[go] cmd/compile: order functions by IR node count
Jakub Ciolek (Gerrit)
Jakub Ciolek has uploaded the change for review![]( "Open in Gerrit")
Commit message
cmd/compile: order functions by IR node count
Count nodes during the existing pre-walk traversal and use that
count instead of the number of top-level statements when ordering
functions for compilation.
compilebench Intel Alder Lake 12600k (P cores):
│ /home/jake/26jul/costmodel.master │ /home/jake/26jul/costmodel.new │
│ sec/op │ sec/op vs base │
Template 87.92m ± 1% 88.16m ± 1% ~ (p=0.341 n=20)
Unicode 61.10m ± 1% 61.17m ± 1% ~ (p=0.429 n=20)
GoTypes 544.1m ± 0% 542.9m ± 1% -0.23% (p=0.046 n=20)
Compiler 83.30m ± 1% 81.97m ± 1% -1.60% (p=0.000 n=20)
SSA 5.170 ± 0% 5.061 ± 0% -2.11% (p=0.000 n=20)
Flate 95.04m ± 0% 94.80m ± 0% ~ (p=0.192 n=20)
GoParser 96.25m ± 0% 84.29m ± 1% -12.43% (p=0.000 n=20)
Reflect 229.2m ± 0% 227.6m ± 1% -0.69% (p=0.000 n=20)
Tar 100.23m ± 1% 99.12m ± 0% -1.11% (p=0.000 n=20)
XML 114.7m ± 1% 113.3m ± 0% -1.29% (p=0.000 n=20)
LinkCompiler 286.8m ± 1% 285.6m ± 1% ~ (p=0.820 n=20)
ExternalLinkCompiler 760.9m ± 0% 760.8m ± 0% ~ (p=0.841 n=20)
LinkWithoutDebugCompiler 182.8m ± 1% 179.5m ± 1% -1.83% (p=0.000 n=20)
StdCmd 36.82 ± 0% 36.60 ± 0% -0.61% (p=0.001 n=20)
geomean 298.5m 293.6m -1.63%
│ /home/jake/26jul/costmodel.master │ /home/jake/26jul/costmodel.new │
│ user-sec/op │ user-sec/op vs base │
Template 225.9m ± 1% 225.2m ± 2% ~ (p=0.846 n=20)
Unicode 67.70m ± 4% 67.59m ± 4% ~ (p=0.841 n=20)
GoTypes 1.791 ± 1% 1.775 ± 1% -0.87% (p=0.038 n=20)
Compiler 164.4m ± 2% 163.8m ± 2% ~ (p=0.142 n=20)
SSA 17.56 ± 1% 17.63 ± 1% ~ (p=0.989 n=20)
Flate 269.2m ± 2% 270.1m ± 1% ~ (p=0.738 n=20)
GoParser 232.4m ± 2% 232.0m ± 2% ~ (p=0.779 n=20)
Reflect 644.1m ± 1% 642.5m ± 0% ~ (p=0.547 n=20)
Tar 267.3m ± 1% 261.6m ± 1% -2.13% (p=0.000 n=20)
XML 313.7m ± 1% 304.0m ± 1% -3.11% (p=0.000 n=20)
LinkCompiler 462.1m ± 1% 468.6m ± 1% +1.41% (p=0.013 n=20)
ExternalLinkCompiler 848.9m ± 1% 846.6m ± 1% ~ (p=0.779 n=20)
LinkWithoutDebugCompiler 213.1m ± 2% 213.3m ± 2% ~ (p=0.429 n=20)
geomean 438.7m 436.9m -0.43%
Change diff
diff --git a/src/cmd/compile/internal/gc/compile.go b/src/cmd/compile/internal/gc/compile.go
index 3d8974e..e1867d8 100644
--- a/src/cmd/compile/internal/gc/compile.go
+++ b/src/cmd/compile/internal/gc/compile.go
@@ -148,7 +148,7 @@
// Since we remove from the end of the slice queue,
// that means shortest to longest.
slices.SortFunc(compilequeue, func(a, b *ir.Func) int {
- return cmp.Compare(len(a.Body), len(b.Body))
+ return cmp.Compare(a.NumPreWalkNodes, b.NumPreWalkNodes)
})
}
diff --git a/src/cmd/compile/internal/ir/func.go b/src/cmd/compile/internal/ir/func.go
index e387927..c6675e2 100644
--- a/src/cmd/compile/internal/ir/func.go
+++ b/src/cmd/compile/internal/ir/func.go
@@ -55,6 +55,11 @@
// if you add or remove a field, don't forget to update sizeof_test.go
miniNode
+
+ // NumPreWalkNodes is the number of IR nodes before Walk.
+ // It is used as an estimate of backend compilation cost.
