basic vectored types support for FIPR and FTRV

src/jit/backend/x64/x64_backend.cc

@@ -73,8 +73,16 @@ const Register x64_registers[] = {
      reinterpret_cast<const void *>(&Xbyak::util::xmm9)},
     {"xmm10", ir::VALUE_FLOAT_MASK,
      reinterpret_cast<const void *>(&Xbyak::util::xmm10)},
-    {"xmm11", ir::VALUE_FLOAT_MASK,
-     reinterpret_cast<const void *>(&Xbyak::util::xmm11)}};
+    {"xmm11", ir::VALUE_VECTOR_MASK,
+     reinterpret_cast<const void *>(&Xbyak::util::xmm11)},
+    {"xmm12", ir::VALUE_VECTOR_MASK,
+     reinterpret_cast<const void *>(&Xbyak::util::xmm12)},
+    {"xmm13", ir::VALUE_VECTOR_MASK,
+     reinterpret_cast<const void *>(&Xbyak::util::xmm13)},
+    {"xmm14", ir::VALUE_VECTOR_MASK,
+     reinterpret_cast<const void *>(&Xbyak::util::xmm14)},
+    {"xmm15", ir::VALUE_VECTOR_MASK,
+     reinterpret_cast<const void *>(&Xbyak::util::xmm15)}};
 
 const int x64_num_registers = sizeof(x64_registers) / sizeof(Register);
 

src/jit/backend/x64/x64_emitter.cc

@@ -577,7 +577,18 @@ EMITTER(STORE_GUEST) {
 EMITTER(LOAD_CONTEXT) {
   int offset = instr->arg0()->i32();
 
-  if (IsFloatType(instr->type())) {
+  if (IsVectorType(instr->type())) {
+    const Xbyak::Xmm result = e.GetXmmRegister(instr);
+
+    switch (instr->type()) {
+      case VALUE_V128:
+        e.movups(result, e.ptr[e.r14 + offset]);
+        break;
+      default:
+        LOG_FATAL("Unexpected result type");
+        break;
+    }
+  } else if (IsFloatType(instr->type())) {
     const Xbyak::Xmm result = e.GetXmmRegister(instr);
 
     switch (instr->type()) {
@@ -638,7 +649,18 @@ EMITTER(STORE_CONTEXT) {
         break;
     }
   } else {
-    if (IsFloatType(instr->arg1()->type())) {
+    if (IsVectorType(instr->arg1()->type())) {
+      const Xbyak::Xmm src = e.GetXmmRegister(instr->arg1());
+
+      switch (instr->arg1()->type()) {
+        case VALUE_V128:
+          e.vmovups(e.ptr[e.r14 + offset], src);
+          break;
+        default:
+          LOG_FATAL("Unexpected result type");
+          break;
+      }
+    } else if (IsFloatType(instr->arg1()->type())) {
       const Xbyak::Xmm src = e.GetXmmRegister(instr->arg1());
 
       switch (instr->arg1()->type()) {
@@ -679,7 +701,18 @@ EMITTER(STORE_CONTEXT) {
 EMITTER(LOAD_LOCAL) {
   int offset = STACK_OFFSET_LOCALS + instr->arg0()->i32();
 
-  if (IsFloatType(instr->type())) {
+  if (IsVectorType(instr->type())) {
+    const Xbyak::Xmm result = e.GetXmmRegister(instr);
+
+    switch (instr->type()) {
+      case VALUE_V128:
+        e.movups(result, e.ptr[e.rsp + offset]);
+        break;
+      default:
+        LOG_FATAL("Unexpected result type");
+        break;
+    }
+  } else if (IsFloatType(instr->type())) {
     const Xbyak::Xmm result = e.GetXmmRegister(instr);
 
     switch (instr->type()) {
@@ -721,7 +754,18 @@ EMITTER(STORE_LOCAL) {
 
   CHECK(!instr->arg1()->constant());
 
