cmd/compile: complicated bounds check elimination #16092
Comments
quentinmit
changed the title
|
/cc @randall77 @dr2chase |
quentinmit
added
the
NeedsDecision
randall77
added
Performance
and removed
NeedsDecision
|
CL https://golang.org/cl/30471 mentions this issue. |
|
I have a fix out for the main bug, doing the immediate load of 1 in the right place. Getting rid of the bounds check is harder. |
Better to just rematerialize them when needed instead of
cross-register spilling or other techniques for keeping them in
registers.
This helps for amd64 code that does 1 << x. It is better to do
loop:
MOVQ $1, AX // materialize arg to SLLQ
SLLQ CX, AX
...
goto loop
than to do
MOVQ $1, AX // materialize outsize of loop
loop:
MOVQ AX, DX // save value that's about to be clobbered
SLLQ CX, AX
MOVQ DX, AX // move it back to the correct register
goto loop
Update #16092
Change-Id: If7ac290208f513061ebb0736e8a79dcb0ba338c0
Reviewed-on: https://go-review.googlesource.com/30471
TryBot-Result: Gobot Gobot <[email protected]>
Run-TryBot: Keith Randall <[email protected]>
Reviewed-by: Cherry Zhang <[email protected]>
quentinmit
added
the
NeedsFix
|
The immediate load seems to be fixed. If you can send a simpler test case for the bounds check, please do so. |
rsc
changed the title
Since version 1.7 is now stable and released, it seems that there is now built-in Bounds Check Elimination (BCE) which does most of what I was talking about as discussed in this Google document and this blog post although there are still many cases where it could eliminate the bounds check and doesn't such as #148058 and #17370 and likely others. With this new BCE capability, I'm sure the loop can be re-written to eliminate bounds checks, or in any case the '-B' compiler option can turn them off; Alternatively, the use of unsafe pointers can also eliminate them so this isn't too much of any issue. Few languages would be smart enough to recognize that the bit index which can be shown to be 32 times the word index in this case will always be less than the array length with the bit index limit tested outside the loop or recognized by the compiler that it is used in the calculation of the slice size when the slice was created.. I downloaded the master and compiled it for x86_64 and the immediate load problem still seems to be there; perhaps the fix has not yet boon committed? Until I can confirm it is gone, I am going to leave this open. golang code for this tight loop still seems to be about 2.5 times faster for some other languages such as Rust, Nim, Haskell, C/C++, etc. on a fast CPU without a cache bottleneck, so I will leave this open until there are some gains; until then array bounds checks are the least of the problems. |
|
I've just checked again that the materialisation of 1 is fixed in Go 1.8. The bound check issue is quite complicated, the compiler would have to prove there are no overflows involved. It sounds unlikely to get there. As per the actual inner loop, it could be faster using the BTS instruction. Basically the hot loop should become: but this is basically a followup to #18943. |
|
The code pasted intp the original post doesn't compile, but the playground link does. I've taken that code, deleted the commented out lines, and converted it into a benchmark. See below. @GordonBGood you might want to double-check that I've preserved the original intent of the benchmark. Using tip at 638ebb0, tip is better than 1.8, which was better than 1.7. It's hard to figure out which of the many things discussed are still relevant, but since the trend is as least correct for 1.9, I'm moving this to 1.10. package main
import (
"math"
"testing"
)
func mkCLUT() [65536]byte {
var arr [65536]byte
for i := 0; i < 65536; i++ {
var cnt byte = 0
for v := (uint16)(i ^ 0xFFFF); v > 0; v &= v - 1 {
cnt++
}
arr[i] = cnt
}
return arr
}
var cnstCLUT [65536]byte = mkCLUT()
func primesTest(top uint) int {
lmtndx := (top - 3) >> 1
lstw := lmtndx >> 5
topsqrtndx := (int(math.Sqrt(float64(top))) - 3) >> 1
cmpsts := make([]uint32, lstw+1)
for i := 0; i <= topsqrtndx; i++ {
if cmpsts[i>>5]&(uint32(1)<<uint(i)) == 0 {
p := (uint(i) << 1) + 3
for j := (p*p - 3) >> 1; j <= lmtndx; j += p {
cmpsts[j>>5] |= 1 << (j & 31)
}
}
}
msk := uint32(0xFFFFFFFE) << (lmtndx & 31)
cmpsts[lstw] |= msk
cnt := 1
for i := uint(0); i <= lstw; i++ {
v := cmpsts[i]
cnt += int(cnstCLUT[v&0xFFFF] + cnstCLUT[0xFFFF&(v>>16)])
}
return cnt
}
var sink int
func BenchmarkPrimesTest(b *testing.B) {
for i := 0; i < b.N; i++ {
sink = primesTest(262146)
}
} |
|
@josharian, yes, the new code tests what I intended to test, with the results calculated by a Look Up Table (LUT) taking a small part of the time to cull the composite numbers, the timing result now reflecting the time taken by the tight culling loop. Also, with the range limited to that required by a 16 Kilobyte CPU L1 buffer size, it is testing raw CPU loop speed and not memory access speed, which testing of tight CPU loop speed is what is desired. |
|
Out of curiosity, I measured where we're at now compared to Go 1.7: (now == tip 1caa062) |
|
This just goes to show that profiling without analysis is of limited use:
What it does seem to show is a 25% improvement for current version for the
particular CPU used.
