Archive/jak-project
1
0
Fork 0
mirror of https://github.com/open-goal/jak-project.git synced 2024-10-20 21:27:52 -04:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity
jak-project/goalc/compiler/compilation/Asm.cpp
water111 cc8801a27b
[goalc] speed up jak3 compilation (#3454) 
I noticed that jak 3's compilation was spending a lot of time accessing
the `unordered_map`s we use to store constants and symbol types.

 
I repurposed the `EnvironmentMap` originally made for GOOS for this. It
turns out that we were copying the entire constant map whenever we
encountered a `deftype`, and fixed that too.

This speeds up jak3 compiles from ~16 to 11 seconds for me.
2024-04-06 16:01:17 -04:00

1484 lines
62 KiB
C++
Raw Blame History

#include "goalc/compiler/Compiler.h"
namespace {
const char* reg_names[] = {
"rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", "r8", "r9", "r10",
"r11", "r12", "r13", "r14", "r15", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
"xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15",
};
}
emitter::Register Compiler::parse_register(const goos::Object& code) {
if (!code.is_symbol()) {
throw_compiler_error(code, "Could not parse {} as a register name", code.print());
}
auto nas = code.as_symbol();
for (int i = 0; i < 32; i++) {
if (std::string_view(nas.name_ptr) == reg_names[i]) {
return emitter::Register(i);
}
}
throw_compiler_error(code, "Could not parse {} as a register name", code.print());
return {};
}
Val* Compiler::compile_rlet(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
if (args.unnamed.size() < 1 || !args.named.empty()) {
throw_compiler_error(form, "Must have an rlet body.");
}
auto defs = args.unnamed.front();
auto fenv = env->function_env();
auto lenv = fenv->alloc_env<LexicalEnv>(env);
std::vector<IRegConstraint> constraints;
std::vector<RegVal*> reset_regs;
for_each_in_list(defs, [&](const goos::Object& o) {
// (new-place [:reg old-place] [:type type-spec] [:class reg-type] [:bind #f|lexical|lambda])
auto def_args = get_va(o, o);
va_check(o, def_args, {goos::ObjectType::SYMBOL},
{{"reg", {false, goos::ObjectType::SYMBOL}},
{"type", {false, {}}},
{"reset-here", {false, {}}},
{"class", {false, goos::ObjectType::SYMBOL}}});
// get the name of the new place
auto new_place_name = def_args.unnamed.at(0);
// get the type of the new place
TypeSpec ts = m_ts.make_typespec("object");
if (def_args.has_named("type")) {
ts = parse_typespec(def_args.named.at("type"), env);
}
// figure out the class
RegClass register_class = RegClass::GPR_64;
if (def_args.has_named("class")) {
auto& class_name = def_args.named.at("class").as_symbol();
if (class_name == "gpr") {
register_class = RegClass::GPR_64;
} else if (class_name == "fpr") {
register_class = RegClass::FLOAT;
} else if (class_name == "vf") {
register_class = RegClass::VECTOR_FLOAT;
} else if (class_name == "i128") {
register_class = RegClass::INT_128;
} else {
throw_compiler_error(o, "Register class {} is unknown.", class_name.name_ptr);
}
}
// alloc a register:
RegVal* new_place_reg = nullptr;
if (def_args.has_named("reg")) {
auto desired_register = parse_register(def_args.named.at("reg"));
// we want to see if we already created a variable for this register, and reuse it.
