[ValueTypes] Automatically assign value numbers (NFC)

[AlexMaclean]

The requirement to specify an index in each ValueType declaration adds a lot of visual clutter to the file and causes the addition of new value types to create somewhat painful chrun, especially if they are in a downstream branch which can create lots of merge conflicts.

[github-actions]

🐧 Linux x64 Test Results

• 186453 tests passed
• 4868 tests skipped

[s-barannikov]

Suggested change

	class ValueType<int size, int value = -1> {
	class ValueType<int size, int value = ?> {

and use !initialized?

[s-barannikov]

Can you confirm that the generated file is the same?
Also, some targets use PtrValueType to define their own types; can this cause issues?

[AlexMaclean]

Can you confirm that the generated file is the same?

The file is largely the same. The only difference is that there used to be a large gap between c128 (the last normal value type) and cPTR (the first non-normal value type). This gap is no longer present after my change. I don't see any reason why we'd want this though.

def c64 : VTCheriCapability<64, 254>;   // 64-bit CHERI capability value
def c128 : VTCheriCapability<128, 255>; // 128-bit CHERI capability value
...
def cPTR : VTAny<503>;

Also, some targets use PtrValueType to define their own types; can this cause issues?

PtrValueType will still have the same value as the underlying scalar type it represents. These should be totally unaffected by this change.

[s-barannikov]

LGTM modulo nit, thanks

[arsenm]

This looks algorithmically terrible. Did you benchmark how long tablegen takes to generate this with and without this patch?

[arsenm]

plus it may influence all of the other backends that include ValueTypes.td

[AlexMaclean]

Yep. This is pretty inefficient, but there doesn't seem to be a better way to implement this in table-gen.

I tested like this:

    git checkout main 
    echo "main"
    time ../cmake/bin/llvm-tblgen llvm/include/llvm/CodeGen/ValueTypes.td &> /dev/null

    git checkout dev/amaclean/td-vts-indices
    echo "dev/amaclean/td-vts-indices"
    time ../cmake/bin/llvm-tblgen llvm/include/llvm/CodeGen/ValueTypes.td &> /dev/null

Here are the results:

Switched to branch 'main'
Your branch is up to date with 'origin/main'.
main

real    0m0.021s
user    0m0.018s
sys     0m0.002s
Switched to branch 'dev/amaclean/td-vts-indices'
dev/amaclean/td-vts-indices

real    0m0.094s
user    0m0.092s
sys     0m0.001s

So, when isolating ValueTypes.td, it makes things nearly 5X slower (which is pretty bad) but we're still only talking about less then 1/10th of a second which seems like it will be pretty negligible on the scale of a complete LLVM build.

[arsenm]

I think @chapuni had a different patch for this at some point

[chapuni]

@arsenm
I haven't had patches for this. I supposed we needed to introduce "autonumber" in tblgen.

I think this is a better workaround. Thanks! @AlexMaclean

[s-barannikov]

Alternative in #169670