[PATCH] Implement aarch64 neon instruction class AdvSIMD (by element) - Clang
liujiangning1 at gmail.com
Mon Sep 30 04:21:45 PDT 2013
Comment at: utils/TableGen/NeonEmitter.cpp:1553-1554
@@ -1517,3 +1552,4 @@
+ case OpFMALane:
s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
Tim Northover wrote:
> Sorry, I should have spotted this earlier, but I think that an fma is semantically distinct from the C expression "a+b*c".
> The latter we *can* transform into an fma in -ffast-math mode, but not generally. So we probably need an intrinsic to make sure that if the user asks for a fused multiply, that's the operation they get.
> Actually, there'a already a generic llvm intrinsic to handle this: "@llvm.fma.*"; we should probably use that rather than an ARM/AArch64-specific one.
I see your point, so do you mean we must generate fmla instruction for intrinsic function vfma_lane_f32(), no matter if it is in -ffast-math mode or not? Then I think we have to generate fmls for intrinsic function vfms_lane_f32() as well. I don't see LLVM IR has @llvm.fms.* defined, so we have to define an aarch64 specific LLVM intrinsic, or we can use an expression containing llvm.fma.* to represent it? Since it's a fused calculation, I doubt we should do the latter. I want to confirm with you before modifying the code.
More information about the llvm-commits