These operations (BT{S,R,C}{Q,L}modify) are quite a bit slower than
other ways of doing the same thing.
Without the BTxmodify operations, there are two fallback ways the compiler
performs these operations: AND/OR/XOR operations directly on memory, or
load-BTx-write sequences. The compiler kinda chooses one arbitrarily
depending on rewrite rule application order. Currently, it uses
load-BTx-write for the Const benchmarks and AND/OR/XOR directly to memory
for the non-Const benchmarks. TBD, someone might investigate which of
the two fallback strategies is really better. For now, they are both
better than BTx ops.
name old time/op new time/op delta
BitSet-8 1.09µs ± 2% 0.64µs ± 5% -41.60% (p=0.000 n=9+10)
BitClear-8 1.15µs ± 3% 0.68µs ± 6% -41.00% (p=0.000 n=10+10)
BitToggle-8 1.18µs ± 4% 0.73µs ± 2% -38.36% (p=0.000 n=10+8)
BitSetConst-8 37.0ns ± 7% 25.8ns ± 2% -30.24% (p=0.000 n=10+10)
BitClearConst-8 30.7ns ± 2% 25.0ns ±12% -18.46% (p=0.000 n=10+10)
BitToggleConst-8 36.9ns ± 1% 23.8ns ± 3% -35.46% (p=0.000 n=9+10)
Fixes#45790
Update #45242
Change-Id: Ie33a72dc139f261af82db15d446cd0855afb4e59
Reviewed-on: https://go-review.googlesource.com/c/go/+/318149
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ben Shi <powerman1st@163.com>
Currently, if we have AX=a and BX=b, and we want to make a call
F(1, a, b), to move arguments into the desired registers it emits
MOVQ AX, CX
MOVL $1, AX // AX=1
MOVQ BX, DX
MOVQ CX, BX // BX=a
MOVQ DX, CX // CX=b
This has a few redundant moves.
This is because we process inputs in order. First, allocate 1 to
AX, which kicks out a (in AX) to CX (a free register at the
moment). Then, allocate a to BX, which kicks out b (in BX) to DX.
Finally, put b to CX.
Notice that if we start with allocating CX=b, then BX=a, AX=1,
we will not have redundant moves. This CL reduces redundant moves
by allocating them in different order: First, for inpouts that are
already in place, keep them there. Then allocate free registers.
Then everything else.
before after
cmd/compile binary size 23703888 23609680
text size 8565899 8533291
(with regabiargs enabled.)
Change-Id: I69e1bdf745f2c90bb791f6d7c45b37384af1e874
Reviewed-on: https://go-review.googlesource.com/c/go/+/311371
Trust: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
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>
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>
Under certain circumstances, the existing rules for bit operations can
produce code that writes beyond its intended bounds. For example,
consider the following code:
func repro(b []byte, addr, bit int32) {
_ = b[3]
v := uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 | 1<<(bit&31)
b[0] = byte(v)
b[1] = byte(v >> 8)
b[2] = byte(v >> 16)
b[3] = byte(v >> 24)
}
Roughly speaking:
1. The expression `1 << (bit & 31)` is rewritten into `(SHLL 1 bit)`
2. The expression `uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 |
uint32(b[3])<<24` is rewritten into `(MOVLload &b[0])`
3. The statements `b[0] = byte(v) ... b[3] = byte(v >> 24)` are
rewritten into `(MOVLstore &b[0], v)`
4. `(ORL (SHLL 1, bit) (MOVLload &b[0]))` is rewritten into
`(BTSL (MOVLload &b[0]) bit)`. This is a valid transformation because
the destination is a register: in this case, the bit offset is masked
by the number of bits in the destination register. This is identical
to the masking performed by `SHL`.
5. `(MOVLstore &b[0] (BTSL (MOVLload &b[0]) bit))` is rewritten into
`(BTSLmodify &b[0] bit)`. This is an invalid transformation because
the destination is memory: in this case, the bit offset is not
masked, and the chosen instruction may write outside its intended
32-bit location.
These changes fix the invalid rewrite performed in step (5) by
explicitly maksing the bit offset operand to `BT(S|R|C)(L|Q)modify`. In
the example above, the adjusted rules produce
`(BTSLmodify &b[0] (ANDLconst [31] bit))` in step (5).
