JIT: Allow more containment opts in Tier0

[saucecontrol]

This enables embedded broadcast of non-const values in Tier0

Diffs are a net improvement, although there are a few regressions where an extra temp ends up being introduced due to arg swapping.

There are also a few 1- or 2-byte regressions where we swapped from containing a full vector load arg to containing a broadcast arg, which then forces EVEX encoding. It would be interesting to look at optimizing around that (separately -- it would impact FullOpts as well)

[saucecontrol]

cc @tannergooding

[tannergooding]

There are also a few 1- or 2-byte regressions where we swapped from containing a full vector load arg to containing a broadcast arg

We view this as an explicit improvement and the real "issue" is more that SPMI doesn't surface any size savings in the data section size. -- That is, while the codegen is 1-2 bytes bigger, we save 8-60 bytes of data section size and improve cache locality.

[saucecontrol]

We view this as an explicit improvement and the real "issue" is more that SPMI doesn't surface any size savings in the data section size. -- That is, while the codegen is 1-2 bytes bigger, we save 8-60 bytes of data section size and improve cache locality.

The cases I'm referring to are like this:

where it's a broadcast either way, and we can contain either the broadcast or the full vector. It's always 2 instructions because they can't both be contained. Switching from containing the full vector to containing the broadcast means you have to switch to EVEX, so it's a net increase in size.

This particular regression only applies to instructions where we swap operands in order to be able to contain one, so I think we could simply give lower preference to CnsVec operands that might be turned into broadcast. Or something like that?

[tannergooding]

This particular regression only applies to instructions where we swap operands in order to be able to contain one, so I think we could simply give lower preference to CnsVec operands that might be turned into broadcast. Or something like that?

Ah, I see.

Yeah, in general we want to prefer loads from arbitrary memory, then broadcastable constants, then regular constants.

[saucecontrol]

Disabled the aligned load containment. Diffs are smaller but still a net improvement.

[saucecontrol]

I've split the TryFoldCnsVecForEmbeddedBroadcast changes out into to #117700

[tannergooding]

LGTM. CC. @dotnet/jit-contrib for secondary review

[tannergooding]

/ba-g unrelated arm64 timeouts