for (auto& constr : fenv->constraints()) {
if (constr.desired_register == desired_register && constr.contrain_everywhere) {
auto reg_val_ptr = std::make_unique<RegVal>(constr.ireg, ts);
new_place_reg = fenv->push_reg_val(std::move(reg_val_ptr));
new_place_reg->mark_as_settable();
break;
}
}
if (!new_place_reg) {
for (auto& constr : constraints) {
if (constr.desired_register == desired_register && constr.contrain_everywhere) {
auto reg_val_ptr = std::make_unique<RegVal>(constr.ireg, ts);
new_place_reg = fenv->push_reg_val(std::move(reg_val_ptr));
new_place_reg->mark_as_settable();
break;
}
}
}
}
if (!new_place_reg) {
new_place_reg = env->make_ireg(ts, register_class);
new_place_reg->mark_as_settable();
if (def_args.has_named("reg")) {
IRegConstraint constraint;
constraint.ireg = new_place_reg->ireg();
constraint.contrain_everywhere = true;
constraint.desired_register = parse_register(def_args.named.at("reg"));
if (def_args.has_named("reset-here") &&
get_true_or_false(form, def_args.get_named("reset-here"))) {
reset_regs.push_back(new_place_reg);
}
new_place_reg->set_rlet_constraint(constraint.desired_register);
constraints.push_back(constraint);
}
}
lenv->vars[new_place_name.as_symbol()] = new_place_reg;
});
if (!reset_regs.empty()) {
lenv->emit_ir<IR_ValueReset>(form, reset_regs);
}
for (auto c : constraints) {
fenv->constrain(c);
}
Val* result = get_none();
for (u64 i = 1; i < args.unnamed.size(); i++) {
auto& o = args.unnamed.at(i);
result = compile_error_guard(o, lenv);
if (!dynamic_cast<None*>(result)) {
result = result->to_reg(o, lenv);
}
}
return result;
}
Val* Compiler::compile_asm_ret(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
env->emit_ir<IR_AsmRet>(form, color);
return get_none();
}
Val* Compiler::compile_asm_pop(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto pop_dest = compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env);
if (!pop_dest->settable()) {
throw_compiler_error(form, "Cannot pop into this destination. Got a {}.", pop_dest->print());
}
env->emit_ir<IR_AsmPop>(form, color, pop_dest);
return get_none();
}
Val* Compiler::compile_asm_push(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
env->emit_ir<IR_AsmPush>(form, color,
compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env));
return get_none();
}
Val* Compiler::compile_asm_sub(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot .sub this. Got a {}.", dest->print());
}
auto src = compile_error_guard(args.unnamed.at(1), env)->to_gpr(form, env);
env->emit_ir<IR_AsmSub>(form, color, dest, src);
return get_none();
}
Val* Compiler::compile_asm_add(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot .add this. Got a {}.", dest->print());
}
auto src = compile_error_guard(args.unnamed.at(1), env)->to_gpr(form, env);
env->emit_ir<IR_AsmAdd>(form, color, dest, src);
return get_none();
}
Val* Compiler::compile_asm_load_sym(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {goos::ObjectType::SYMBOL}},
{{"sext", {false, goos::ObjectType::SYMBOL}}, {"color", {false, goos::ObjectType::SYMBOL}}});
auto& sym_name = args.unnamed.at(1).as_symbol();
const auto* sym_ts = m_symbol_types.lookup(sym_name);
if (!sym_ts) {
throw_compiler_error(form, "Cannot find a symbol named {}.", sym_name.name_ptr);
}
bool sext = m_ts.lookup_type(*sym_ts)->get_load_signed();
if (args.has_named("sext")) {
sext = get_true_or_false(form, args.named.at("sext"));
}
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot .load-sym this. Got a {}.", dest->print());
}
env->emit_ir<IR_GetSymbolValueAsm>(form, color, dest, sym_name.name_ptr, sext);
return get_none();
}
Val* Compiler::compile_asm_jr(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto src = compile_error_guard(args.unnamed.at(0), env)->to_gpr(form, env);
env->emit_ir<IR_JumpReg>(form, color, src);
return get_none();
}
Val* Compiler::compile_asm_mov(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot .mov this. Got a {}.", dest->print());
}
auto src = compile_error_guard(args.unnamed.at(1), env)->to_reg(form, env);
env->emit_ir<IR_RegSetAsm>(form, color, dest, src);
return get_none();
}
Val* Compiler::compile_asm_nop_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {}, {});
env->emit_ir<IR_AsmFNop>(form);
return get_none();
}
Val* Compiler::compile_asm_wait_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {}, {});
env->emit_ir<IR_AsmFWait>(form);
return get_none();
}
/*!
* Load a vector float from memory. Does an aligned load.
*/
Val* Compiler::compile_asm_lvf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}},
{"offset", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
if (!dest->settable() || dest->ireg().reg_class != RegClass::VECTOR_FLOAT) {
throw_compiler_error(form, "Cannot .lvf into this. Got a {}.", dest->print());
}
auto src = compile_error_guard(args.unnamed.at(1), env);
auto as_sv = dynamic_cast<StaticVal*>(src);
// Loading directly from a static value is not supported!
if (as_sv && args.has_named("offset")) {
throw_compiler_error(form, "Cannot .lvf from a static value");
} else if (as_sv && !args.has_named("offset")) {
env->emit_ir<IR_StaticVarLoad>(form, dest, as_sv->obj);
return get_none();
}
auto as_co = dynamic_cast<MemoryOffsetConstantVal*>(src);
RegVal* baseReg = as_co ? as_co->base->to_gpr(form, env) : src->to_gpr(form, env);
int offset = 0;
if (as_co) {
if (!args.has_named("offset")) {
offset = as_co->offset;
} else {
offset = as_co->offset + args.named.at("offset").as_int();
}
} else if (args.has_named("offset")) {
offset = args.named.at("offset").as_int();
}
MemLoadInfo info;
info.sign_extend = false;
info.size = 16;
info.reg = RegClass::VECTOR_FLOAT;
env->emit_ir<IR_LoadConstOffset>(form, dest, offset, baseReg, info, color);
return get_none();
}
/*!
* Store a vector float into memory. Does an aligned load.
*/
Val* Compiler::compile_asm_svf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}},
{"offset", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_reg(form, env);
if (!src->settable() || (src->ireg().reg_class != RegClass::VECTOR_FLOAT &&
src->ireg().reg_class != RegClass::INT_128)) {
throw_compiler_error(form, "Cannot .svf from this. Got a {}.", src->print());
}
auto as_co = dynamic_cast<MemoryOffsetConstantVal*>(dest);
RegVal* baseReg = as_co ? as_co->base->to_gpr(form, env) : dest->to_gpr(form, env);
int offset = 0;
if (as_co) {
if (!args.has_named("offset")) {
offset = as_co->offset;
} else {
offset = as_co->offset + args.named.at("offset").as_int();
}
} else if (args.has_named("offset")) {
offset = args.named.at("offset").as_int();
}
MemLoadInfo info;
info.sign_extend = false;
info.size = 16;
info.reg = RegClass::VECTOR_FLOAT;
env->emit_ir<IR_StoreConstOffset>(form, src, offset, baseReg, info.size, color);
return get_none();
}
void Compiler::check_vector_float_regs(const goos::Object& form,
Env*,
std::vector<std::pair<std::string, RegVal*>> args) {
for (std::pair<std::string, RegVal*> arg : args) {
if (!arg.second->settable() || arg.second->ireg().reg_class != RegClass::VECTOR_FLOAT) {
throw_compiler_error(form, "Invalid {} register for a vector float operation form. Got a {}.",
arg.first, arg.second->print());
}
}
}
Val* Compiler::compile_asm_mov_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
check_vector_float_regs(form, env, {{"destination", dest}, {"source", src}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
env->emit_ir<IR_BlendVF>(form, color, dest, dest, src, mask);
return get_none();
}
Val* Compiler::compile_asm_blend_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
env->emit_ir<IR_BlendVF>(form, color, dest, src1, src2, mask);
return get_none();
}
Val* Compiler::compile_asm_vf_math3(const goos::Object& form,
const goos::Object& rest,
IR_VFMath3Asm::Kind kind,
emitter::Register::VF_ELEMENT broadcastElement,
Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
// If there is a broadcast register, splat that float across the entire src2 register before
// performing the operation For example vaddx.xyzw vf10, vf20, vf30
// vf10[x] = vf20[x] + vf30[x]
// vf10[y] = vf20[y] + vf30[x]
// vf10[z] = vf20[z] + vf30[x]
// vf10[w] = vf20[w] + vf30[x]
if (broadcastElement != emitter::Register::VF_ELEMENT::NONE) {
auto temp_reg = env->make_vfr(dest->type());
env->emit_ir<IR_SplatVF>(form, color, temp_reg, src2, broadcastElement);
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_VFMath3Asm>(form, color, dest, src1, temp_reg, kind);
} else {
// Perform the arithmetic operation on the two vectors into a temporary register
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src1, temp_reg, kind);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
} else {
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_VFMath3Asm>(form, color, dest, src1, src2, kind);
} else {
auto temp_reg = env->make_vfr(dest->type());
// Perform the arithmetic operation on the two vectors into a temporary register
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src1, src2, kind);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
}
return get_none();
}
Val* Compiler::compile_asm_int128_math3(const goos::Object& form,
const goos::Object& rest,
IR_Int128Math3Asm::Kind kind,
Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot set destination");
}
env->emit_ir<IR_Int128Math3Asm>(form, color, dest, src1, src2, kind);
return get_none();
}
Val* Compiler::compile_asm_vf_math2(const goos::Object& form,
const goos::Object& rest,
IR_VFMath2Asm::Kind kind,
Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
check_vector_float_regs(form, env, {{"destination", dest}, {"source", src}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_VFMath2Asm>(form, color, dest, src, kind);
} else {
auto temp_reg = env->make_vfr(dest->type());
// Perform the arithmetic operation on the two vectors into a temporary register
env->emit_ir<IR_VFMath2Asm>(form, color, temp_reg, src, kind);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
return get_none();
}
Val* Compiler::compile_asm_int128_math2_imm_u8(const goos::Object& form,
const goos::Object& rest,
IR_Int128Math2Asm::Kind kind,
Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
s64 imm = get_constant_integer_or_error(args.unnamed.at(2), env);
if (imm < 0 || imm > 255) {
throw_compiler_error(form, "Immediate {} is invalid. The value {} is out of range for a uint8.",
args.unnamed.at(2).print(), imm);
}
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_Int128Math2Asm>(form, color, dest, src, kind, imm);
} else {
auto temp_reg = env->make_vfr(dest->type());
// Perform the arithmetic operation on the two vectors into a temporary register
env->emit_ir<IR_Int128Math2Asm>(form, color, temp_reg, src, kind, imm);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
return get_none();
}
Val* Compiler::compile_asm_pw_sll(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math2_imm_u8(form, rest, IR_Int128Math2Asm::Kind::PW_SLL, env);
}
Val* Compiler::compile_asm_pw_srl(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math2_imm_u8(form, rest, IR_Int128Math2Asm::Kind::PW_SRL, env);
}
Val* Compiler::compile_asm_pw_sra(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math2_imm_u8(form, rest, IR_Int128Math2Asm::Kind::PW_SRA, env);
}
Val* Compiler::compile_asm_por(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::POR, env);
}
Val* Compiler::compile_asm_pxor(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PXOR, env);
}
bool ireg_is_128_ok(const IRegister& ireg) {
return ireg.reg_class == RegClass::VECTOR_FLOAT || ireg.reg_class == RegClass::INT_128;
}
Val* Compiler::compile_asm_pnor(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {});
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env); // rs
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env); // rt
auto temp = env->make_ireg(TypeSpec("uint128"), RegClass::INT_128);
if (!ireg_is_128_ok(dest->ireg())) {
throw_compiler_error(args.unnamed.at(0), "bad destination register kind");
}
if (!ireg_is_128_ok(src1->ireg())) {
throw_compiler_error(args.unnamed.at(1), "bad src1 register kind");
}
if (!ireg_is_128_ok(src2->ireg())) {
throw_compiler_error(args.unnamed.at(2), "bad src2 register kind");
}
if (!dest->settable()) {
throw_compiler_error(form, "Cannot set destination");
}
// A NOR is equivalent to an inverted input AND
// Use the destination register and a temp register.
// There is also no PNOT instruction, so the easiest way to is to XOR with an all-ones register
// Remember - we do not want to mutate the source registers!
// How do we create an all-ones-register? Compare it with itself
env->emit_ir<IR_Int128Math3Asm>(form, true, temp, temp, temp, IR_Int128Math3Asm::Kind::PCEQW);
// Then NOT the first input, store in destination
env->emit_ir<IR_Int128Math3Asm>(form, true, dest, temp, src1, IR_Int128Math3Asm::Kind::PXOR);
// NOT the second input, we not longer require the all-ones-register
env->emit_ir<IR_Int128Math3Asm>(form, true, temp, temp, src2, IR_Int128Math3Asm::Kind::PXOR);
// Preform the AND aka NOR
env->emit_ir<IR_Int128Math3Asm>(form, true, dest, dest, temp, IR_Int128Math3Asm::Kind::PAND);
return get_none();
}
Val* Compiler::compile_asm_pand(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PAND, env);
}
Val* Compiler::compile_asm_pceqb(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCEQB, env);
}
Val* Compiler::compile_asm_pceqh(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCEQH, env);
}
Val* Compiler::compile_asm_pceqw(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCEQW, env);
}
Val* Compiler::compile_asm_pcgtb(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCGTB, env);
}
Val* Compiler::compile_asm_pcgth(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCGTH, env);
}
Val* Compiler::compile_asm_pcgtw(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCGTW, env);
}
Val* Compiler::compile_asm_pextub(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTUB, env);
}
Val* Compiler::compile_asm_pextuh(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTUH, env);
}
Val* Compiler::compile_asm_pextuw(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTUW, env);
}
Val* Compiler::compile_asm_pextlb(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTLB, env);
}
Val* Compiler::compile_asm_pextlh(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTLH, env);
}
Val* Compiler::compile_asm_pextlw(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PEXTLW, env);
}
Val* Compiler::compile_asm_pcpyud(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCPYUD, env);
}
Val* Compiler::compile_asm_pcpyld(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PCPYLD, env);
}
Val* Compiler::compile_asm_psubw(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PSUBW, env);
}
Val* Compiler::compile_asm_paddb(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_int128_math3(form, rest, IR_Int128Math3Asm::Kind::PADDB, env);
}
Val* Compiler::compile_asm_ppach(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {});
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env); // rs
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env); // rt
auto temp = env->make_ireg(TypeSpec("uint128"), RegClass::INT_128);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot set destination");
}
env->emit_ir<IR_Int128Math2Asm>(form, true, temp, src1, IR_Int128Math2Asm::Kind::VPSHUFLW, 0x88);
env->emit_ir<IR_Int128Math2Asm>(form, true, dest, src2, IR_Int128Math2Asm::Kind::VPSHUFLW, 0x88);
env->emit_ir<IR_Int128Math2Asm>(form, true, temp, temp, IR_Int128Math2Asm::Kind::VPSHUFHW, 0x88);
env->emit_ir<IR_Int128Math2Asm>(form, true, dest, dest, IR_Int128Math2Asm::Kind::VPSHUFHW, 0x88);
env->emit_ir<IR_Int128Math2Asm>(form, true, temp, temp, IR_Int128Math2Asm::Kind::VPSRLDQ, 4);
env->emit_ir<IR_Int128Math2Asm>(form, true, dest, dest, IR_Int128Math2Asm::Kind::VPSRLDQ, 4);
// is actually a VPUNPCKLQDQ with srcs swapped.