What it fails to consider is the runtime conditions and the impact on those
results.
The analysis in terms of clock cycles per cull loop is instructive: This
algorithm when compiled efficiently with gcc C/C++, Nim, Rust, etc., takes
about 5.5 clock cycles per cull (for about 200K cull operations) whether
run on a high end modern CPU or a basic one such as the Intel Atom x5-Z8350
in my Windows tablet; thus a run time of about 0.3 second doesn't sound too
bad assuming about a 3.5 GHz clock speed.
However, I ran the benchmark on golang version 1.10.3 (which I don't know
if includes the tweak or not) on my Atom taking about 13.75 clocks per
cull! Even if new code would produce a 25% improvement, that would still
take about 10.5 to 11 clocks per cull.
For comparison, the same algorithm run with F# (which produces very
unoptimized code for this loop) takes about this same number of clocks of
10.5 per cull.
What's the difference between high end CPU's and my Atom other than amount
and sophistication of caches unused by this benchmark (and obviously the
price - grin)? High end processors have more sophisticated branch
prediction, better Out of Order Execution (OOE), and simultaneous
instruction execution than my poor Atom; this can serve to hide some of the
deficiencies of crap code as in this case.
I didn't bother doing the full autopsy on the code as I did in the OP, but
it seems obvious that it is still producing poor quality native code just
as F# does (for which I have also done the autopsy) other than tip has
likely (finally) eliminated a redundant register load I identified in the
OP.
So it's too early to close this issue and I would say you likely have a
long way to go before golang speed approaches the leaders of the pack, in
many use cases by a factor of two or more as for this benchmark.
It also shows the deficiencies of benchmarking/profiling in judging quality
of code given modern high-end CPU's abilities to make the best of the bad.
It also shows that benchmark results should be given in numbers of clock
cycles to eliminate CPU clock speed, but that CPU model and generation can
still skew results of benchmarks as here.
A case could be made that all benchmarks should be verified on a lower end
processor as I did.
…On Tue, Jun 19, 2018, 11:50 Brad Fitzpatrick ***@***.***> wrote:
Out of curiosity, I measured where we're at now compared to Go 1.7: (now
== tip 1caa062
<1caa062>
)
***@***.***:~/src/issue_16092$ benchstat 1.7 tip
name old time/op new time/op delta
PrimesTest-4 399µs ± 1% 301µs ± 0% -24.46% (p=0.000 n=10+8)
***@***.***:~/src/issue_16092$ benchstat 1.8 tip
name old time/op new time/op delta
PrimesTest-4 395µs ± 1% 301µs ± 0% -23.66% (p=0.000 n=9+8)
***@***.***:~/src/issue_16092$ benchstat 1.7 1.8
name old time/op new time/op delta
PrimesTest-4 399µs ± 1% 395µs ± 1% -1.06% (p=0.000 n=10+9)
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#16092 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AKM2TV3uSQQN78yddXVWJWLdWjasnSD9ks5t-Tn3gaJpZM4I4IVh>
.
|
|
I should have clarified what I meant by "crap" native code produced by F#
and golang: once array bounds checks are turned off or elided, the main
restriction is not using read/modify/write instructions that can make the
loop much more efficient as discussed previously.
This is what is wrong with the F# loops, although in addition array bounds
checks can't be turned off or easily elided for F# as they can be/could be
in golang: F# never uses r/m/w as I assume golang still doesn't.
Sophisticated simultaneous execution and OOE can mitigate this on high end
processors just as advanced branch prediction can reduce the cost of bounds
checks.
On Tue, Jul 17, 2018, 05:06 W. Gordon Goodsman <[email protected]>
wrote:
… This just goes to show that profiling without analysis is of limited use:
What it does seem to show is a 25% improvement for current version for the
particular CPU used.
What it fails to consider is the runtime conditions and the impact on
those results.
The analysis in terms of clock cycles per cull loop is instructive: This
algorithm when compiled efficiently with gcc C/C++, Nim, Rust, etc., takes
about 5.5 clock cycles per cull (for about 200K cull operations) whether
run on a high end modern CPU or a basic one such as the Intel Atom x5-Z8350
in my Windows tablet; thus a run time of about 0.3 second doesn't sound too
bad assuming about a 3.5 GHz clock speed.
However, I ran the benchmark on golang version 1.10.3 (which I don't know
if includes the tweak or not) on my Atom taking about 13.75 clocks per
cull! Even if new code would produce a 25% improvement, that would still
take about 10.5 to 11 clocks per cull.
For comparison, the same algorithm run with F# (which produces very
unoptimized code for this loop) takes about this same number of clocks of
10.5 per cull.
What's the difference between high end CPU's and my Atom other than amount
and sophistication of caches unused by this benchmark (and obviously the
price - grin)? High end processors have more sophisticated branch
prediction, better Out of Order Execution (OOE), and simultaneous
instruction execution than my poor Atom; this can serve to hide some of the
deficiencies of crap code as in this case.
I didn't bother doing the full autopsy on the code as I did in the OP, but
it seems obvious that it is still producing poor quality native code just
as F# does (for which I have also done the autopsy) other than tip has
likely (finally) eliminated a redundant register load I identified in the
OP.
So it's too early to close this issue and I would say you likely have a
long way to go before golang speed approaches the leaders of the pack, in
many use cases by a factor of two or more as for this benchmark.
It also shows the deficiencies of benchmarking/profiling in judging
quality of code given modern high-end CPU's abilities to make the best of
the bad.
It also shows that benchmark results should be given in numbers of clock
cycles to eliminate CPU clock speed, but that CPU model and generation can
still skew results of benchmarks as here.
A case could be made that all benchmarks should be verified on a lower end
processor as I did.
On Tue, Jun 19, 2018, 11:50 Brad Fitzpatrick ***@***.***>
wrote:
> Out of curiosity, I measured where we're at now compared to Go 1.7: (now
> == tip 1caa062
> <1caa062>
> )
>
> ***@***.***:~/src/issue_16092$ benchstat 1.7 tip
> name old time/op new time/op delta
> PrimesTest-4 399µs ± 1% 301µs ± 0% -24.46% (p=0.000 n=10+8)
>
> ***@***.***:~/src/issue_16092$ benchstat 1.8 tip
> name old time/op new time/op delta
> PrimesTest-4 395µs ± 1% 301µs ± 0% -23.66% (p=0.000 n=9+8)
>
> ***@***.***:~/src/issue_16092$ benchstat 1.7 1.8
> name old time/op new time/op delta
> PrimesTest-4 399µs ± 1% 395µs ± 1% -1.06% (p=0.000 n=10+9)
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#16092 (comment)>, or mute
> the thread
> <https://github.com/notifications/unsubscribe-auth/AKM2TV3uSQQN78yddXVWJWLdWjasnSD9ks5t-Tn3gaJpZM4I4IVh>
> .
>
|
1. go version go1.7beta1 windows/amd64
2.
set GOARCH=amd64
set GOBIN=
set GOEXE=.exe
set GOHOSTARCH=amd64
set GOHOSTOS=windows
set GOOS=windows
set GOPATH=F:\Go
set GORACE=
set GOROOT=F:\Go
set GOTOOLDIR=F:\Go\pkg\tool\windows_amd64
set CC=gcc
set GOGCCFLAGS=-m64 -mthreads -fmessage-length=0 -fdebug-prefix-map=C:\Users\Super\AppData\Local\Temp\go-build211527254=/tmp/go-build -gno-record-gcc-switches
set CXX=g++
set CGO_ENABLED=1
3. Runnable program:
play.golang.org link: https://play.golang.org/p/_E5R5JAlGW
4. When "go tool compile -S PrimeSpeed.go > PrimeSpeed.s" is run, the inner tight composite number culling loop as quoted below:
looks like the following assembly code from PrimeSpeed.s:
5. I expected to see:
Even better, without recalculating p = 2 * i + 3 thus j += j + 2 * i + 3 inside the inner loop:
Includes changing order of instructions for processors without OOE:
The following is the same loop without bounds checks generated for C/C++ with the Visual Studio compiler (intel assembler format):
Note that the above uses a total of seven registers for this inner loop, so the same code is generated for x86 and x64 compilations. Unfortunately, it takes another register to hold the upper array bound for a range check and the x86 architecture can only have seven available; however, it is possible to slightly change the code as follows:
which for Visual Studio C/C++ generates the following same number of instructions in the loop:
and it can be seen that the array bounds check is now done at the same time as the loop completion check; it should be a simple matter to clue the compiler that 'bnds' contains the array length, perhaps by assigning it inside the loop as len(cmpsts) as is done for C# x86 code so that it recognizes that the bounds check is already done. The start point of the loop could be the line after the "or" line at the $Start: label or an external check could be implemented to ensure that the bounds check is done for the first loop before the array is accessed as is done for the Visual Studio C/C++ compiler.
As demonstrated above, the golang code runs slower than C/C++ code by almost a factor of two on some x86 processors and more than that factor for x86 processors. It also runs slightly slower than C#/Java for both architectures.
The text was updated successfully, but these errors were encountered: