[CIR][CIRGen][Builtin][X86] lower undef intrinsics

[RiverDave]

We have seem to be generating extra load/store instructions as noted here for all null values encoded in this patch:

CIR:

define dso_local <8 x bfloat> @test_mm_undefined_pbh() #0 {
  %1 = alloca <8 x bfloat>, i64 1, align 16
  %2 = alloca <8 x bfloat>, i64 1, align 16
  store <8 x bfloat> zeroinitializer, ptr %1, align 16
  %3 = load <8 x bfloat>, ptr %1, align 16
  store <8 x bfloat> %3, ptr %2, align 16
  %4 = load <8 x bfloat>, ptr %2, align 16
  ret <8 x bfloat> %4
}

whereas OG:

define dso_local <8 x bfloat> @test_mm_undefined_pbh() #0 {
entry:
  ret <8 x bfloat> zeroinitializer
}

[andykaylor]

This looks reasonable to me.

The behavior you mention in the comments with the extra alloca instructions is an artifact of CIR always generating an alloca for the return value, combined with most optimizations being disabled. The x86 instrinsics are defined as functions with the always_inline attribute set, so unless you pass -disable-llvm-passes on the command-line, the always inline pass will inline the function, but no other optimizations are being run, so the allocas aren't cleaned up. You can see that with a simple case here:

https://godbolt.org/z/KcWdvP3r4

It's a benign behavior, generally speaking. At some point we may want to do something about the return allocas in a pre-lowering transform.