+ NumPreWalkNodes int32
+
Body Nodes
Nname *Name // ONAME node
diff --git a/src/cmd/compile/internal/ir/sizeof_test.go b/src/cmd/compile/internal/ir/sizeof_test.go
index b805155..047de93 100644
--- a/src/cmd/compile/internal/ir/sizeof_test.go
+++ b/src/cmd/compile/internal/ir/sizeof_test.go
@@ -20,7 +20,7 @@
_32bit uintptr // size on 32bit platforms
_64bit uintptr // size on 64bit platforms
}{
- {Func{}, 184, 312},
+ {Func{}, 188, 312},
{Name{}, 96, 160},
{miniExpr{}, 32, 48},
{miniNode{}, 12, 12},
diff --git a/src/cmd/compile/internal/walk/walk.go b/src/cmd/compile/internal/walk/walk.go
index 68e8545..ac81661 100644
--- a/src/cmd/compile/internal/walk/walk.go
+++ b/src/cmd/compile/internal/walk/walk.go
@@ -471,7 +471,9 @@
ro.Init(fn)
sv := make(map[ir.Node]ir.Node)
scs := make(map[*ir.Name]*types.Type)
+ var numNodes int32
ir.Visit(fn, func(n ir.Node) {
+ numNodes++
switch n.Op() {
case ir.OCONVIFACE:
x := n.(*ir.ConvExpr).X
@@ -496,4 +498,5 @@
})
staticValues = sv
shapeConvSources = scs
+ fn.NumPreWalkNodes = numNodes
}
Change information
- M src/cmd/compile/internal/gc/compile.go
- M src/cmd/compile/internal/ir/func.go
- M src/cmd/compile/internal/ir/sizeof_test.go
- M src/cmd/compile/internal/walk/walk.go
Related details
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Jakub Ciolek (Gerrit)
Jakub Ciolek voted Commit-Queue+1![]( "Open in Gerrit")
| Commit-Queue | +1 |
Related details
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob voted and added 1 comment![]( "Open in Gerrit")
Votes added by Michael Matloob
| Code-Review | +2 |
1 comment
Ooh thanks! This looks good! I'll try running benchmarks on some other machines.
Related details
- Jakub Ciolek
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Carlos Amedee (Gerrit)
Carlos Amedee voted Code-Review+1![]( "Open in Gerrit")
| Code-Review | +1 |
Related details
- Jakub Ciolek
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob voted Code-Review+1![]( "Open in Gerrit")
| Code-Review | +1 |
Related details
- Jakub Ciolek
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob added 1 comment![]( "Open in Gerrit")
There's not much of a difference on the amd64 GoBuild sweet benchmarks, but the c4 72 core builder does show a noticeable improvement.
Jakub Ciolek (Gerrit)
Jakub Ciolek added 1 comment![]( "Open in Gerrit")
There's not much of a difference on the amd64 GoBuild sweet benchmarks, but the c4 72 core builder does show a noticeable improvement.
Thanks for doing the benches, how much of a difference was it on the 72 core one?
Related details
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob added 1 comment![]( "Open in Gerrit")
Jakub CiolekThere's not much of a difference on the amd64 GoBuild sweet benchmarks, but the c4 72 core builder does show a noticeable improvement.
Thanks for doing the benches, how much of a difference was it on the 72 core one?
Sorry for the delay: here are the ARM 72 core builder numbers:
```
│ baseline │ experiment │
│ sec/op │ sec/op vs base │
GoBuildKubelet-72 17.48 ± 1% 17.44 ± 1% -0.23% (p=0.023 n=10)
GoBuildKubeletLink-72 5.710 ± 0% 5.712 ± 0% ~ (p=0.684 n=10)
GoBuildIstioctl-72 18.83 ± 0% 18.69 ± 0% -0.75% (p=0.001 n=10)
GoBuildIstioctlLink-72 5.917 ± 1% 5.909 ± 0% ~ (p=0.393 n=10)
GoBuildFrontend-72 6.145 ± 1% 6.130 ± 1% ~ (p=0.353 n=10)
GoBuildFrontendLink-72 885.0m ± 2% 870.5m ± 1% -1.64% (p=0.015 n=10)
GoBuildTsgo-72 24.20 ± 1% 24.19 ± 1% ~ (p=1.000 n=10)
GoBuildTsgoLink-72 471.0m ± 1% 470.4m ± 1% ~ (p=0.481 n=10)
```
(I ran the builders as trybots so you can go look at their numbers by going to the "Checks" tab, clicking on one of the perf_vs_parent builders, and then clicking on "benchmark results". The relevant numbers are the "GoBuild" numbers around the middle of the page. Note that the machines are named based on the google cloud machine types, so the c2s16 and c3h88 builders are c2 Intel amd64 machines, the c4dh96 is an AMD amd64 machine and the c4ah72 and c4as16 are Google Axion arm64 machines)
Related details
- Jakub Ciolek
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob submitted the change![]( "Open in Gerrit")
Change information
- M src/cmd/compile/internal/gc/compile.go
- M src/cmd/compile/internal/ir/func.go
- M src/cmd/compile/internal/ir/sizeof_test.go
- M src/cmd/compile/internal/walk/walk.go
Code-Review: +1 by Carlos Amedee, +1 by Michael Matloob, +2 by Michael Matloob
TryBots-Pass: LUCI-TryBot-Result+1 by golang...@luci-project-accounts.iam.gserviceaccount.com
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Jakub Ciolek (Gerrit)
Jakub Ciolek added 1 comment![]( "Open in Gerrit")
Jakub CiolekThere's not much of a difference on the amd64 GoBuild sweet benchmarks, but the c4 72 core builder does show a noticeable improvement.
