ARM64 vmmov experiment: Reduce precision by using FMUL+FADD instead o…

…f FMADD. May help #12082 and thus also #11179 and #9843.
hrydgard · Jun 4, 2019 · 649b7a5 · 649b7a5 · hrydgard · Jun 4, 2019
1 parent c4d26dc
commit 649b7a5
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/Core/MIPS/ARM64/Arm64CompVFPU.cpp b/Core/MIPS/ARM64/Arm64CompVFPU.cpp
@@ -1240,7 +1240,8 @@ namespace MIPSComp {
 					fp.FMUL(fpr.V(dregs[a * 4 + b]), fpr.V(sregs[b * 4]), fpr.V(tregs[a * 4]));
 					for (int c = 1; c < n; c++) {
 						fpr.MapDirtyInInV(dregs[a * 4 + b], sregs[b * 4 + c], tregs[a * 4 + c], false);
-						fp.FMADD(fpr.V(dregs[a * 4 + b]), fpr.V(sregs[b * 4 + c]), fpr.V(tregs[a * 4 + c]), fpr.V(dregs[a * 4 + b]));
+						fp.FMUL(S0, fpr.V(sregs[b * 4 + c]), fpr.V(tregs[a * 4 + c]));
+						fp.FADD(fpr.V(dregs[a * 4 + b]), fpr.V(dregs[a * 4 + b]), S0);
 					}
 				}
 			}