From 6c2c68bce6c3589a576c461bfb343185db42f68d Mon Sep 17 00:00:00 2001
From: Lioncash <mathew1800@gmail.com>
Date: Thu, 11 Apr 2019 16:44:15 -0400
Subject: [PATCH] A64: Implement FCMLA's indexed element variant
With this, all of the instructions introduced with ARMv8.3-CompNum have
an implementation.
---
src/frontend/A64/decoder/a64.inc | 2 +-
.../impl/simd_vector_x_indexed_element.cpp | 89 +++++++++++++++++++
2 files changed, 90 insertions(+), 1 deletion(-)
diff --git a/src/frontend/A64/decoder/a64.inc b/src/frontend/A64/decoder/a64.inc
index 5bdca7e8..6d0ae98e 100644
--- a/src/frontend/A64/decoder/a64.inc
+++ b/src/frontend/A64/decoder/a64.inc
@@ -896,7 +896,7 @@ INST(UDOT_elt, "UDOT (by element)", "0Q101
//INST(SQRDMLSH_elt_2, "SQRDMLSH (by element)", "0Q101111zzLMmmmm1111H0nnnnnddddd")
//INST(FMULX_elt_3, "FMULX (by element)", "0Q10111100LMmmmm1001H0nnnnnddddd")
INST(FMULX_elt_4, "FMULX (by element)", "0Q1011111zLMmmmm1001H0nnnnnddddd")
-//INST(FCMLA_elt, "FCMLA (by element)", "0Q101111zzLMmmmm0rr1H0nnnnnddddd")
+INST(FCMLA_elt, "FCMLA (by element)", "0Q101111zzLMmmmm0rr1H0nnnnnddddd")
// Data Processing - FP and SIMD - Cryptographic three register
INST(SM3TT1A, "SM3TT1A", "11001110010mmmmm10ii00nnnnnddddd")
diff --git a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp
index 2fc62bab..8eda641d 100644
--- a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp
+++ b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp
@@ -187,6 +187,95 @@ bool TranslatorVisitor::MUL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4>
return MultiplyByElement(*this, Q, size, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::None);
}
+bool TranslatorVisitor::FCMLA_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<2> rot, Imm<1> H, Vec Vn, Vec Vd) {
+ if (size == 0b00 || size == 0b11) {
+ return ReservedValue();
+ }
+
+ if (size == 0b01 && H == 1 && Q == 0) {
+ return ReservedValue();
+ }
+
+ if (size == 0b10 && (L == 1 || Q == 0)) {
+ return ReservedValue();
+ }
+
+ const size_t esize = 8U << size.ZeroExtend();
+
+ // TODO: We don't support the half-precision floating point variant yet.
+ if (esize == 16) {
+ return InterpretThisInstruction();
+ }
+
+ const size_t index = [=] {
+ if (size == 0b01) {
+ return concatenate(H, L).ZeroExtend();
+ }
+ return H.ZeroExtend();
+ }();
+
+ const Vec Vm = concatenate(M, Vmlo).ZeroExtend<Vec>();
+
+ const size_t datasize = Q ? 128 : 64;
+ const size_t num_elements = datasize / esize;
+ const size_t num_iterations = num_elements / 2;
+
+ const IR::U128 operand1 = V(datasize, Vn);
+ const IR::U128 operand2 = V(datasize, Vm);
+ const IR::U128 operand3 = V(datasize, Vd);
+ IR::U128 result = ir.ZeroVector();
+
+ IR::U32U64 element1;
+ IR::U32U64 element2;
+ IR::U32U64 element3;
+ IR::U32U64 element4;
+ for (size_t e = 0; e < num_iterations; ++e) {
+ const size_t first = e * 2;
+ const size_t second = first + 1;
+
+ const size_t index_first = index * 2;
+ const size_t index_second = index_first + 1;
+
+ switch (rot.ZeroExtend()) {
+ case 0b00: // 0 degrees
+ element1 = ir.VectorGetElement(esize, operand2, index_first);
+ element2 = ir.VectorGetElement(esize, operand1, first);
+ element3 = ir.VectorGetElement(esize, operand2, index_second);
+ element4 = ir.VectorGetElement(esize, operand1, first);
+ break;
+ case 0b01: // 90 degrees
+ element1 = ir.FPNeg(ir.VectorGetElement(esize, operand2, index_second));
+ element2 = ir.VectorGetElement(esize, operand1, second);
+ element3 = ir.VectorGetElement(esize, operand2, index_first);
+ element4 = ir.VectorGetElement(esize, operand1, second);
+ break;
+ case 0b10: // 180 degrees
+ element1 = ir.FPNeg(ir.VectorGetElement(esize, operand2, index_first));
+ element2 = ir.VectorGetElement(esize, operand1, first);
+ element3 = ir.FPNeg(ir.VectorGetElement(esize, operand2, index_second));
+ element4 = ir.VectorGetElement(esize, operand1, first);
+ break;
+ case 0b11: // 270 degrees
+ element1 = ir.VectorGetElement(esize, operand2, index_second);
+ element2 = ir.VectorGetElement(esize, operand1, second);
+ element3 = ir.FPNeg(ir.VectorGetElement(esize, operand2, index_first));
+ element4 = ir.VectorGetElement(esize, operand1, second);
+ break;
+ }
+
+ const IR::U32U64 operand3_elem1 = ir.VectorGetElement(esize, operand3, first);
+ const IR::U32U64 operand3_elem2 = ir.VectorGetElement(esize, operand3, second);
+
+ result = ir.VectorSetElement(esize, result, first,
+ ir.FPMulAdd(operand3_elem1, element2, element1, true));
+ result = ir.VectorSetElement(esize, result, second,
+ ir.FPMulAdd(operand3_elem2, element4, element3, true));
+ }
+
+ ir.SetQ(Vd, result);
+ return true;
+}
+
bool TranslatorVisitor::FMLA_elt_4(bool Q, bool sz, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return FPMultiplyByElement(*this, Q, sz, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Accumulate);
}