[llvm-branch-commits] [llvm] [AMDGPU] Prefer fma over fmac for f16 operands in the high half (PR #242072)

Dmitry Sidorov via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Nov 12 15:50:07 PST 2026


================
@@ -4948,6 +4948,17 @@ AMDGPUInstructionSelector::selectVOP3NoMods(MachineOperand &Root) const {
   }};
 }
 
+InstructionSelector::ComplexRendererFns
+AMDGPUInstructionSelector::selectVOP3NoOpSelMods(MachineOperand &Root) const {
+  Register Reg = Root.getReg();
+  Register Unused;
+  if (!Subtarget->useRealTrue16Insts() &&
+      MRI->getType(Reg).getSizeInBits() == 16 &&
----------------
MrSidims wrote:

indeed, thanks!

https://github.com/llvm/llvm-project/pull/242072


More information about the llvm-branch-commits mailing list