-  if (IsFloatType(instr->arg1()->type())) {
+  if (IsVectorType(instr->arg1()->type())) {
+    const Xbyak::Xmm src = e.GetXmmRegister(instr->arg1());
+
+    switch (instr->arg1()->type()) {
+      case VALUE_V128:
+        e.vmovups(e.ptr[e.rsp + offset], src);
+        break;
+      default:
+        LOG_FATAL("Unexpected result type");
+        break;
+    }
+  } else if (IsFloatType(instr->arg1()->type())) {
     const Xbyak::Xmm src = e.GetXmmRegister(instr->arg1());
 
     switch (instr->arg1()->type()) {
@@ -1152,6 +1196,37 @@ EMITTER(SQRT) {
   }
 }
 
+EMITTER(VBROADCAST) {
+  const Xbyak::Xmm result = e.GetXmmRegister(instr);
+  const Xbyak::Xmm a = e.GetXmmRegister(instr->arg0());
+
+  e.vbroadcastss(result, a);
+}
+
+EMITTER(VADD) {
+  const Xbyak::Xmm result = e.GetXmmRegister(instr);
+  const Xbyak::Xmm a = e.GetXmmRegister(instr->arg0());
+  const Xbyak::Xmm b = e.GetXmmRegister(instr->arg1());
+
+  e.vaddps(result, a, b);
+}
+
+EMITTER(VDOT) {
+  const Xbyak::Xmm result = e.GetXmmRegister(instr);
+  const Xbyak::Xmm a = e.GetXmmRegister(instr->arg0());
+  const Xbyak::Xmm b = e.GetXmmRegister(instr->arg1());
+
+  e.vdpps(result, a, b, 0b11110001);
+}
+
+EMITTER(VMUL) {
+  const Xbyak::Xmm result = e.GetXmmRegister(instr);
+  const Xbyak::Xmm a = e.GetXmmRegister(instr->arg0());
+  const Xbyak::Xmm b = e.GetXmmRegister(instr->arg1());
+
+  e.vmulps(result, a, b);
+}
+
 EMITTER(AND) {
   const Xbyak::Reg result = e.GetRegister(instr);
   const Xbyak::Reg a = e.GetRegister(instr->arg0());

src/jit/frontend/sh4/sh4_builder.h

@@ -26,12 +26,12 @@ class SH4Builder : public ir::IRBuilder {
 
   void Emit(uint32_t addr, int max_instrs);
 
-  ir::Value *LoadRegister(int n, ir::ValueType type);
-  void StoreRegister(int n, ir::Value *v);
-  ir::Value *LoadRegisterF(int n, ir::ValueType type);
-  void StoreRegisterF(int n, ir::Value *v);
-  ir::Value *LoadRegisterXF(int n, ir::ValueType type);
-  void StoreRegisterXF(int n, ir::Value *v);
+  ir::Value *LoadGPR(int n, ir::ValueType type);
+  void StoreGPR(int n, ir::Value *v);
+  ir::Value *LoadFPR(int n, ir::ValueType type);
+  void StoreFPR(int n, ir::Value *v);
+  ir::Value *LoadXFR(int n, ir::ValueType type);
+  void StoreXFR(int n, ir::Value *v);
   ir::Value *LoadSR();
   void StoreSR(ir::Value *v);
   ir::Value *LoadT();

src/jit/frontend/sh4/sh4_context.h

@@ -74,6 +74,11 @@ struct SH4Context {
   // mapping for each pair of single-precision registers is instead swapped by
   // XOR'ing the actual index with 1. for example, fr2 becomes fr[3] and fr3
   // becomes fr[2], enabling dr2 to perfectly alias fr[2]
+
+  // note note, this incorrectly causes fv registers to be swizzled. fv0 should
+  // be loaded as {fr0, fr1, fr2, fr3} but it's actually loaded as
+  // {fr1, fr0, fr3, fr2}. however, due to the way the FV registers are
+  // used (FIPR and FTRV) this doesn't actually affect the results
   uint32_t fr[16], xf[16];
 };
 }