Michael MatloobThanks for doing the benches, how much of a difference was it on the 72 core one?
Sorry for the delay: here are the ARM 72 core builder numbers:
```
│ baseline │ experiment │
│ sec/op │ sec/op vs base │
GoBuildKubelet-72 17.48 ± 1% 17.44 ± 1% -0.23% (p=0.023 n=10)
GoBuildKubeletLink-72 5.710 ± 0% 5.712 ± 0% ~ (p=0.684 n=10)
GoBuildIstioctl-72 18.83 ± 0% 18.69 ± 0% -0.75% (p=0.001 n=10)
GoBuildIstioctlLink-72 5.917 ± 1% 5.909 ± 0% ~ (p=0.393 n=10)
GoBuildFrontend-72 6.145 ± 1% 6.130 ± 1% ~ (p=0.353 n=10)
GoBuildFrontendLink-72 885.0m ± 2% 870.5m ± 1% -1.64% (p=0.015 n=10)
GoBuildTsgo-72 24.20 ± 1% 24.19 ± 1% ~ (p=1.000 n=10)
GoBuildTsgoLink-72 471.0m ± 1% 470.4m ± 1% ~ (p=0.481 n=10)
```(I ran the builders as trybots so you can go look at their numbers by going to the "Checks" tab, clicking on one of the perf_vs_parent builders, and then clicking on "benchmark results". The relevant numbers are the "GoBuild" numbers around the middle of the page. Note that the machines are named based on the google cloud machine types, so the c2s16 and c3h88 builders are c2 Intel amd64 machines, the c4dh96 is an AMD amd64 machine and the c4ah72 and c4as16 are Google Axion arm64 machines)
Ohh, cool. Thanks Michael! Do you know if I'm able to run those with a try-bot permission or does one need higher privileges? It would be nice to run those for codegen changes in the future.
Related details
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Michael Matloob (Gerrit)
Michael Matloob added 1 comment![]( "Open in Gerrit")
Jakub CiolekThere's not much of a difference on the amd64 GoBuild sweet benchmarks, but the c4 72 core builder does show a noticeable improvement.
Michael MatloobThanks for doing the benches, how much of a difference was it on the 72 core one?
Jakub CiolekSorry for the delay: here are the ARM 72 core builder numbers:
```
│ baseline │ experiment │
│ sec/op │ sec/op vs base │
GoBuildKubelet-72 17.48 ± 1% 17.44 ± 1% -0.23% (p=0.023 n=10)
GoBuildKubeletLink-72 5.710 ± 0% 5.712 ± 0% ~ (p=0.684 n=10)
GoBuildIstioctl-72 18.83 ± 0% 18.69 ± 0% -0.75% (p=0.001 n=10)
GoBuildIstioctlLink-72 5.917 ± 1% 5.909 ± 0% ~ (p=0.393 n=10)
GoBuildFrontend-72 6.145 ± 1% 6.130 ± 1% ~ (p=0.353 n=10)
GoBuildFrontendLink-72 885.0m ± 2% 870.5m ± 1% -1.64% (p=0.015 n=10)
GoBuildTsgo-72 24.20 ± 1% 24.19 ± 1% ~ (p=1.000 n=10)
GoBuildTsgoLink-72 471.0m ± 1% 470.4m ± 1% ~ (p=0.481 n=10)
```(I ran the builders as trybots so you can go look at their numbers by going to the "Checks" tab, clicking on one of the perf_vs_parent builders, and then clicking on "benchmark results". The relevant numbers are the "GoBuild" numbers around the middle of the page. Note that the machines are named based on the google cloud machine types, so the c2s16 and c3h88 builders are c2 Intel amd64 machines, the c4dh96 is an AMD amd64 machine and the c4ah72 and c4as16 are Google Axion arm64 machines)
Ohh, cool. Thanks Michael! Do you know if I'm able to run those with a try-bot permission or does one need higher privileges? It would be nice to run those for codegen changes in the future.
Yeah that should be enough!
Related details
Code-Review
No-Unresolved-Comments
Review-Enforcement
TryBots-Pass
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |