[llvm] [SDAG] Match BUILD_VECTOR in INSERT_SUBVECTOR to SPLAT_VECTOR fold (PR #163984)

[Paul Walker via llvm-commits]

================
@@ -8,7 +8,7 @@
 define void @pr35443() {
 ; CHECK-LABEL: pr35443:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vpbroadcastb ac+4(%rip), %xmm0
+; CHECK-NEXT:    vpbroadcastb ac+4(%rip), %ymm0
----------------
paulwalker-arm wrote:

I'm pretty sure we've tried this before with the same outcome, because the combine is removing potentially valuable information. I wonder if instead of removing it you can teach the relevant is_splat/get_splat functions that take allow_undefs like parameters to do a better job of this idiom for scalable vectors?

Same goes for matching splats during isel as most of the relevant patterns should just work by looking-through the insert_subvector.

https://github.com/llvm/llvm-project/pull/163984