Changes to constant handling - faster deduplication, more compact represtntation

[FractalFir]

This PR changes a few things.

First of all, it caches ConstAllocations, ensuring that for a given ConstAllocation(or one an identical one with a different aligement), you will get the exact same Rvalue.

This allows us to simplify the deduplication code in static_addr_of, and remove the formating (which caused memory leaks).

Additonally, this allows constant byte sequences to sometimes be represented in a more compact fashion. From a few checks, it looks like ~80 % of constant byte sequences can be optimized this way.

[antoyo]

Very nice work!

Here's a first review for the ConstAllocation part.

[FractalFir]

As a nice bonus, this PR seems to drastically reduce the ammount of memory used to build regex-automata.

I can now build rustc with 10 threads, and run on ~28 GB of RAM. This is still suboptimal, but a fair bit better than before.

[antoyo]

As a nice bonus, this PR seems to drastically reduce the ammount of memory used to build regex-automata.

I can now build rustc with 10 threads, and run on ~28 GB of RAM. This is still suboptimal, but a fair bit better than before.

Very nice improvement!
Is this with -Copt-level=0 and do you still disable some GCC passes?

[FractalFir]

As a nice bonus, this PR seems to drastically reduce the ammount of memory used to build regex-automata.
I can now build rustc with 10 threads, and run on ~28 GB of RAM. This is still suboptimal, but a fair bit better than before.

Very nice improvement! Is this with -Copt-level=0 and do you still disable some GCC passes?

On -Copt-level=1

[FractalFir]

Should be good to merge now.

[antoyo]

After this, it should be good to merge.
Thanks!

[antoyo]

A few nitpicks:

[antoyo]

After this last nitpick is fixed, I believe I can merge.
Thanks!

[bjorn3]

In your commit message: *representation

[FractalFir]

In your commit message: *representation

Did not see that, thanks :)!

[antoyo]

Thanks a lot for this nice improvement!