[llvm] [RegisterCoalescer]: Try inflated RC for coalescing reg->subreg (PR #134438)

[Jeffrey Byrnes via llvm-commits]

================
@@ -72,46 +72,94 @@ bb:
 }
 
 define <4 x float> @test_smfmac_f32_16x16x64_f16(<8 x half> %arg0, <16 x half> %arg1, <4 x float> %arg2, i32 %arg3) {
-; GCN-LABEL: test_smfmac_f32_16x16x64_f16:
-; GCN:       ; %bb.0:
-; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    v_smfmac_f32_16x16x64_f16 v[12:15], v[0:3], v[4:11], v16
-; GCN-NEXT:    s_nop 7
-; GCN-NEXT:    v_mov_b32_e32 v0, v12
-; GCN-NEXT:    v_mov_b32_e32 v1, v13
-; GCN-NEXT:    v_mov_b32_e32 v2, v14
-; GCN-NEXT:    v_mov_b32_e32 v3, v15
-; GCN-NEXT:    s_setpc_b64 s[30:31]
+; SDAG-LABEL: test_smfmac_f32_16x16x64_f16:
----------------
jrbyrnes wrote:

As mentioned in https://github.com/llvm/llvm-project/pull/130870 , splitting the PR up introduces regressions that wouldn't have been observed othterwise. For example, many of the regressions here don't exist in the combined version.

https://github.com/llvm/llvm-project/pull/134438