[llvm] [AMDGPU][True16][CodeGen] build_vector pattern in true16 (PR #118904)

[Brox Chen via llvm-commits]

================
@@ -3424,6 +3441,8 @@ def : GCNPat <
 
 // Take the upper 16 bits from V[0] and the lower 16 bits from V[1]
 // Special case, can use V_ALIGNBIT (always uses encoded literal)
+foreach p = [NotHasTrue16BitInsts, UseFakeTrue16Insts] in
+let True16Predicate = p in
 def : GCNPat <
   (vecTy (DivergentBinFrag<build_vector>
     (Ty !if(!eq(Ty, i16),
----------------
broxigarchen wrote:

Took a closer look at this. I think these vector patch all might be moving to just fake16/non-16 mode. probably get a patch up in downstream first

https://github.com/llvm/llvm-project/pull/118904