env->emit_ir<IR_Int128Math3Asm>(form, true, dest, temp, dest, IR_Int128Math3Asm::Kind::PCPYLD);
return get_none();
}
Val* Compiler::compile_asm_ppacb(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {});
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env); // rs
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env); // rt
auto temp = env->make_ireg(TypeSpec("uint128"), RegClass::INT_128);
if (!dest->settable()) {
throw_compiler_error(form, "Cannot set destination");
}
if (dest->ireg().reg_class != RegClass::VECTOR_FLOAT &&
dest->ireg().reg_class != RegClass::INT_128) {
throw_compiler_error(form, "Destination must be vector float or int128");
}
env->emit_ir<IR_RegSet>(form, temp, src1);
env->emit_ir<IR_RegSet>(form, dest, src2);
env->emit_ir<IR_Int128Math2Asm>(form, true, temp, temp, IR_Int128Math2Asm::Kind::PH_SLL, 8);
env->emit_ir<IR_Int128Math2Asm>(form, true, dest, dest, IR_Int128Math2Asm::Kind::PH_SLL, 8);
env->emit_ir<IR_Int128Math2Asm>(form, true, temp, temp, IR_Int128Math2Asm::Kind::PH_SRL, 8);
env->emit_ir<IR_Int128Math2Asm>(form, true, dest, dest, IR_Int128Math2Asm::Kind::PH_SRL, 8);
// swapped on purpose, ps2 is backward (first arg goes low on x86, first arg goes high on ps2)
env->emit_ir<IR_Int128Math3Asm>(form, true, dest, dest, temp, IR_Int128Math3Asm::Kind::PACKUSWB);
return get_none();
}
Val* Compiler::compile_asm_xorp(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {});
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env); // rs
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env); // rt
if (!dest->settable()) {
throw_compiler_error(form, "Cannot set destination");
}
env->emit_ir<IR_VFMath3Asm>(form, true, dest, src1, src2, IR_VFMath3Asm::Kind::XOR);
return get_none();
}
Val* Compiler::compile_asm_itof_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math2(form, rest, IR_VFMath2Asm::Kind::ITOF, env);
}
Val* Compiler::compile_asm_ftoi_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math2(form, rest, IR_VFMath2Asm::Kind::FTOI, env);
}
Val* Compiler::compile_asm_xor_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::XOR,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_max_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MAX,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_max_x_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MAX,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_max_y_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MAX,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_max_z_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MAX,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_max_w_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MAX,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_min_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MIN,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_min_x_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MIN,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_min_y_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MIN,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_min_z_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MIN,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_min_w_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MIN,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_sub_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_sub_x_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_sub_y_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_sub_z_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_sub_w_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_add_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_add_x_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_add_y_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_add_z_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_add_w_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_mul_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MUL,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_mul_x_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MUL,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_mul_y_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MUL,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_mul_z_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MUL,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_mul_w_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
return compile_asm_vf_math3(form, rest, IR_VFMath3Asm::Kind::MUL,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_vf_math4_two_operation(const goos::Object& form,
const goos::Object& rest,
IR_VFMath3Asm::Kind first_op_kind,
IR_VFMath3Asm::Kind second_op_kind,
emitter::Register::VF_ELEMENT broadcastElement,
Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}, {}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
// This third register is intended for the ACC/Q/ETC, and is used to temporarily store the value
// that eventually goes into the destination
//
// For example VMADDA:
// > ACC += src1 * src2
// > DEST = ACC
auto src3 = compile_error_guard(args.unnamed.at(3), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest},
{"first source", src1},
{"second source", src2},
{"third source", src3}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(form, "The value {} is out of range for a blend mask (0-15 inclusive).",
mask);
}
}
// First we clear a temporary register
auto temp_reg = env->make_vfr(dest->type());
// If there is a broadcast register, splat that float across the entire src2 register before
