Jacalz [Fri, 9 Apr 2021 14:57:59 +0000 (14:57 +0000)]
mime: support reading shared mime-info database on unix systems

This adds support for reading the FreeDesktop Shared MIME-info Database on Unix systems, if it exists.
It should make lookups work on systems where the mime.types files are not present and
should lead to better mimetype lookup in general. If the shared mimetype database does not exist,
we will fall back to reading mime.types files in common locations.

Related to a bug on Solus bugtracker: https://dev.getsol.us/T9394
This change makes the mime package work on Solus.

Change-Id: If330c22ffe523bf31f7f10807a54fc8858517055
GitHub-Last-Rev: d5fbe8c41a9d975029b35498183a0f5a40df8e6a
GitHub-Pull-Request: golang/go#45271
Reviewed-on: https://go-review.googlesource.com/c/go/+/305230
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Tobias Klauser <tobias.klauser@gmail.com>

4 years agoruntime: consolidate "is sweep done" conditions
Austin Clements [Tue, 6 Apr 2021 23:25:28 +0000 (19:25 -0400)]
runtime: consolidate "is sweep done" conditions

The runtime currently has two different notions of sweep completion:

1. All spans are either swept or have begun sweeping.

2. The sweeper has *finished* sweeping all spans.

Having both is confusing (it doesn't help that the documentation is
often unclear or wrong). Condition 2 is stronger and the theoretical
slight optimization that condition 1 could impact is never actually
useful. Hence, this CL consolidates both conditions down to condition 2.

Updates #45315.

Change-Id: I55c84d767d74eb31a004a5619eaba2e351162332
Reviewed-on: https://go-review.googlesource.com/c/go/+/307916
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agoruntime: block sweep completion on all sweep paths
Austin Clements [Fri, 2 Apr 2021 19:54:24 +0000 (15:54 -0400)]
runtime: block sweep completion on all sweep paths

The runtime currently has two different notions of sweep completion:

1. All spans are either swept or have begun sweeping.

2. The sweeper has *finished* sweeping all spans.

Most things depend on condition 1. Notably, GC correctness depends on
condition 1, but since all sweep operations a non-preemptible, the STW
at the beginning of GC forces condition 1 to become condition 2.

runtime.GC(), however, depends on condition 2, since the intent is to
complete a complete GC cycle, and also update the heap profile (which
can only be done after sweeping is complete).

However, the way we compute condition 2 is racy right now and may in
fact only indicate condition 1. Specifically, sweepone blocks
condition 2 until all sweepone calls are done, but there are many
other ways to enter the sweeper that don't block this. Hence, sweepone
may see that there are no more spans in the sweep list and see that
it's the last sweepone and declare sweeping done, while there's some
other sweeper still working on a span.

Fix this by making sure every entry to the sweeper participates in the
protocol that blocks condition 2. To make sure we get this right, this
CL introduces a type to track sweep blocking and (lightly) enforces
span sweep ownership via the type system. This has the nice
side-effect of abstracting the pattern of acquiring sweep ownership
that's currently repeated in many different places.

Fixes #45315.

Change-Id: I7fab30170c5ae14c8b2f10998628735b8be6d901
Reviewed-on: https://go-review.googlesource.com/c/go/+/307915
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
4 years agocmd/link: fix TestLargeText
Lynn Boger [Fri, 9 Apr 2021 15:41:27 +0000 (10:41 -0500)]
cmd/link: fix TestLargeText

This test is not run in short mode so it was getting
failures that didn't happen with default testing. See
the issue for details on the failures.

Fixes #45406

Change-Id: I51d97cc4c910fe3ba2bc0a12742023a57d101f44
Reviewed-on: https://go-review.googlesource.com/c/go/+/308935
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Paul Murphy <murp@ibm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>

4 years agoruntime: port performance-critical functions to regabi
Austin Clements [Thu, 8 Apr 2021 21:43:51 +0000 (17:43 -0400)]
runtime: port performance-critical functions to regabi

This CL ports a few performance-critical runtime assembly functions to
use register arguments directly. While using the faster ABI is nice,
the real win here is that we avoid ABI wrappers: since these are
"builtin" functions in the compiler, it can generate calls to them
without knowing that their native implementation is ABI0. Hence, it
generates ABIInternal calls that go through ABI wrappers. By porting
them to use ABIInternal natively, we avoid the overhead of the ABI
wrapper.