These changes also add several new rules to rewrite bit sets, toggles,
and clears that are rooted at `(OR|XOR|AND)(L|Q)modify` operators into
appropriate `BT(S|R|C)(L|Q)modify` operators. These rules catch cases
where `MOV(L|Q)store ((OR|XOR|AND)(L|Q) ...)` is rewritten to
`(OR|XOR|AND)(L|Q)modify` before the `(OR|XOR|AND)(L|Q) ...` can be
rewritten to `BT(S|R|C)(L|Q) ...`.
Overall, compilecmp reports small improvements in code size on
darwin/amd64 when the changes to the compiler itself are exlcuded:
file before after Δ %
runtime.s 536464 536412 -52 -0.010%
bytes.s 32629 32593 -36 -0.110%
strings.s 44565 44529 -36 -0.081%
os/signal.s 7967 7959 -8 -0.100%
cmd/vendor/golang.org/x/sys/unix.s 81686 81678 -8 -0.010%
math/big.s 188235 188253 +18 +0.010%
cmd/link/internal/loader.s 89295 89056 -239 -0.268%
cmd/link/internal/ld.s 633551 633232 -319 -0.050%
cmd/link/internal/arm.s 18934 18928 -6 -0.032%
cmd/link/internal/arm64.s 31814 31801 -13 -0.041%
cmd/link/internal/riscv64.s 7347 7345 -2 -0.027%
cmd/compile/internal/ssa.s 4029173 4033066 +3893 +0.097%
total 21298280 21301472 +3192 +0.015%
Change-Id: I2e560548b515865129e1724e150e30540e9d29ce
GitHub-Last-Rev: 9a42bd29a5
GitHub-Pull-Request: golang/go#45242
Reviewed-on: https://go-review.googlesource.com/c/go/+/304869
Reviewed-by: Keith Randall <khr@golang.org>
Trust: Josh Bleecher Snyder <josharian@gmail.com>
Optimize some patterns into rev16/rev16w instruction.
Pattern1:
(c & 0xff00ff00)>>8 | (c & 0x00ff00ff)<<8
To:
rev16w c
Pattern2:
(c & 0xff00ff00ff00ff00)>>8 | (c & 0x00ff00ff00ff00ff)<<8
To:
rev16 c
This patch is a copy of CL 239637, contributed by Alice Xu(dianhong.xu@arm.com).
Change-Id: I96936c1db87618bc1903c04221c7e9b2779455b3
Reviewed-on: https://go-review.googlesource.com/c/go/+/268377
Trust: fannie zhang <Fannie.Zhang@arm.com>
Run-TryBot: fannie zhang <Fannie.Zhang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
When -clobberdeadreg flag is set, the compiler inserts code that
clobbers integer registers at call sites. This may be helpful for
debugging register ABI.
Only implemented on AMD64 for now.
Change-Id: Ia203d3f891c30fd95d0103489056fe01d63a2899
Reviewed-on: https://go-review.googlesource.com/c/go/+/302809
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>
This CL adds rewrite rules for CSETM, CSINC, CSINV, and CSNEG. By adding
these rules, we can save one instruction.
For example,
func test(cond bool, a int) int {
if cond {
a++
}
return a
}
Before:
MOVD "".a+8(RSP), R0
ADD $1, R0, R1
MOVBU "".cond(RSP), R2
CMPW $0, R2
CSEL NE, R1, R0, R0
After:
MOVBU "".cond(RSP), R0
CMPW $0, R0
MOVD "".a+8(RSP), R0
CSINC EQ, R0, R0, R0
This patch is a copy of CL 285694. Co-authored-by: JunchenLi
<junchen.li@arm.com>
Change-Id: Ic1a79e8b8ece409b533becfcb7950f11e7b76f24
Reviewed-on: https://go-review.googlesource.com/c/go/+/302231
Trust: fannie zhang <Fannie.Zhang@arm.com>
Run-TryBot: fannie zhang <Fannie.Zhang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
This CL resurrects the clobberdead debugging mode (CL 23924).
When -clobberdead flag is set (TODO: make it GOEXPERIMENT?), the
compiler inserts code that clobbers all dead stack slots that
contains pointers.
Mark windows syscall functions cgo_unsafe_args, as the code
actually does that, by taking the address of one argument and
passing it to cgocall.
