From 69a1d58a2b53709c7d257d7757dd3ea48823f8f7 Mon Sep 17 00:00:00 2001
From: MerryMage <MerryMage@users.noreply.github.com>
Date: Sun, 21 Jun 2020 10:00:24 +0100
Subject: [PATCH] A32: Implement ASIMD VMULL

---
 src/frontend/A32/decoder/asimd.inc            |  2 +-
 .../A32/translate/impl/asimd_three_same.cpp   | 23 +++++++++++++++++++
 .../A32/translate/impl/translate_arm.h        |  1 +
 tests/A32/fuzz_arm.cpp                        |  1 +
 4 files changed, 26 insertions(+), 1 deletion(-)

diff --git a/src/frontend/A32/decoder/asimd.inc b/src/frontend/A32/decoder/asimd.inc
index 99cdfb37..6ff00de6 100644
--- a/src/frontend/A32/decoder/asimd.inc
+++ b/src/frontend/A32/decoder/asimd.inc
@@ -30,7 +30,7 @@ INST(asimd_VCEQ_reg,        "VCEG (register)",          "111100110Dzznnnndddd100
 INST(asimd_VMLA,            "VMLA/VMLS",                "1111001o0Dzznnnndddd1001NQM0mmmm") // ASIMD
 //INST(asimd_VMLAL,           "VMLAL/VMLSL",              "1111001U1Dzznnnndddd10o0N0M0mmmm") // ASIMD
 INST(asimd_VMUL,            "VMUL",                     "1111001P0Dzznnnndddd1001NQM1mmmm") // ASIMD
-//INST(asimd_VMULL,           "VMULL",                    "1111001U1Dzznnnndddd11o0N0M0mmmm") // ASIMD
+INST(asimd_VMULL,           "VMULL",                    "1111001U1Dzznnnndddd11P0N0M0mmmm") // ASIMD
 //INST(asimd_VPMAX,           "VPMAX/VPMIN",              "1111001U0-CC--------1010---B----") // ASIMD
 //INST(asimd_VQDMULH,         "VQDMULH",                  "111100100-CC--------1011---0----") // ASIMD
 //INST(asimd_VQRDMULH,        "VQRDMULH",                 "111100110-CC--------1011---0----") // ASIMD
diff --git a/src/frontend/A32/translate/impl/asimd_three_same.cpp b/src/frontend/A32/translate/impl/asimd_three_same.cpp
index 8aef3d28..95f34713 100644
--- a/src/frontend/A32/translate/impl/asimd_three_same.cpp
+++ b/src/frontend/A32/translate/impl/asimd_three_same.cpp
@@ -550,6 +550,29 @@ bool ArmTranslatorVisitor::asimd_VMUL(bool P, bool D, size_t sz, size_t Vn, size
     return true;
 }
 
+bool ArmTranslatorVisitor::asimd_VMULL(bool U, bool D, size_t sz, size_t Vn, size_t Vd, bool P, bool N, bool M, size_t Vm) {
+    if (sz == 0b11 || (P & (U || sz == 0b10)) || Common::Bit<0>(Vd)) {
+        return UndefinedInstruction();
+    }
+
+    const size_t esize = P ? (sz == 0b00 ? 8 : 64) : 8U << sz;
+    const auto d = ToVector(true, Vd, D);
+    const auto m = ToVector(false, Vm, M);
+    const auto n = ToVector(false, Vn, N);
+
+    const auto extend_reg = [&](const auto& reg) {
+        return U ? ir.VectorZeroExtend(esize, reg) : ir.VectorSignExtend(esize, reg);
+    };
+
+    const auto reg_n = ir.GetVector(n);
+    const auto reg_m = ir.GetVector(m);
+    const auto result = P ? ir.VectorPolynomialMultiplyLong(esize, reg_m, reg_n)
+                          : ir.VectorMultiply(2 * esize, extend_reg(reg_m), extend_reg(reg_n));
+
+    ir.SetVector(d, result);
+    return true;
+}
+
 bool ArmTranslatorVisitor::asimd_VPADD(bool D, size_t sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm) {
     if (Q || sz == 0b11) {
         return UndefinedInstruction();
diff --git a/src/frontend/A32/translate/impl/translate_arm.h b/src/frontend/A32/translate/impl/translate_arm.h
index a9d3b690..ea4db9eb 100644
--- a/src/frontend/A32/translate/impl/translate_arm.h
+++ b/src/frontend/A32/translate/impl/translate_arm.h
@@ -476,6 +476,7 @@ struct ArmTranslatorVisitor final {
     bool asimd_VCEQ_reg(bool D, size_t sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
     bool asimd_VMLA(bool op, bool D, size_t sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
     bool asimd_VMUL(bool P, bool D, size_t sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
+    bool asimd_VMULL(bool U, bool D, size_t sz, size_t Vn, size_t Vd, bool P, bool N, bool M, size_t Vm);
     bool asimd_VPADD(bool D, size_t sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
     bool asimd_VADD_float(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
     bool asimd_VSUB_float(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
diff --git a/tests/A32/fuzz_arm.cpp b/tests/A32/fuzz_arm.cpp
index fb14b7ff..47a2bfac 100644
--- a/tests/A32/fuzz_arm.cpp
+++ b/tests/A32/fuzz_arm.cpp
@@ -64,6 +64,7 @@ bool ShouldTestInst(u32 instruction, u32 pc, bool is_last_inst) {
         // Currently unimplemented in Unicorn
         case IR::Opcode::FPVectorRecipEstimate16:
         case IR::Opcode::FPVectorRSqrtEstimate16:
+        case IR::Opcode::VectorPolynomialMultiplyLong64:
             return false;
         default:
             continue;