[llvm] [AMDGPU] Vectorize more 16 bit shuffles (PR #90648)

[Matt Arsenault via llvm-commits]

================
@@ -463,8 +463,8 @@ define void @shuffle() {
 ; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8 = shufflevector <8 x i8> undef, <8 x i8> undef, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
 ; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16 = shufflevector <16 x i8> undef, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
 ; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i16_2 = shufflevector <2 x i16> undef, <2 x i16> undef, <2 x i32> <i32 1, i32 0>
-; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %v4i16_4 = shufflevector <4 x i16> undef, <4 x i16> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
-; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %v8i16_8 = shufflevector <8 x i16> undef, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
+; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v4i16_4 = shufflevector <4 x i16> undef, <4 x i16> undef, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
----------------
arsenm wrote:

This blend between different registers will take at least 4 instructions. Right now it looks like it does this in 2 v_perm_b32, but still requires the constant materializes 

https://github.com/llvm/llvm-project/pull/90648