// performing the operation For example vaddx.xyzw vf10, vf20, vf30
// vf10[x] = vf20[x] + vf30[x]
// vf10[y] = vf20[y] + vf30[x]
// vf10[z] = vf20[z] + vf30[x]
// vf10[w] = vf20[w] + vf30[x]
if (broadcastElement != emitter::Register::VF_ELEMENT::NONE) {
env->emit_ir<IR_SplatVF>(form, color, temp_reg, src2, broadcastElement);
// Perform the first operation
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src1, temp_reg, first_op_kind);
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_VFMath3Asm>(form, color, dest, src3, temp_reg, second_op_kind);
} else {
// Perform the second operation on the two vectors into the temporary register
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src3, temp_reg, second_op_kind);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
} else {
// Perform the first operation
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src1, src2, first_op_kind);
// If the entire destination is to be copied, we can optimize out the blend
if (mask == 0b1111) {
env->emit_ir<IR_VFMath3Asm>(form, color, dest, src3, temp_reg, second_op_kind);
} else {
// Perform the second operation on the two vectors into the temporary register
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src3, temp_reg, second_op_kind);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
}
return get_none();
}
Val* Compiler::compile_asm_mul_add_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_mul_add_x_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_mul_add_y_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_mul_add_z_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_mul_add_w_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::ADD,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_mul_sub_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::NONE, env);
}
Val* Compiler::compile_asm_mul_sub_x_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::X, env);
}
Val* Compiler::compile_asm_mul_sub_y_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::Y, env);
}
Val* Compiler::compile_asm_mul_sub_z_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::Z, env);
}
Val* Compiler::compile_asm_mul_sub_w_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
return compile_asm_vf_math4_two_operation(form, rest, IR_VFMath3Asm::Kind::MUL,
IR_VFMath3Asm::Kind::SUB,
emitter::Register::VF_ELEMENT::W, env);
}
Val* Compiler::compile_asm_abs_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"mask", {false, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
check_vector_float_regs(form, env, {{"destination", dest}, {"source", src}});
u8 mask = 0b1111;
if (args.has_named("mask")) {
mask = args.named.at("mask").as_int();
if (mask > 15) {
throw_compiler_error(
form, "The value {} is out of range for a destination mask (0-15 inclusive).", mask);
}
}
// There is no single instruction ABS on AVX, so there are a number of ways to do it manually,
// this is one of them. For example, assume the original vec = <1, -2, -3, 4>
// First we clear a temporary register, XOR'ing itself
auto temp_reg = env->make_vfr(dest->type());
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, temp_reg, temp_reg, IR_VFMath3Asm::Kind::XOR);
// Next, find the difference between our source operand and 0, use the same temp register, no need
// to use another <0, 0, 0, 0> - <1, -2, -3, 4> = <-1, 2, 3, 4>
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, temp_reg, src, IR_VFMath3Asm::Kind::SUB);
// Finally, find the maximum between our difference, and the original value
// MAX_OF(<-1, 2, 3, 4>, <1, -2, -3, 4>) = <1, 2, 3, 4>
if (mask == 0b1111) { // If the entire destination is to be copied, we can optimize out the blend
env->emit_ir<IR_VFMath3Asm>(form, color, dest, src, temp_reg, IR_VFMath3Asm::Kind::MAX);
} else {
env->emit_ir<IR_VFMath3Asm>(form, color, temp_reg, src, temp_reg, IR_VFMath3Asm::Kind::MAX);
// Blend the result back into the destination register using the mask
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp_reg, mask);
}
return get_none();
}
u8 Compiler::ftf_fsf_to_blend_mask(u8 val) {
// 00 -> x
// ...
// 11 -> w
return 0b0001 << val;
}
emitter::Register::VF_ELEMENT Compiler::ftf_fsf_to_vector_element(u8 val) {
// 00 -> x
// ...
// 11 -> w
switch (val) {
case 0b00:
return emitter::Register::VF_ELEMENT::X;
case 0b01:
return emitter::Register::VF_ELEMENT::Y;
case 0b10:
return emitter::Register::VF_ELEMENT::Z;
case 0b11:
return emitter::Register::VF_ELEMENT::W;
default:
throw_compiler_error_no_code("Invalid vector element {}", (int)val);
return emitter::Register::VF_ELEMENT::NONE;
}
}
Val* Compiler::compile_asm_div_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}},
{
{"color", {false, goos::ObjectType::SYMBOL}},
{"fsf", {true, goos::ObjectType::INTEGER}},
{"ftf", {true, goos::ObjectType::INTEGER}},
});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
u8 fsf = args.named.at("fsf").as_int();
if (fsf > 3) {
throw_compiler_error(form, "The value {} is out of range for fsf (0-3 inclusive).", fsf);
}
u8 ftf = args.named.at("ftf").as_int();
if (ftf > 3) {
throw_compiler_error(form, "The value {} is out of range for ftf (0-3 inclusive).", ftf);
}
// VDIV in the VU stores its result in a single 32bit `Q` Register, it does not compute the packed
// division result
//
// Further more, you can mix and match the vector elements (ex. src1's X component divided by
// src2's Y) Because of this, we need to blend the two components into corresponding locations,
// perform the divide then place into the cleared dest. register.
