[llvm] [AMDGPU] Fix isCanonicalized for IEEE-mode-dependent sNaN quieting (PR #208733)

[Matt Arsenault via llvm-commits]

================
@@ -15960,8 +15960,37 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
   if (MaxDepth == 0)
     return false;
 
+  // Prevent llvm.canonicalize from being silently dropped when the hardware
+  // won't quiet sNaN based on the following hardware sNaN quieting behavior.
+  //
+  // Pre-GFX12:
+  //   Always quiet: v_cvt_f16_f32, v_cvt_pkrtz_f16_f32, v_dot2_f32_f16,
+  //                 v_cvt_f32_ubyte{0-3} (integer->float, never produces NaN).
+  //   IEEE-dependent: all other float ops quiet sNaN only when MODE.IEEE=1.
+  //   Never quiets: DIV_SCALE passes sNaN through even at MODE.IEEE=1.
+  //
+  // GFX12+: all exception-producing ops unconditionally quiet sNaN regardless
+  // of MODE.IEEE.
+  const bool IsGFX12Plus = Subtarget->getGeneration() >= AMDGPUSubtarget::GFX12;
----------------
arsenm wrote:

The feature you would really want is DX10ClampAndIEEEMode

https://github.com/llvm/llvm-project/pull/208733