Description
When using the g++ compiler on the windows-latest GitHub Actions runner to compile a simple C++ program, the first compilation time is highly unstable (4s to 60s). The second and third compilations in the same job are consistently under 1 second.
Platforms affected
Runner images affected
Image version and build link
I will provide more infomations below.
Is it regression?
idk
Expected behavior
I hope that windows-latest runner's g++ first compile should be fast.
Actual behavior
It is very slow and unstable.
Repro steps
The C++ source code compiled is as follows:
#include <cstring>
#include <iostream>
#include <vector>
using namespace std;
int f(int n) {
vector<int> a(1);
for (int i = 0; i < n; i++) {
a[i] = 100;
}
}
int *p;
int main() {
int s;
cin >> s;
switch (s) {
case 1:
cout << "AC" << endl;
break;
case 2:
cout << "WA" << endl;
break;
case 3:
for (int i = 0; i < 200; i++)
p = new int[1 << 24], memset(p, 0x3f, sizeof(int) * (1 << 24));
break;
case 4:
while (1)
;
break;
case 5:
f(100);
break;
default:
break;
}
}It should be noted that although the code contains a function with no return statement, we only compile it, not execute it.
The CI workflow (including the C++ source code) can be found here:
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/c3b734815e56a5115b714fba15d10c3063797ab4/.github/workflows/test.yml
Example run output:
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25832441342/job/75900199272
In that run, compilation took 11 seconds. Attached below are a few more extreme cases.
Next I tried adding -ftime-report (https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/dd5f131ea30de0438f86ebe3d387c415fdc5c7c0/.github/workflows/test.yml) and obtained:
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25833002457/job/75902175044
In this run, we reached an extreme compilation time of 60 seconds (1st: 61557.9172 ms). The log file includes timestamps for each command. From these timestamps we observed:
cc1plus.exe started, but its first version information was printed only 18.5 seconds later. This 18.5 seconds is not accounted for in -ftime-report.- The
TOTAL wall time reported by -ftime-report was 30.19 seconds, broken down as:
- phase setup: 5.36 s
- phase parsing: 11.88 s
- phase opt and generate: 12.86 s
- preprocessing: 8.06 s
- callgraph functions expansion: 8.60 s
- Additionally:
- as.exe: 0.54 s
- collect2.exe: 7.33 s
We have tried disabling Windows Defender (although according to the official repo it should already be disabled by default). This did not improve the first compilation speed at all. The second link below provides evidence that Defender is indeed disabled by default.
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/2bf29752f3df3afa16691f1112512a7505d6a1d9/.github/workflows/test.yml
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25833408494/job/75903332461
Switching to the windows-2025-vs2026 runner, the first compilation time stabilised at 5–8 seconds.
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/cd2ee615f3ac08645f909b812220cf9316eb2976/.github/workflows/test.yml
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25832535087/job/75900585355
It should be noted that I cannot be certain whether the issue is truly resolved on windows-2025-vs2026 (I can only judge by running multiple times to see if the time fluctuates significantly).
Summary: The first-time g++ compilation performance on windows-latest is very poor and unstable, and disabling Windows Defender does not help.
Description
When using the g++ compiler on the
windows-latestGitHub Actions runner to compile a simple C++ program, the first compilation time is highly unstable (4s to 60s). The second and third compilations in the same job are consistently under 1 second.Platforms affected
Runner images affected
Image version and build link
I will provide more infomations below.
Is it regression?
idk
Expected behavior
I hope that
windows-latestrunner's g++ first compile should be fast.Actual behavior
It is very slow and unstable.
Repro steps
The C++ source code compiled is as follows:
It should be noted that although the code contains a function with no return statement, we only compile it, not execute it.
The CI workflow (including the C++ source code) can be found here:
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/c3b734815e56a5115b714fba15d10c3063797ab4/.github/workflows/test.yml
Example run output:
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25832441342/job/75900199272
In that run, compilation took 11 seconds. Attached below are a few more extreme cases.
Next I tried adding
-ftime-report(https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/dd5f131ea30de0438f86ebe3d387c415fdc5c7c0/.github/workflows/test.yml) and obtained:https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25833002457/job/75902175044
In this run, we reached an extreme compilation time of 60 seconds (
1st: 61557.9172 ms). The log file includes timestamps for each command. From these timestamps we observed:cc1plus.exestarted, but its first version information was printed only 18.5 seconds later. This 18.5 seconds is not accounted for in-ftime-report.TOTAL walltime reported by-ftime-reportwas 30.19 seconds, broken down as:We have tried disabling Windows Defender (although according to the official repo it should already be disabled by default). This did not improve the first compilation speed at all. The second link below provides evidence that Defender is indeed disabled by default.
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/2bf29752f3df3afa16691f1112512a7505d6a1d9/.github/workflows/test.yml
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25833408494/job/75903332461
Switching to the
windows-2025-vs2026runner, the first compilation time stabilised at 5–8 seconds.https://github.com/ZnPdCo/windows-ci-gcc-compile-test/blob/cd2ee615f3ac08645f909b812220cf9316eb2976/.github/workflows/test.yml
https://github.com/ZnPdCo/windows-ci-gcc-compile-test/actions/runs/25832535087/job/75900585355
It should be noted that I cannot be certain whether the issue is truly resolved on
windows-2025-vs2026(I can only judge by running multiple times to see if the time fluctuates significantly).Summary: The first-time
g++compilation performance onwindows-latestis very poor and unstable, and disabling Windows Defender does not help.