[llvm] [AMDGPU] Legalize 64bit elements for BUILD_VECTOR on gfx942 (PR #145052)

[Janek van Oirschot via llvm-commits]

================
@@ -2435,39 +2435,40 @@ define amdgpu_kernel void @test_mfma_nested_loop_zeroinit(ptr addrspace(1) %arg)
 ;
 ; GFX942-LABEL: test_mfma_nested_loop_zeroinit:
----------------
JanekvO wrote:

Materializes `0` as b64 which is done through VGPR with v_mov_b64, results in regression of +1 instruction instead of materializing into a0 and re-using that to set the remainder.

https://github.com/llvm/llvm-project/pull/145052