speed of encoding

[rgrig]

The following are times of running parts of topl on dacapo:

• 7m30s for zip+decode+encode+topl instrumentation
• 6m33s for zip+decode+encode
• 1m41s for zip+decode+inherit
• 1m35s for zip+decode
• 13s for zip

Encoding is more than twice slower than decoding. There's no good reason for this difference; should be fixed.

[yoff]

Does encoding not do static analysis?

[rgrig]

I don't expect that part to be slow on average, unless the algorithm isn't implemented properly.

[rgrig]

Hmm ... You are right: Changing unify to always return Top brings down the runtime from 7m30s to 5m12s. So, the encoding is about 4m58s, out of which about 2m18s are spent spinning in symbolic execution. Meanwhile, decoding is about 1m22s. I still think encoding should be faster, but symbolic execution takes more time than I expected.

Here are the estimates again, for ease of reading later:

• decoding 1m20s
• encoding 5m, out of which
  • 2m20s searching for fixed-point in type inference
  • 2m40s other stuff

[yoff]

Wow, so even the other stuff is twice the time of decoding...does that include writing the files? I share your feeling that the symbolic execution should be faster as well, but I am not really sure why...

[rgrig]

Decoding includes reading the files; encoding includes writing the files.

[yoff]

Ok, then the "other stuff" is definitely too slow, I wonder if it is worth profiling...

[rgrig]

It is worth profiling, but not this week. :)