[llvm] [RISCV] Default to MicroOpBufferSize = 1 for scheduling purposes (PR #126608)

[Philip Reames via llvm-commits]

preames wrote:

Additional context because the review description was getting rather long...

I believe this fixes the same issue as @BeMg 's https://github.com/llvm/llvm-project/pull/125468.  The root cause discussion on that was useful for confirming my own understanding, thanks!  I also benefited from offline conversation with both @lukel97 and @topperc when thinking about how to approach this.

This isn't a strict improvement.  There are a couple cases in the tests where we actually see increased spills, but a) this are outweighed by the improvements and b) represent cases where -mcpu=something-with-a-model would already see regressions today.  I think these are easily ignorable, but have only skimmed the tests, please point out anything you find particularly concerning.

This parallels what is done in X86, though the code structure differs a bit.  There both the default (which uses the sandybridge model), and GenericX86Model use values for MicroOpBufferSize which imply out of order designs.  I deliberately exclude that from this change semantic diff.  

This also helps significantly when we end up scalarizing a vector math lib call.  Without this change, we end up interweaving the element libc calls and the extracts, which results in some pretty terrible codegen.  With this change, we do explode vector, then element wise scalar calls, then buildvector.  This is a definite benefit for the default and rva23u configurations.  (As above, codegen with -mcpu for a cpu with a cost model doesn't change.)

https://github.com/llvm/llvm-project/pull/126608