[llvm] [AMDGPU] Recognise bitmask operations as srcmods (PR #149110)

[Janek van Oirschot via llvm-commits]

================
@@ -3036,6 +3036,36 @@ bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
     Src = Src.getOperand(0);
   }
 
+  // Convert various sign-bit masks to src mods. Currently disabled for 16-bit
+  // types as the codegen replaces the operand without adding a srcmod.
+  // Recognise (xor a, 0x80000000) as NEG SrcMod.
----------------
JanekvO wrote:

Might want to add explicit note that we're talking about float neg/abs operations being done on ints, otherwise people might misinterpret it as 2s compliment neg/abs.

https://github.com/llvm/llvm-project/pull/149110