//
// Why do we even bother using VDIVPS instead of FDIV? Because otherwise in x86, you have to use
// the FPU stack Registers are nicer.
auto temp_reg1 = env->make_vfr(dest->type());
auto temp_reg2 = env->make_vfr(dest->type());
// Splat src1's value into a temp reg, keep it simple, this way no matter which vector component
// is accessed from the final result will be the correct answer
env->emit_ir<IR_SplatVF>(form, color, temp_reg1, src1, ftf_fsf_to_vector_element(fsf));
// Splat src1's value into the the temp reg
env->emit_ir<IR_SplatVF>(form, color, temp_reg2, src2, ftf_fsf_to_vector_element(ftf));
// Perform the Division
env->emit_ir<IR_VFMath3Asm>(form, color, dest, temp_reg1, temp_reg2, IR_VFMath3Asm::Kind::DIV);
return get_none();
}
Val* Compiler::compile_asm_sqrt_vf(const goos::Object& form, const goos::Object& rest, Env* env) {
auto args = get_va(form, rest);
va_check(
form, args, {{}, {}},
{{"color", {false, goos::ObjectType::SYMBOL}}, {"ftf", {true, goos::ObjectType::INTEGER}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
check_vector_float_regs(form, env, {{"destination", dest}, {"source", src}});
u8 ftf = args.named.at("ftf").as_int();
if (ftf > 3) {
throw_compiler_error(form, "The value {} is out of range for ftf (0-3 inclusive).", ftf);
}
// VSQRT in the VU stores its result in a single 32bit `Q` Register, it does not compute the
// packed division result
//
// Because of this, we need to blend the relevent component into a cleared register and then
// perform the SQRT
//
// Why do we even bother using VSQRTPS instead of FSQRT? Because otherwise in x86, you have to use
// the FPU stack Registers are nicer.
// Splat src's value into the dest reg, keep it simple, this way no matter which vector component
// is accessed from the final result will be the correct answer
env->emit_ir<IR_SplatVF>(form, color, dest, src, ftf_fsf_to_vector_element(ftf));
env->emit_ir<IR_SqrtVF>(form, color, dest, dest);
return get_none();
}
Val* Compiler::compile_asm_inv_sqrt_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}},
{
{"color", {false, goos::ObjectType::SYMBOL}},
{"fsf", {true, goos::ObjectType::INTEGER}},
{"ftf", {true, goos::ObjectType::INTEGER}},
});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
u8 fsf = args.named.at("fsf").as_int();
if (fsf > 3) {
throw_compiler_error(form, "The value {} is out of range for fsf (0-3 inclusive).", fsf);
}
u8 ftf = args.named.at("ftf").as_int();
if (ftf > 3) {
throw_compiler_error(form, "The value {} is out of range for ftf (0-3 inclusive).", ftf);
}
// Save one temp reg, use the destination as one
auto temp_reg = env->make_vfr(dest->type());
// Splat src1's value into the dest reg, keep it simple, this way no matter which vector component
// is accessed from the final result will be the correct answer
env->emit_ir<IR_SplatVF>(form, color, dest, src1, ftf_fsf_to_vector_element(fsf));
// Splat src1's value into the the temp reg
env->emit_ir<IR_SplatVF>(form, color, temp_reg, src2, ftf_fsf_to_vector_element(ftf));
// Square Root the temp reg
env->emit_ir<IR_SqrtVF>(form, color, temp_reg, temp_reg);
// Perform the Division
env->emit_ir<IR_VFMath3Asm>(form, color, dest, dest, temp_reg, IR_VFMath3Asm::Kind::DIV);
return get_none();
}
Val* Compiler::compile_asm_outer_product_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
// Given 2 vectors V1 = <1,2,3,4> and V2 = <5,6,7,8> and assume VDEST = <0, 0, 0, 999>
// The outer product is computed like so (only x,y,z components are operated on):
// x = (V1y * V2z) - (V2y * V1z) => (2 * 7) - (6 * 3) => -4
// y = (V1z * V2x) - (V2z * V1x) => (3 * 5) - (7 * 1) => 8
// z = (V1x * V2y) - (V2x * V1y) => (1 * 6) - (5 * 2) => -4
// w = N/A, left alone => 999
//
// There is probably a more optimized alg for this, but we can just do this in two stages
// First swizzle the first two vectors accordingly, and store in `dest`
// Then follow up with the second half.