This significantly improves performance on several benchmarks,
comparing regabiwrappers before and after this change:

name                                old time/op  new time/op  delta
BiogoIgor                            15.7s ± 2%   15.7s ± 2%    ~     (p=0.617 n=25+25)
BiogoKrishna                         18.5s ± 5%   17.7s ± 2%  -4.61%  (p=0.000 n=25+25)
BleveIndexBatch100                   5.91s ± 3%   5.82s ± 3%  -1.60%  (p=0.000 n=25+25)
BleveQuery                           6.76s ± 0%   6.60s ± 1%  -2.31%  (p=0.000 n=22+25)
CompileTemplate                      248ms ± 5%   245ms ± 1%    ~     (p=0.643 n=25+20)
CompileUnicode                      94.4ms ± 3%  93.9ms ± 2%    ~     (p=0.152 n=24+23)
CompileGoTypes                       1.60s ± 2%   1.59s ± 2%    ~     (p=0.059 n=24+24)
CompileCompiler                      104ms ± 3%   103ms ± 1%    ~     (p=0.056 n=25+22)
CompileSSA                           10.9s ± 1%   10.9s ± 1%    ~     (p=0.052 n=25+25)
CompileFlate                         156ms ± 8%   152ms ± 1%  -2.49%  (p=0.008 n=25+21)
CompileGoParser                      248ms ± 1%   249ms ± 2%    ~     (p=0.058 n=21+20)
CompileReflect                       595ms ± 3%   601ms ± 4%    ~     (p=0.182 n=25+25)
CompileTar                           211ms ± 2%   211ms ± 1%    ~     (p=0.663 n=23+23)
CompileXML                           282ms ± 2%   284ms ± 5%    ~     (p=0.456 n=21+23)
CompileStdCmd                        13.6s ± 2%   13.5s ± 2%    ~     (p=0.112 n=25+24)
FoglemanFauxGLRenderRotateBoat       8.69s ± 2%   8.67s ± 0%    ~     (p=0.094 n=22+25)
FoglemanPathTraceRenderGopherIter1   20.2s ± 2%   20.7s ± 3%  +2.53%  (p=0.000 n=24+24)
GopherLuaKNucleotide                 31.4s ± 1%   31.0s ± 1%  -1.28%  (p=0.000 n=25+24)
MarkdownRenderXHTML                  246ms ± 1%   244ms ± 1%  -0.79%  (p=0.000 n=20+21)
Tile38WithinCircle100kmRequest       843µs ± 4%   818µs ± 4%  -2.93%  (p=0.000 n=25+25)
Tile38IntersectsCircle100kmRequest  1.06ms ± 5%  1.05ms ± 3%  -1.19%  (p=0.021 n=24+25)
Tile38KNearestLimit100Request       1.01ms ± 1%  1.01ms ± 2%    ~     (p=0.335 n=22+25)
[Geo mean]                           596ms        592ms       -0.71%

(https://perf.golang.org/search?q=upload:20210411.5)

It also significantly reduces the performance penalty of enabling
regabiwrappers, though it doesn't yet fully close the gap on all
benchmarks:

name                                old time/op  new time/op  delta
BiogoIgor                            15.7s ± 1%   15.7s ± 2%    ~     (p=0.366 n=24+25)
BiogoKrishna                         17.7s ± 2%   17.7s ± 2%    ~     (p=0.315 n=23+25)
BleveIndexBatch100                   5.86s ± 4%   5.82s ± 3%    ~     (p=0.137 n=24+25)
BleveQuery                           6.55s ± 0%   6.60s ± 1%  +0.83%  (p=0.000 n=24+25)
CompileTemplate                      244ms ± 1%   245ms ± 1%    ~     (p=0.208 n=21+20)
CompileUnicode                      94.0ms ± 4%  93.9ms ± 2%    ~     (p=0.666 n=24+23)
CompileGoTypes                       1.60s ± 2%   1.59s ± 2%    ~     (p=0.154 n=25+24)
CompileCompiler                      103ms ± 1%   103ms ± 1%    ~     (p=0.905 n=24+22)
CompileSSA                           10.9s ± 2%   10.9s ± 1%    ~     (p=0.803 n=25+25)
CompileFlate                         153ms ± 1%   152ms ± 1%    ~     (p=0.182 n=23+21)
CompileGoParser                      250ms ± 2%   249ms ± 2%    ~     (p=0.843 n=24+20)
CompileReflect                       595ms ± 4%   601ms ± 4%    ~     (p=0.141 n=25+25)
CompileTar                           212ms ± 3%   211ms ± 1%    ~     (p=0.499 n=23+23)
CompileXML                           282ms ± 1%   284ms ± 5%    ~     (p=0.129 n=20+23)
CompileStdCmd                        13.5s ± 2%   13.5s ± 2%    ~     (p=0.480 n=24+24)
FoglemanFauxGLRenderRotateBoat       8.66s ± 1%   8.67s ± 0%    ~     (p=0.325 n=25+25)
FoglemanPathTraceRenderGopherIter1   20.6s ± 3%   20.7s ± 3%    ~     (p=0.137 n=25+24)
GopherLuaKNucleotide                 30.5s ± 2%   31.0s ± 1%  +1.68%  (p=0.000 n=23+24)
MarkdownRenderXHTML                  243ms ± 1%   244ms ± 1%  +0.51%  (p=0.000 n=23+21)
Tile38WithinCircle100kmRequest       801µs ± 2%   818µs ± 4%  +2.11%  (p=0.000 n=25+25)
Tile38IntersectsCircle100kmRequest  1.01ms ± 2%  1.05ms ± 3%  +4.34%  (p=0.000 n=24+25)
Tile38KNearestLimit100Request       1.00ms ± 1%  1.01ms ± 2%  +0.81%  (p=0.008 n=21+25)
[Geo mean]                           589ms        592ms       +0.50%

