[llvm] [AMDGPU][DAGCombiner][GlobalISel] Prevent FMA contraction when multiply cannot be eliminated (PR #169735)

[Matt Arsenault via llvm-commits]

================
@@ -797,8 +745,9 @@ define float @fsub_muladd_fpext_mul_f16_to_f32_commute(float %x, float %y, float
 ; GFX9-F32FLUSH-LABEL: fsub_muladd_fpext_mul_f16_to_f32_commute:
 ; GFX9-F32FLUSH:       ; %bb.0: ; %entry
 ; GFX9-F32FLUSH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-F32FLUSH-NEXT:    v_mad_mix_f32 v0, -v3, v4, v0 op_sel_hi:[1,1,0]
-; GFX9-F32FLUSH-NEXT:    v_mad_f32 v0, -v1, v2, v0
+; GFX9-F32FLUSH-NEXT:    v_mul_f16_e32 v3, v3, v4
----------------
arsenm wrote:

regression 

https://github.com/llvm/llvm-project/pull/169735