A64: Implement SQRDMULH's by-index vector variant

This commit is contained in:
Lioncash 2019-03-20 14:05:41 -04:00 committed by MerryMage
parent 42b090d234
commit 224ff0afaa
2 changed files with 27 additions and 1 deletions

View file

@ -874,7 +874,7 @@ INST(MUL_elt, "MUL (by element)", "0Q001
INST(SMULL_elt, "SMULL, SMULL2 (by element)", "0Q001111zzLMmmmm1010H0nnnnnddddd") INST(SMULL_elt, "SMULL, SMULL2 (by element)", "0Q001111zzLMmmmm1010H0nnnnnddddd")
//INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001111zzLMmmmm1011H0nnnnnddddd") //INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001111zzLMmmmm1011H0nnnnnddddd")
INST(SQDMULH_elt_2, "SQDMULH (by element)", "0Q001111zzLMmmmm1100H0nnnnnddddd") INST(SQDMULH_elt_2, "SQDMULH (by element)", "0Q001111zzLMmmmm1100H0nnnnnddddd")
//INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd") INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd")
INST(SDOT_elt, "SDOT (by element)", "0Q001111zzLMmmmm1110H0nnnnnddddd") INST(SDOT_elt, "SDOT (by element)", "0Q001111zzLMmmmm1110H0nnnnnddddd")
//INST(FMLA_elt_3, "FMLA (by element)", "0Q00111100LMmmmm0001H0nnnnnddddd") //INST(FMLA_elt_3, "FMLA (by element)", "0Q00111100LMmmmm0001H0nnnnnddddd")
INST(FMLA_elt_4, "FMLA (by element)", "0Q0011111zLMmmmm0001H0nnnnnddddd") INST(FMLA_elt_4, "FMLA (by element)", "0Q0011111zLMmmmm0001H0nnnnnddddd")

View file

@ -246,6 +246,32 @@ bool TranslatorVisitor::SQDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, I
return true; return true;
} }
bool TranslatorVisitor::SQRDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
if (size == 0b00 || size == 0b11) {
return UnallocatedEncoding();
}
const size_t idxsize = H == 1 ? 128 : 64;
const size_t esize = 8 << size.ZeroExtend();
const size_t datasize = Q ? 128 : 64;
const auto [index, Vmhi] = [=] {
if (size == 0b01) {
return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0});
}
return std::make_pair(concatenate(H, L).ZeroExtend(), M);
}();
const IR::U128 operand1 = V(datasize, Vn);
const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend<Vec>());
const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index));
const IR::UpperAndLower multiply = ir.VectorSignedSaturatedDoublingMultiply(esize, operand1, index_vector);
const IR::U128 result = ir.VectorAdd(esize, multiply.upper, ir.VectorLogicalShiftRight(esize, multiply.lower, static_cast<u8>(esize - 1)));
V(datasize, Vd, result);
return true;
}
bool TranslatorVisitor::SDOT_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { bool TranslatorVisitor::SDOT_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return DotProduct(*this, Q, size, L, M, Vmlo, H, Vn, Vd, &IREmitter::SignExtendToWord); return DotProduct(*this, Q, size, L, M, Vmlo, H, Vn, Vd, &IREmitter::SignExtendToWord);
} }