[clang][X86] Wrong result for __builtin_elementwise_fma on _Float16

[SEt-t]

Godbolt: https://godbolt.org/z/Ydj17K17b

Clang uses single-precision FMA to emulate half-precision FMA, what is wrong as it doesn't have enough precision.

Example, round to even: 0x1.400p+8 * 0x1.008p+7 + 0x1.000p-24
Precise result: 0x1.40a0000002p+15
Half-precision FMA: 0x1.40cp+15
Single-precision FMA: 0x1.40a000p+15
(clang) Single-precision FMA -> half-precision: 0x1.408p+15

Another example: 0x1.eb8p-12 * 0x1.9p-11 - 0x1p-11

To produce correct result single-precision multiplication, then double-precision addition seems to be enough.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: SEt (SEt-t)

Godbolt: https://godbolt.org/z/Ydj17K17b

Clang uses single-precision FMA to emulate half-precision FMA, what is wrong as it doesn't have enough precision.

Example, round to even: 0x1.400p+8 * 0x1.008p+7 + 0x1.000p-24
Precise result: 0x1.40a0000002p+15
Half-precision FMA: 0x1.40cp+15
Single-precision FMA: 0x1.40a000p+15
(clang) Single-precision FMA -> half-precision: 0x1.408p+15

Another example: 0x1.eb8p-12 * 0x1.9p-11 - 0x1p-11

To produce correct result single-precision multiplication, then double-precision addition seems to be enough.

[akshaykumars614]

I want to work on this issue. Any input or suggestions on where to start would be really helpful!

[beetrees]

This bug is a duplicate of #98389.

@akshaykumars614 AFAICT this is where the half FMA operation is lowered incorrectly:

llvm-project/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

Lines 3396 to 3414 in 5cc2ae0

	SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FMAD(SDNode *N) {
	EVT OVT = N->getValueType(0);
	EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
	SDValue Op0 = GetSoftPromotedHalf(N->getOperand(0));
	SDValue Op1 = GetSoftPromotedHalf(N->getOperand(1));
	SDValue Op2 = GetSoftPromotedHalf(N->getOperand(2));
	SDLoc dl(N);
	// Promote to the larger FP type.
	auto PromotionOpcode = GetPromotionOpcode(OVT, NVT);
	Op0 = DAG.getNode(PromotionOpcode, dl, NVT, Op0);
	Op1 = DAG.getNode(PromotionOpcode, dl, NVT, Op1);
	Op2 = DAG.getNode(PromotionOpcode, dl, NVT, Op2);
	SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, Op0, Op1, Op2);
	// Convert back to FP16 as an integer.
	return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
	}