src/jit/ir/ir_builder.cc

@@ -401,6 +401,44 @@ Instr *IRBuilder::Sqrt(Value *a) {
   return instr;
 }
 
+Instr *IRBuilder::VBroadcast(Value *a) {
+  CHECK(a->type() == VALUE_F32);
+
+  Instr *instr = AppendInstr(OP_VBROADCAST, VALUE_V128);
+  instr->set_arg0(a);
+  return instr;
+}
+
+Instr *IRBuilder::VAdd(Value *a, Value *b, ValueType el_type) {
+  CHECK(IsVectorType(a->type()) && IsVectorType(b->type()));
+  CHECK_EQ(el_type, VALUE_F32);
+
+  Instr *instr = AppendInstr(OP_VADD, a->type());
+  instr->set_arg0(a);
+  instr->set_arg1(b);
+  return instr;
+}
+
+Instr *IRBuilder::VDot(Value *a, Value *b, ValueType el_type) {
+  CHECK(IsVectorType(a->type()) && IsVectorType(b->type()));
+  CHECK_EQ(el_type, VALUE_F32);
+
+  Instr *instr = AppendInstr(OP_VDOT, el_type);
+  instr->set_arg0(a);
+  instr->set_arg1(b);
+  return instr;
+}
+
+Instr *IRBuilder::VMul(Value *a, Value *b, ValueType el_type) {
+  CHECK(IsVectorType(a->type()) && IsVectorType(b->type()));
+  CHECK_EQ(el_type, VALUE_F32);
+
+  Instr *instr = AppendInstr(OP_VMUL, a->type());
+  instr->set_arg0(a);
+  instr->set_arg1(b);
+  return instr;
+}
+
 Instr *IRBuilder::And(Value *a, Value *b) {
   CHECK(IsIntType(a->type()) && a->type() == b->type());
 

src/jit/ir/ir_builder.h

@@ -30,6 +30,7 @@ enum ValueType {
   VALUE_I64,
   VALUE_F32,
   VALUE_F64,
+  VALUE_V128,
   VALUE_NUM,
 };
 
@@ -40,9 +41,11 @@ enum {
   VALUE_I64_MASK = 1 << VALUE_I64,
   VALUE_F32_MASK = 1 << VALUE_F32,
   VALUE_F64_MASK = 1 << VALUE_F64,
+  VALUE_V128_MASK = 1 << VALUE_V128,
   VALUE_INT_MASK =
       VALUE_I8_MASK | VALUE_I16_MASK | VALUE_I32_MASK | VALUE_I64_MASK,
   VALUE_FLOAT_MASK = VALUE_F32_MASK | VALUE_F64_MASK,
+  VALUE_VECTOR_MASK = VALUE_V128_MASK,
   VALUE_ALL_MASK = VALUE_INT_MASK | VALUE_FLOAT_MASK,
 };
 
@@ -53,11 +56,16 @@ enum {
 class Instr;
 class Use;
 
+static inline bool IsIntType(ValueType type) {
+  return type == VALUE_I8 || type == VALUE_I16 || type == VALUE_I32 ||
+         type == VALUE_I64;
+}
+
 static inline bool IsFloatType(ValueType type) {
   return type == VALUE_F32 || type == VALUE_F64;
 }
 
-static inline bool IsIntType(ValueType type) { return !IsFloatType(type); }
+static inline bool IsVectorType(ValueType type) { return type == VALUE_V128; }
 
 static inline int SizeForType(ValueType type) {
   switch (type) {
@@ -73,6 +81,8 @@ static inline int SizeForType(ValueType type) {
       return 4;
     case VALUE_F64:
       return 8;
+    case VALUE_V128:
+      return 16;
     default:
       LOG_FATAL("Unexpected value type");
       break;
@@ -373,7 +383,7 @@ class IRBuilder {
   Instr *FCmpLE(Value *a, Value *b);
   Instr *FCmpLT(Value *a, Value *b);
 
