[llvm] [AMDGPU] Update patterns for v_cvt_flr and v_cvt_rpi (PR #177962)

Mon Jan 26 07:16:33 PST 2026

================
@@ -806,17 +809,20 @@ class DwordAddrPat<ValueType vt, RegisterClass rc> : AMDGPUPat <
 
 // Special conversion patterns
 
-def cvt_rpi_i32_f32 : PatFrag <
-  (ops node:$src),
-  (fp_to_sint (ffloor (fadd $src, FP_HALF))),
-  [{ (void) N; return TM.Options.NoNaNsFPMath; }]
->;
+class NNanOpPat<dag frag> : PatFrag<(ops node:$src), frag,
+  [{ return N->getOperand(0).getNode()->getFlags().hasNoNaNs() ||
----------------
mbrkusanin wrote:

Added `ffloor_nnan`. Changed all tests to use `nnan`

https://github.com/llvm/llvm-project/pull/177962