Change-Id: Ie09a015f4bd14ae6053cc707866e30ae509b9d6f
Reviewed-on: https://go-review.googlesource.com/c/go/+/301791
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>
These instructions are actually 5 argument opcodes as specified
by the ISA. Prior to this patch, the MB and ME arguments were
merged into a single bitmask operand to workaround the limitations
of the ppc64 assembler backend.
This limitation no longer exists. Thus, we can pass operands for
these opcodes without having to merge the MB and ME arguments in
the assembler frontend or compiler backend.
Likewise, support for 4 operand variants is unchanged.
Change-Id: Ib086774f3581edeaadfd2190d652aaaa8a90daeb
Reviewed-on: https://go-review.googlesource.com/c/go/+/298750
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>
Add generic rule to rewrite the single-precision square root expression
with one single-precision instruction. The optimization will reduce two
times of precision converting between double-precision and single-precision.
On arm64 flatform.
previous:
FCVTSD F0, F0
FSQRTD F0, F0
FCVTDS F0, F0
optimized:
FSQRTS S0, S0
And this patch adds the test case to check the correctness.
This patch refers to CL 241877, contributed by Alice Xu
(dianhong.xu@arm.com)
Change-Id: I6de5d02281c693017ac4bd4c10963dd55989bd7e
Reviewed-on: https://go-review.googlesource.com/c/go/+/276873
Trust: fannie zhang <Fannie.Zhang@arm.com>
Run-TryBot: fannie zhang <Fannie.Zhang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
The code generated when storing eight bytes loaded from memory created a
series of small writes instead of a single, large one. The specific
pattern of instructions generated stored 1 byte, then 2 bytes, then 4
bytes, and finally 1 byte.
The new rules match this specific pattern both for amd64 and for s390x,
and convert it into a single instruction to store the 8 bytes. arm64 and
ppc64le already generated the right code, but the new codegen test
covers also those architectures.
Fixes#41663
Change-Id: Ifb9b464be2d59c2ed5034acf7b9c3e473f344030
Reviewed-on: https://go-review.googlesource.com/c/go/+/280456
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
Trust: Josh Bleecher Snyder <josharian@gmail.com>
Trust: Jason A. Donenfeld <Jason@zx2c4.com>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Some bit test instruction generation stopped triggering after
the change to addressing modes. I suspect this was just because
ANDQload was being generated before the rewrite rules could discover
the BTQ. Fix that by decomposing the ANDQload when it is surrounded
by a TESTQ (thus re-enabling the BTQ rules).
Fixes#44228
Change-Id: I489b4a5a7eb01c65fc8db0753f8cec54097cadb2
Reviewed-on: https://go-review.googlesource.com/c/go/+/291749
Trust: Keith Randall <khr@golang.org>
Trust: Josh Bleecher Snyder <josharian@gmail.com>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
In ABIInternal, reserve X15 as constant zero, and use it to zero
memory. (Maybe there can be more use of it?)
The register is zeroed when transition to ABIInternal from ABI0.
Caveat: using X15 generates longer instructions than using X0.
Maybe we want to use X0?
Change-Id: I12d5ee92a01fc0b59dad4e5ab023ac71bc2a8b7d
Reviewed-on: https://go-review.googlesource.com/c/go/+/288093
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>
Current many architectures use a rule along the lines of
// Canonicalize the order of arguments to comparisons - helps with CSE.
((CMP|CMPW) x y) && x.ID > y.ID => (InvertFlags ((CMP|CMPW) y x))
to normalize comparisons as much as possible for CSE. Replace the
ID comparison with something less variable across compiler changes.
This helps avoid spurious failures in some of the codegen-comparison
tests (though the current choice of comparison is sensitive to Op
ordering).
Two tests changed to accommodate modified instruction choice.
Change-Id: Ib35f450bd2bae9d4f9f7838ceaf7ec682bcf1e1a
Reviewed-on: https://go-review.googlesource.com/c/go/+/280155
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>
Some rules for PPC64 were checking for a case
where a shift followed by an 'and' of a mask could
be lowered, depending on the format of the mask. The
function to verify if the mask was valid for this purpose
was not checking if the mask was 0 which we don't want to
allow. This case can happen if previous optimizations
resulted in that mask value.
This fixes isPPC64ValidShiftMask to check for a mask of 0 and return
false.
This also adds a codegen testcase to verify it doesn't try to
match the rules in the future.
