[llvm] [AMDGPU] Use FPImmLeaf for float constants in tablegen. NFCI. (PR #178018)

[Matt Arsenault via llvm-commits]

================
@@ -3776,7 +3776,7 @@ def : GCNPat <
 >;
 
 def : GCNPat <
-  (v2f16 (DivergentBinFrag<build_vector> (f16 VGPR_32:$src1), (f16 FP_ZERO))),
+  (v2f16 (DivergentBinFrag<build_vector> (f16 VGPR_32:$src1), (f16 fpimm_zero))),
----------------
arsenm wrote:

These cases actually had and still have a bug. These require the constant to be a positive zero. The matcher is improperly selecting + or - 0 

https://github.com/llvm/llvm-project/pull/178018