[llvm] [ADMGPU] Replace `isInlinableLiteral16` with specific version (PR #84402)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -112,7 +112,7 @@ bb:
 define amdgpu_ps void @test_wmma_bf16_16x16x16_bf16_imm(<8 x i16> %A, <8 x i16> %B, ptr addrspace(1) %out) {
 ; GFX12-LABEL: test_wmma_bf16_16x16x16_bf16_imm:
 ; GFX12:       ; %bb.0: ; %bb
-; GFX12-NEXT:    v_wmma_bf16_16x16x16_bf16 v[10:13], v[0:3], v[4:7], 1.0
+; GFX12-NEXT:    v_wmma_bf16_16x16x16_bf16 v[10:13], v[0:3], v[4:7], 0
----------------
rampitec wrote:

What has happened here again? I have not seen this downstream. I would understand a code bloat with the initialization with 0x3f80, but not outright 0 inline.

Also just as a note pulled from downstream: this shall use BF16 in the profile. As a minimum in the profile, but then also in the intrinsic. But downstream it still maintains 1.0, so something is missing.

https://github.com/llvm/llvm-project/pull/84402