Fixes#42610
Change-Id: I565d94e88495f51321ab365d6388c01e791b4dbb
Reviewed-on: https://go-review.googlesource.com/c/go/+/270358
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: Carlos Eduardo Seo <carlos.seo@linaro.org>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
Fixes#42445
Change-Id: I9653ef094dba2a1ac2e3daaa98279d10df17a2a1
Reviewed-on: https://go-review.googlesource.com/c/go/+/268257
Trust: Alberto Donizetti <alb.donizetti@gmail.com>
Trust: Martin Möhrmann <moehrmann@google.com>
Run-TryBot: Alberto Donizetti <alb.donizetti@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Martin Möhrmann <moehrmann@google.com>
Optimize combinations of left and right shifts by a constant value
into a 'rotate then insert selected bits [into zero]' instruction.
Use the same instruction for contiguous masks since it has some
benefits over 'and immediate' (not restricted to 32-bits, does not
overwrite source register).
To keep the complexity of this change under control I've only
implemented 64 bit operations for now.
There are a lot more optimizations that can be done with this
instruction family. However, since their function overlaps with other
instructions we need to be somewhat careful not to break existing
optimization rules by creating optimization dead ends. This is
particularly true of the load/store merging rules which contain lots
of zero extensions and shifts.
This CL does interfere with the store merging rules when an operand
is shifted left before it is stored:
binary.BigEndian.PutUint64(b, x << 1)
This is unfortunate but it's not critical and somewhat complex so
I plan to fix that in a follow up CL.
file before after Δ %
addr2line 4117446 4117282 -164 -0.004%
api 4945184 4942752 -2432 -0.049%
asm 4998079 4991891 -6188 -0.124%
buildid 2685158 2684074 -1084 -0.040%
cgo 4553732 4553394 -338 -0.007%
compile 19294446 19245070 -49376 -0.256%
cover 4897105 4891319 -5786 -0.118%
dist 3544389 3542785 -1604 -0.045%
doc 3926795 3927617 +822 +0.021%
fix 3302958 3293868 -9090 -0.275%
link 6546274 6543456 -2818 -0.043%
nm 4102021 4100825 -1196 -0.029%
objdump 4542431 4548483 +6052 +0.133%
pack 2482465 2416389 -66076 -2.662%
pprof 13366541 13363915 -2626 -0.020%
test2json 2829007 2761515 -67492 -2.386%
trace 10216164 10219684 +3520 +0.034%
vet 6773956 6773572 -384 -0.006%
total 107124151 106917891 -206260 -0.193%
Change-Id: I7591cce41e06867ba10a745daae9333513062746
Reviewed-on: https://go-review.googlesource.com/c/go/+/233317
Run-TryBot: Michael Munday <mike.munday@ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Trust: Michael Munday <mike.munday@ibm.com>
We already remove racefuncenter and racefuncexit if they are not
needed (i.e. the function doesn't have any other race calls).
racefuncenterfp is like racefuncenter but used on LR machines.
Remove unnecessary racefuncenterfp as well.
Change-Id: I65edb00e19c6d9ab55a204cbbb93e9fb710559f1
Reviewed-on: https://go-review.googlesource.com/c/go/+/267099
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>
Combine (AND m (SRWconst x)) or (SRWconst (AND m x)) when mask m is
and the shift value produce constant which can be encoded into an
RLWINM instruction.
Combine (CLRLSLDI (SRWconst x)) if the combining of the underling rotate
masks produces a constant which can be encoded into RLWINM.
Likewise for (SLDconst (SRWconst x)) and (CLRLSDI (RLWINM x)).
Combine rotate word + and operations which can be encoded as a single
RLWINM/RLWNM instruction.
The most notable performance improvements arise from the crypto
benchmarks below (GOARCH=power8 on a ppc64le/linux):
pkg:golang.org/x/crypto/blowfish goos:linux goarch:ppc64le
ExpandKeyWithSalt 52.2µs ± 0% 47.5µs ± 0% -8.88%
ExpandKey 44.4µs ± 0% 40.3µs ± 0% -9.15%
pkg:golang.org/x/crypto/ssh/internal/bcrypt_pbkdf goos:linux goarch:ppc64le
Key 57.6ms ± 0% 52.3ms ± 0% -9.13%
pkg:golang.org/x/crypto/bcrypt goos:linux goarch:ppc64le
Equal 90.9ms ± 0% 82.6ms ± 0% -9.13%
DefaultCost 91.0ms ± 0% 82.7ms ± 0% -9.12%
Change-Id: I59a0ca29face38f4ab46e37124c32906f216c4ce
Reviewed-on: https://go-review.googlesource.com/c/go/+/260798
Run-TryBot: Carlos Eduardo Seo <carlos.seo@linaro.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.com>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
Backstop support for non-sse2 chips now that 387 is gone.
