JIT: Avoid cloning cold descendant loops quadratically #96580
Conversation
When we have a loop nest where we clone multiple ancestor loops we would clone descendant loops, ending up with a total quadratic number of cloned loops in terms of the depth of the loop nest. Instead of producing these multiple versions of cold loops (each of which is optimized for a particular hot ancestor loop) we can reuse the first cold loop we cloned. This has the effect of exiting all the "hot" versions of ancestor loops if the cloning conditions fail, but it gets rid of the quadratic loop cloning problem.
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsWhen we have a loop nest where we clone multiple ancestor loops we would clone descendant loops, ending up with a total quadratic number of cloned loops in terms of the depth of the loop nest. Instead of producing these multiple versions of cold loops (each of which is optimized for a particular hot ancestor loop) we can reuse the first cold loop we cloned. This has the effect of exiting all the "hot" versions of ancestor loops if the cloning conditions fail, but it gets rid of the quadratic loop cloning problem. Size improvements expected, with a few regressions where RBO was able to fully optimize away some cold loop nests before, and now cannot because they are entered in new, strange ways. Example: [MethodImpl(MethodImplOptions.NoInlining)]
private static long Foo(int[] a, int[] b, int[] c, int[] d, int[] e, int n)
{
long sum = 0;
for (int i = 0; i < n; i++)
{
sum += a[i];
for (int j = 0; j < n; j++)
{
sum += b[j];
for (int k = 0; k < n; k++)
{
sum += c[k];
for (int l = 0; l < n; l++)
{
sum += d[l];
for (int m = 0; m < n; m++)
{
sum += e[m];
}
}
}
}
}
return sum;
}Flow graph before; a total of 20 loops after cloning has run. -; Total bytes of code 756, prolog size 27, PerfScore 11557.63, instruction count 234, allocated bytes for code 756 (MethodHash=9f186a76) for method Program:Foo(int[],int[],int[],int[],int[],int):long (FullOpts)
+; Total bytes of code 423, prolog size 27, PerfScore 11001.65, instruction count 134, allocated bytes for code 423 (MethodHash=9f186a76) for method Program:Foo(int[],int[],int[],int[],int[],int):long (FullOpts)
|
ghost
closed this
|
Draft Pull Request was automatically closed for 30 days of inactivity. Please let us know if you'd like to reopen it. |
When we have a loop nest where we clone multiple ancestor loops we would clone descendant loops every time, ending up with a total quadratic number of cloned loops in terms of the depth of the loop nest.
Instead of producing these multiple versions of cold loops (each of which is optimized for a particular hot ancestor loop) we can reuse the first cold loop we cloned. This has the effect of exiting all the "hot" versions of ancestor loops if the cloning conditions of a descendant loop fails, but it gets rid of the quadratic loop cloning problem.
Size improvements expected, with a few regressions where RBO was able to fully optimize away some cold loop nests before, and now cannot because they are entered in new, strange ways.
Example:
Flow graph before; a total of 20 loops after cloning has run.
Flow graph after; a total of 5 natural loops after cloning has run, and 5 more unnatural ones.
Assembly diff:
(Easiest to review with whitespace disabled.)
Addresses part of #8558