[llvm] DAG: Use fast variants of fast math libcalls (PR #147481)

[Brian Cain via llvm-commits]

================
@@ -1411,27 +1422,26 @@ def __hexagon_moddi3 : RuntimeLibcallImpl<SREM_I64>;
 def __hexagon_umodsi3 : RuntimeLibcallImpl<UREM_I32>;
 def __hexagon_umoddi3 : RuntimeLibcallImpl<UREM_I64>;
 
-// FIXME: "Fast" versions should be treated as a separate RTLIB::FAST_* function
 def __hexagon_adddf3 : RuntimeLibcallImpl<ADD_F64>;
-def __hexagon_fast_adddf3 : RuntimeLibcallImpl<ADD_F64>;
+def __hexagon_fast_adddf3 : RuntimeLibcallImpl<FAST_ADD_F64>;
 
 def __hexagon_subdf3 : RuntimeLibcallImpl<SUB_F64>;
-def __hexagon_fast_subdf3 : RuntimeLibcallImpl<SUB_F64>;
+def __hexagon_fast_subdf3 : RuntimeLibcallImpl<FAST_SUB_F64>;
 
 def __hexagon_muldf3 : RuntimeLibcallImpl<MUL_F64>;
-def __hexagon_fast_muldf3 : RuntimeLibcallImpl<MUL_F64>;
+def __hexagon_fast_muldf3 : RuntimeLibcallImpl<FAST_MUL_F64>;
 
 def __hexagon_divdf3 : RuntimeLibcallImpl<DIV_F64>;
-def __hexagon_fast_divdf3 : RuntimeLibcallImpl<DIV_F64>;
+def __hexagon_fast_divdf3 : RuntimeLibcallImpl<FAST_DIV_F64>;
 
 def __hexagon_divsf3 : RuntimeLibcallImpl<DIV_F32>;
-def __hexagon_fast_divsf3 : RuntimeLibcallImpl<DIV_F32>;
+def __hexagon_fast_divsf3 : RuntimeLibcallImpl<FAST_DIV_F32>;
 
 def __hexagon_sqrtf : RuntimeLibcallImpl<SQRT_F32>;
-def __hexagon_fast2_sqrtf : RuntimeLibcallImpl<SQRT_F32>;
+def __hexagon_fast2_sqrtf : RuntimeLibcallImpl<FAST_SQRT_F32>;
 
 // This is the only fast library function for sqrtd.
-def __hexagon_fast2_sqrtdf2 : RuntimeLibcallImpl<SQRT_F64>;
----------------
androm3da wrote:

IIRC these functions in the library are aliased to the same implementations.

But this change is still acceptable IMO, the library could conceivably distinguish between them at some point.

@aankit-ca can you run any more comprehensive testing on this change, that would give us more confidence about it?

https://github.com/llvm/llvm-project/pull/147481