RELNOTE=yes
Change-Id: Ib10e69c4a3654c15a03568f93393437e1939e013
Reviewed-on: https://go-review.googlesource.com/c/go/+/260017
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
My last 387 CL. So sad ... ... ... ... not!
Fixes#40255
Change-Id: I8d4ddb744b234b8adc735db2f7c3c7b6d8bbdfa4
Reviewed-on: https://go-review.googlesource.com/c/go/+/258957
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
A recent change to improve shifts was generating some
invalid cases when the rule was based on an AND. The
extended mnemonics CLRLSLDI and CLRLSLWI only allow
certain values for the operands and in the mask case
those values were not being checked properly. This
adds a check to those rules to verify that the
'b' and 'n' values used when an AND was part of the rule
have correct values.
There was a bug in some diag messages in asm9. The
message expected 3 values but only provided 2. Those are
corrected here also.
The test/codegen/shift.go was updated to add a few more
cases to check for the case mentioned here.
Some of the comments that mention the order of operands
in these extended mnemonics were wrong and those have been
corrected.
Fixes#41683.
Change-Id: If5bb860acaa5051b9e0cd80784b2868b85898c31
Reviewed-on: https://go-review.googlesource.com/c/go/+/258138
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Paul Murphy <murp@ibm.com>
Reviewed-by: Carlos Eduardo Seo <carlos.seo@gmail.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
This adds support for the extswsli instruction which combines
extsw followed by a shift.
New benchmark demonstrates the improvement:
name old time/op new time/op delta
ExtShift 1.34µs ± 0% 1.30µs ± 0% -3.15% (p=0.057 n=4+3)
Change-Id: I21b410676fdf15d20e0cbbaa75d7c6dcd3bbb7b0
Reviewed-on: https://go-review.googlesource.com/c/go/+/257017
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Carlos Eduardo Seo <carlos.seo@gmail.com>
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
This change adds rules to find pairs of instructions that can
be combined into a single shifts. These instruction sequences
are common in array addressing within loops. Improvements can
be seen in many crypto packages and the hash packages.
These are based on the extended mnemonics found in the ISA
sections C.8.1 and C.8.2.
Some rules in PPC64.rules were moved because the ordering prevented
some matching.
The following results were generated on power9.
hash/crc32:
CRC32/poly=Koopman/size=40/align=0 195ns ± 0% 163ns ± 0% -16.41%
CRC32/poly=Koopman/size=40/align=1 200ns ± 0% 163ns ± 0% -18.50%
CRC32/poly=Koopman/size=512/align=0 1.98µs ± 0% 1.67µs ± 0% -15.46%
CRC32/poly=Koopman/size=512/align=1 1.98µs ± 0% 1.69µs ± 0% -14.80%
CRC32/poly=Koopman/size=1kB/align=0 3.90µs ± 0% 3.31µs ± 0% -15.27%
CRC32/poly=Koopman/size=1kB/align=1 3.85µs ± 0% 3.31µs ± 0% -14.15%
CRC32/poly=Koopman/size=4kB/align=0 15.3µs ± 0% 13.1µs ± 0% -14.22%
CRC32/poly=Koopman/size=4kB/align=1 15.4µs ± 0% 13.1µs ± 0% -14.79%
CRC32/poly=Koopman/size=32kB/align=0 137µs ± 0% 105µs ± 0% -23.56%
CRC32/poly=Koopman/size=32kB/align=1 137µs ± 0% 105µs ± 0% -23.53%
crypto/rc4:
RC4_128 733ns ± 0% 650ns ± 0% -11.32% (p=1.000 n=1+1)
RC4_1K 5.80µs ± 0% 5.17µs ± 0% -10.89% (p=1.000 n=1+1)
RC4_8K 45.7µs ± 0% 40.8µs ± 0% -10.73% (p=1.000 n=1+1)
crypto/sha1:
Hash8Bytes 635ns ± 0% 613ns ± 0% -3.46% (p=1.000 n=1+1)
Hash320Bytes 2.30µs ± 0% 2.18µs ± 0% -5.38% (p=1.000 n=1+1)
Hash1K 5.88µs ± 0% 5.38µs ± 0% -8.62% (p=1.000 n=1+1)
Hash8K 42.0µs ± 0% 37.9µs ± 0% -9.75% (p=1.000 n=1+1)
There are other improvements found in golang.org/x/crypto which are all in the
range of 5-15%.