(https://perf.golang.org/search?q=upload:20210411.6)

Change-Id: I8f77f010b0abc658064df569a27a9c7a7b1c7bf9
Reviewed-on: https://go-review.googlesource.com/c/go/+/308931
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/compile: do not allocate space for unspilled in-register results
Cherry Zhang [Fri, 9 Apr 2021 22:02:48 +0000 (18:02 -0400)]
cmd/compile: do not allocate space for unspilled in-register results

For function results, if in register, we allocate spill slots
within the frame like locals. Currently, even if we never spill
to it the slot is still allocated. This CL makes it not allocate
the slot if it is never used.

Change-Id: Idbd4e3096cfac6d2bdfb501d8efde48ee2191d7b
Reviewed-on: https://go-review.googlesource.com/c/go/+/309150
Trust: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agointernal/poll: ensure that newPoolPipe doesn't return a nil pointer
Ian Lance Taylor [Mon, 12 Apr 2021 17:02:37 +0000 (10:02 -0700)]
internal/poll: ensure that newPoolPipe doesn't return a nil pointer

The function could occasionally return a nil pointer as a non-nil
interface, confusing the calling code.

Fixes #45520

Change-Id: Ifd35613728efa2cee9903177e85d369155074804
Reviewed-on: https://go-review.googlesource.com/c/go/+/309429
Trust: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Andy Pan <panjf2000@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>

4 years agocmd/compile: look for newobject in register ABI for write barrier elision
Cherry Zhang [Thu, 8 Apr 2021 21:46:21 +0000 (17:46 -0400)]
cmd/compile: look for newobject in register ABI for write barrier elision

If we are assigning a global address to an object that is
immediately returned from runtime.newobject, we omit the write
barrier because we know that both the source (static address) and
the destination (zeroed memory) do not need to be tracked by the
GC. Currently, the code that matches runtime.newobject's result
is specific to ABI0 layout. Update the code to work with register
ABI as well.

Change-Id: I7ab0833c6f745329271881ee4169956928a3a948
Reviewed-on: https://go-review.googlesource.com/c/go/+/308709
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
4 years agoruntime: replace outdated documentation link in Windows' nanotime
Paschalis Tsilias [Sun, 11 Apr 2021 10:13:11 +0000 (13:13 +0300)]
runtime: replace outdated documentation link in Windows' nanotime

Fixes #45498

Change-Id: I89365f3517bc84376f0f580c64a57f38aaba0cbb
Reviewed-on: https://go-review.googlesource.com/c/go/+/308997
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Trust: Alex Brainman <alex.brainman@gmail.com>

4 years agocmd/compile: preserve name association when eliding copies in expand_calls
Cherry Zhang [Sun, 11 Apr 2021 16:07:33 +0000 (12:07 -0400)]
cmd/compile: preserve name association when eliding copies in expand_calls

If v is a Copy of x, we will rewrite v to x. If v has a name
associated to it, let the name associate to x.

Under register ABI, this helps associate in-register Arg values
to the parameters' names. (But does not address all cases.)

Change-Id: I47c779e56c9d0823a88890497e32326bc0290f82
Reviewed-on: https://go-review.googlesource.com/c/go/+/309330
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
4 years agocmd/compile: support memmove inlining with register args
Cherry Zhang [Sun, 11 Apr 2021 18:33:28 +0000 (14:33 -0400)]
cmd/compile: support memmove inlining with register args

The rule that inlines memmove expects SSA ops that calls memmove
with arguments in memory. This CL adds a version that matches
it with arguments in registers, so the optimization works for
both situations.

Change-Id: Ideb64f65b7521481ab2ca7c9975a6cf7b70d5966
Reviewed-on: https://go-review.googlesource.com/c/go/+/309332
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: David Chase <drchase@google.com>