-  // math operators
+  // integer math operators
   Instr *Add(Value *a, Value *b);
   Instr *Sub(Value *a, Value *b);
   Instr *SMul(Value *a, Value *b);
@@ -381,6 +391,8 @@ class IRBuilder {
   Instr *Div(Value *a, Value *b);
   Instr *Neg(Value *a);
   Instr *Abs(Value *a);
+
+  // floating point math operators
   Instr *FAdd(Value *a, Value *b);
   Instr *FSub(Value *a, Value *b);
   Instr *FMul(Value *a, Value *b);
@@ -389,6 +401,12 @@ class IRBuilder {
   Instr *FAbs(Value *a);
   Instr *Sqrt(Value *a);
 
+  // vector math operators
+  Instr *VBroadcast(Value *a);
+  Instr *VAdd(Value *a, Value *b, ValueType el_type);
+  Instr *VDot(Value *a, Value *b, ValueType el_type);
+  Instr *VMul(Value *a, Value *b, ValueType el_type);
+
   // bitwise operations
   Instr *And(Value *a, Value *b);
   Instr *Or(Value *a, Value *b);

src/jit/ir/ir_ops.inc

@@ -30,6 +30,10 @@ IR_OP(FDIV)
 IR_OP(FNEG)
 IR_OP(FABS)
 IR_OP(SQRT)
+IR_OP(VBROADCAST)
+IR_OP(VADD)
+IR_OP(VDOT)
+IR_OP(VMUL)
 IR_OP(AND)
 IR_OP(OR)
 IR_OP(XOR)

src/jit/ir/passes/register_allocation_pass.cc

@@ -82,7 +82,8 @@ void RegisterSet::InsertInterval(Interval *interval) {
 
 RegisterAllocationPass::RegisterAllocationPass(const Backend &backend)
     : int_registers_(backend.num_registers()),
-      float_registers_(backend.num_registers()) {
+      float_registers_(backend.num_registers()),
+      vector_registers_(backend.num_registers()) {
   registers_ = backend.registers();
   num_registers_ = backend.num_registers();
 
@@ -139,12 +140,17 @@ RegisterSet &RegisterAllocationPass::GetRegisterSet(ValueType type) {
     return float_registers_;
   }
 
+  if (IsVectorType(type)) {
+    return vector_registers_;
+  }
+
   LOG_FATAL("Unexpected value type");
 }
 
 void RegisterAllocationPass::Reset() {
   int_registers_.Clear();
   float_registers_.Clear();
+  vector_registers_.Clear();
 
   for (int i = 0; i < num_registers_; i++) {
     const Register &r = registers_[i];
@@ -153,10 +159,10 @@ void RegisterAllocationPass::Reset() {
       int_registers_.PushRegister(i);
     } else if (r.value_types == VALUE_FLOAT_MASK) {
       float_registers_.PushRegister(i);
+    } else if (r.value_types == VALUE_VECTOR_MASK) {
+      vector_registers_.PushRegister(i);
     } else {
-      LOG_FATAL(
-          "Unsupported register value mask, expected VALUE_INT_MASK or "
-          "VALUE_FLOAT_MASK");
+      LOG_FATAL("Unsupported register value mask");
     }
   }
 }
@@ -220,6 +226,7 @@ void RegisterAllocationPass::ExpireOldIntervals(Instr *instr) {
 
   expire_set(int_registers_);
   expire_set(float_registers_);
+  expire_set(vector_registers_);
 }
 
 // If the first argument isn't used after this instruction, its register

src/jit/ir/passes/register_allocation_pass.h

@@ -59,6 +59,7 @@ class RegisterAllocationPass : public Pass {
 
   RegisterSet int_registers_;
   RegisterSet float_registers_;
+  RegisterSet vector_registers_;
 
   // intervals, keyed by register
   Interval *intervals_;