Change-Id: I193471fbcf674151ffe2edab212799d9b08dfb8c
Reviewed-on: https://go-review.googlesource.com/c/go/+/252097
Trust: Lynn Boger <laboger@linux.vnet.ibm.com>
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Carlos Eduardo Seo <cseo@linux.vnet.ibm.com>
The order array was zero initialized by the compiler, but ends up being
overwritten by the runtime anyway.
So let the runtime takes full responsibility for initializing, save us
one instruction per select.
Fixes#40399
Change-Id: Iec1eca27ad7180d4fcb3cc9ef97348206b7fe6b8
Reviewed-on: https://go-review.googlesource.com/c/go/+/251517
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
This merges an lis + subf into subfic, and for 32b constants
lwa + subf into oris + ori + subf.
The carry bit is no longer used in code generation, therefore
I think we can clobber it as needed. Note, lowered borrow/carry
arithmetic is self-contained and thus is not affected.
A few extra rules are added to ensure early transformations to
SUBFCconst don't trip up earlier rules, fold constant operations,
or otherwise simplify lowering. Likewise, tests are added to
ensure all rules are hit. Generic constant folding catches
trivial cases, however some lowering rules insert arithmetic
which can introduce new opportunities (e.g BitLen or Slicemask).
I couldn't find a specific benchmark to demonstrate noteworthy
improvements, but this is generating subfic in many of the default
bent test binaries, so we are at least saving a little code space.
Change-Id: Iad7c6e5767eaa9dc24dc1c989bd1c8cfe1982012
Reviewed-on: https://go-review.googlesource.com/c/go/+/249461
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Carlos Eduardo Seo <cseo@linux.vnet.ibm.com>
With this patch, opt pass can expose more obvious constant-folding
opportunites.
Example:
func test(i int) int {return (i+8)-(i+4)}
The previous version:
MOVD "".i(FP), R0
ADD $8, R0, R1
ADD $4, R0, R0
SUB R0, R1, R0
MOVD R0, "".~r1+8(FP)
RET (R30)
The optimized version:
MOVD $4, R0
MOVD R0, "".~r1+8(FP)
RET (R30)
This patch removes some existing reassociation rules, such as "x+(z-C)",
because the current generic rewrite rules will canonicalize "x-const"
to "x+(-const)", making "x+(z-C)" equal to "x+(z+(-C))".
This patch also adds test cases.
Change-Id: I857108ba0b5fcc18a879eeab38e2551bc4277797
Reviewed-on: https://go-review.googlesource.com/c/go/+/237137
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Add a new lowering rule to match and replace such instances
with the MADDLD instruction available on power9 where
possible.
Likewise, this plumbs in a new ppc64 ssa opcode to house
the newly generated MADDLD instructions.
When testing ed25519, this reduced binary size by 936B.
Similarly, MADDLD combination occcurs in a few other less
obvious cases such as division by constant.
Testing of golang.org/x/crypto/ed25519 shows non-trivial
speedup during keygeneration:
name old time/op new time/op delta
KeyGeneration 65.2µs ± 0% 63.1µs ± 0% -3.19%
Signing 64.3µs ± 0% 64.4µs ± 0% +0.16%
Verification 147µs ± 0% 147µs ± 0% +0.11%
Similarly, this test binary has shrunk by 66488B.
Change-Id: I077aeda7943119b41f07e4e62e44a648f16e4ad0
Reviewed-on: https://go-review.googlesource.com/c/go/+/248723
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
Some of the existing optimizations aren't triggered because they
are handled by the generic rules so this CL removes them. Also
some constraints were copied without much thought from the amd64
rules and they don't make sense on s390x, so we remove those
constraints.
Finally, add a 'multiply by the sum of two powers of two'
optimization. This makes sense on s390x as shifts are low latency
and can also sometimes be optimized further (especially if we add
support for RISBG instructions).
name old time/op new time/op delta
IntMulByConst/3-8 1.70ns ±11% 1.10ns ± 5% -35.26% (p=0.000 n=10+10)
IntMulByConst/5-8 1.64ns ± 7% 1.10ns ± 4% -32.94% (p=0.000 n=10+9)
IntMulByConst/12-8 1.65ns ± 6% 1.20ns ± 4% -27.16% (p=0.000 n=10+9)
IntMulByConst/120-8 1.66ns ± 4% 1.22ns ±13% -26.43% (p=0.000 n=10+10)
IntMulByConst/-120-8 1.65ns ± 7% 1.19ns ± 4% -28.06% (p=0.000 n=9+10)
IntMulByConst/65537-8 0.86ns ± 9% 1.12ns ±12% +30.41% (p=0.000 n=10+10)
IntMulByConst/65538-8 1.65ns ± 5% 1.23ns ± 5% -25.11% (p=0.000 n=10+10)
Change-Id: Ib196e6bff1e97febfd266134d0a2b2a62897989f
Reviewed-on: https://go-review.googlesource.com/c/go/+/248937
Run-TryBot: Michael Munday <mike.munday@ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
For an unsigned integer, it's useful to convert its order test with 0/1
to its equality test with 0. We can save a comparison instruction that
followed by a conditional branch on arm64 since it supports
compare-with-zero-and-branch instructions. For example,
if x > 0 { ... } else { ... }
the original version:
CMP $0, R0
BLS 9
the optimized version:
CBZ R0, 8
Updates #21439
Change-Id: Id1de6f865f6aa72c5d45b29f7894818857288425
Reviewed-on: https://go-review.googlesource.com/c/go/+/246857
Reviewed-by: Keith Randall <khr@golang.org>
If the AND has other uses, we end up saving an argument to the AND
in another register, so we can use it for the TEST. No point in doing that.
Change-Id: I73444a6aeddd6f55e2328ce04d77c3e6cf4a83e0
Reviewed-on: https://go-review.googlesource.com/c/go/+/241280
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
There are some architecture-independent rules in #21439, since an
unsigned integer >= 0 is always true and < 0 is always false. This CL
adds these optimizations to generic rules.
Updates #21439
Change-Id: Iec7e3040b761ecb1e60908f764815fdd9bc62495
Reviewed-on: https://go-review.googlesource.com/c/go/+/246617
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
They were missed as part of the refactoring to use a separate
addressing modes pass.
Fixes#40426
Change-Id: Ie0418b2fac4ba1ffe720644ac918f6d728d5e420
Reviewed-on: https://go-review.googlesource.com/c/go/+/244859
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Taking over Zach's CL 212277. Just cleaned up and added a test.
For a positive, signed integer, an arithmetic right shift of count
(bit-width - 1) equals zero. e.g. int64(22) >> 63 -> 0. This CL makes
prove replace these right shifts with a zero-valued constant.
These shifts may arise in source code explicitly, but can also be
created by the generic rewrite of signed division by a power of 2.
// Signed divide by power of 2.
// n / c = n >> log(c) if n >= 0
// = (n+c-1) >> log(c) if n < 0
// We conditionally add c-1 by adding n>>63>>(64-log(c))
(first shift signed, second shift unsigned).
(Div64 <t> n (Const64 [c])) && isPowerOfTwo(c) ->
(Rsh64x64
(Add64 <t> n (Rsh64Ux64 <t>
(Rsh64x64 <t> n (Const64 <typ.UInt64> [63]))
(Const64 <typ.UInt64> [64-log2(c)])))
(Const64 <typ.UInt64> [log2(c)]))
If n is known to be positive, this rewrite includes an extra Add and 2
extra Rsh. This CL will allow prove to replace one of the extra Rsh with
a 0. That replacement then allows lateopt to remove all the unneccesary
fixups from the generic rewrite.
There is a rewrite rule to handle this case directly:
(Div64 n (Const64 [c])) && isNonNegative(n) && isPowerOfTwo(c) ->
(Rsh64Ux64 n (Const64 <typ.UInt64> [log2(c)]))
But this implementation of isNonNegative really only handles constants
and a few special operations like len/cap. The division could be
handled if the factsTable version of isNonNegative were available.
Unfortunately, the first opt pass happens before prove even has a
chance to deduce the numerator is non-negative, so the generic rewrite
has already fired and created the extra Ops discussed above.
Fixes#36159
By Printf count, this zeroes 137 right shifts when building std and cmd.
Change-Id: Iab486910ac9d7cfb86ace2835456002732b384a2
Reviewed-on: https://go-review.googlesource.com/c/go/+/232857
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
match:
m = make([]T, x); copy(m, s)
for pointer free T and x==len(s) rewrite to:
m = mallocgc(x*elemsize(T), nil, false); memmove(&m, &s, x*elemsize(T))
otherwise rewrite to:
m = makeslicecopy([]T, x, s)
This avoids memclear and shading of pointers in the newly created slice
before the copy.
With this CL "s" is only be allowed to bev a variable and not a more
complex expression. This restriction could be lifted in future versions
of this optimization when it can be proven that "s" is not referencing "m".
Triggers 450 times during make.bash..
Reduces go binary size by ~8 kbyte.
name old time/op new time/op delta
MakeSliceCopy/mallocmove/Byte 71.1ns ± 1% 65.8ns ± 0% -7.49% (p=0.000 n=10+9)
MakeSliceCopy/mallocmove/Int 71.2ns ± 1% 66.0ns ± 0% -7.27% (p=0.000 n=10+8)
MakeSliceCopy/mallocmove/Ptr 104ns ± 4% 99ns ± 1% -5.13% (p=0.000 n=10+10)
MakeSliceCopy/makecopy/Byte 70.3ns ± 0% 68.0ns ± 0% -3.22% (p=0.000 n=10+9)
MakeSliceCopy/makecopy/Int 70.3ns ± 0% 68.5ns ± 1% -2.59% (p=0.000 n=9+10)
MakeSliceCopy/makecopy/Ptr 102ns ± 0% 99ns ± 1% -2.97% (p=0.000 n=9+9)
MakeSliceCopy/nilappend/Byte 75.4ns ± 0% 74.9ns ± 2% -0.63% (p=0.015 n=9+9)
MakeSliceCopy/nilappend/Int 75.6ns ± 0% 76.4ns ± 3% ~ (p=0.245 n=9+10)
MakeSliceCopy/nilappend/Ptr 107ns ± 0% 108ns ± 1% +0.93% (p=0.005 n=9+10)
Fixes#26252
Change-Id: Iec553dd1fef6ded16197216a472351c8799a8e71
Reviewed-on: https://go-review.googlesource.com/c/go/+/146719
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Martin Möhrmann <moehrmann@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
We set up static symbols during walk that
we later make copies of to initialize local variables.
It is difficult to ascertain at that time exactly
when copying a symbol is profitable vs locally
initializing an autotmp.
During SSA, we are much better placed to optimize.
This change recognizes when we are copying from a
global readonly all-zero symbol and replaces it with
direct zeroing.
This often allows the all-zero symbol to be
deadcode eliminated at link time.
This is not ideal--it makes for large object files,
and longer link times--but it is the cleanest fix I could find.
This makes the final binary for the program in #38554
shrink from >500mb to ~2.2mb.
It also shrinks the standard binaries:
file before after Δ %
addr2line 4412496 4404304 -8192 -0.186%
buildid 2893816 2889720 -4096 -0.142%
cgo 4841048 4832856 -8192 -0.169%
compile 19926480 19922432 -4048 -0.020%
cover 5281816 5277720 -4096 -0.078%
link 6734648 6730552 -4096 -0.061%
nm 4366240 4358048 -8192 -0.188%
objdump 4755968 4747776 -8192 -0.172%
pprof 14653060 14612100 -40960 -0.280%
trace 11805940 11777268 -28672 -0.243%
vet 7185560 7181416 -4144 -0.058%
total 113588440 113465560 -122880 -0.108%
And not just by removing unnecessary symbols;
the program text shrinks a bit as well.
Fixes#38554
Change-Id: I8381ae6084ae145a5e0cd9410c451e52c0dc51c8
Reviewed-on: https://go-review.googlesource.com/c/go/+/229704
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
Triggers a handful of times in std+cmd.
Change-Id: I9bb8ce9a5f8bae2547cb61157cd8f256e1b63e76
Reviewed-on: https://go-review.googlesource.com/c/go/+/229602
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