//
// Some temporary regs are required AND its important to not modify dest's `w` or the source
// registers at all
// Init two temp registers
auto temp1 = env->make_vfr(dest->type());
auto temp2 = env->make_vfr(dest->type());
auto temp_dst = env->make_vfr(dest->type());
// First Portion
// - Swizzle src1 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp1, src1, 0b00001001);
// - Move it into 'dest' safely (avoid mutating `w`)
env->emit_ir<IR_BlendVF>(form, color, temp_dst, temp_dst, temp1, 0b0111);
// - Swizzle src2 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp1, src2, 0b00010010);
// - Multiply - Result in `dest`
env->emit_ir<IR_VFMath3Asm>(form, color, temp1, temp_dst, temp1, IR_VFMath3Asm::Kind::MUL);
// - Move it into 'dest' safely (avoid mutating `w`)
env->emit_ir<IR_BlendVF>(form, color, temp_dst, temp_dst, temp1, 0b0111);
// Second Portion
// - Swizzle src2 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp1, src2, 0b00001001);
// - Swizzle src1 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp2, src1, 0b00010010);
// - Multiply - Result in `temp1`
env->emit_ir<IR_VFMath3Asm>(form, color, temp1, temp1, temp2, IR_VFMath3Asm::Kind::MUL);
// Finalize
// - Subtract
env->emit_ir<IR_VFMath3Asm>(form, color, temp2, temp_dst, temp1, IR_VFMath3Asm::Kind::SUB);
// - Blend result, as to avoid not modifying dest's `w` component
env->emit_ir<IR_BlendVF>(form, color, dest, dest, temp2, 0b0111);
return get_none();
}
Val* Compiler::compile_asm_outer_product_a_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest}, {"first source", src1}, {"second source", src2}});
// Given 2 vectors V1 = <1,2,3,4> and V2 = <5,6,7,8> and assume VDEST = <0, 0, 0, 999>
// The outer product is computed like so (only x,y,z components are operated on):
// x = V1y * V2z => (2 * 7) => 14
// y = V1z * V2x => (3 * 5) => 15
// z = V1x * V2y => (1 * 6) => 6
// w = N/A, left alone => 999
//
// Some temporary regs are required AND its important to not modify dest's `w` or the source
// registers at all
// Init two temp registers
auto temp1 = env->make_vfr(dest->type());
auto temp2 = env->make_vfr(dest->type());
// - Swizzle src1 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp1, src1, 0b00001001);
// - Swizzle src2 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp2, src2, 0b00010010);
// - Multiply - Result in `dest`
env->emit_ir<IR_VFMath3Asm>(form, color, temp1, temp1, temp2, IR_VFMath3Asm::Kind::MUL);
// - Move it into 'dest' safely (avoid mutating `w`)
env->emit_ir<IR_BlendVF>(form, color, dest, temp1, temp1, 0b0111);
return get_none();
}
Val* Compiler::compile_asm_outer_product_b_vf(const goos::Object& form,
const goos::Object& rest,
Env* env) {
auto args = get_va(form, rest);
va_check(form, args, {{}, {}, {}, {}}, {{"color", {false, goos::ObjectType::SYMBOL}}});
bool color = true;
if (args.has_named("color")) {
color = get_true_or_false(form, args.named.at("color"));
}
auto dest = compile_error_guard(args.unnamed.at(0), env)->to_reg(form, env);
auto src1 = compile_error_guard(args.unnamed.at(1), env)->to_xmm128(form, env);
auto src2 = compile_error_guard(args.unnamed.at(2), env)->to_xmm128(form, env);
auto acc = compile_error_guard(args.unnamed.at(3), env)->to_xmm128(form, env);
check_vector_float_regs(form, env,
{{"destination", dest},
{"first source", src1},
{"second source", src2},
{"acc source", acc}});
// Given 2 vectors V1 = <1,2,3,4> and V2 = <5,6,7,8> and assume VDEST = <0, 0, 0, 999>
// also assume ACC = <14, 15, 6>
// The outer product is computed like so (only x,y,z components are operated on):
// x = ACCx - (V1y * V2z) => 14 - (2 * 7) => 0
// y = ACCy - (V1z * V2x) => 15 - (3 * 5) => 0
// z = ACCz - (V1x * V2y) => 6 - (1 * 6) => 0
// w = N/A, left alone => 999
//
// Some temporary regs are required AND its important to not modify dest's `w` or the source
// registers at all
// Init two temp registers
auto temp1 = env->make_vfr(dest->type());
auto temp2 = env->make_vfr(dest->type());
// - Swizzle src1 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp1, src1, 0b00001001);
// - Swizzle src2 appropriately
env->emit_ir<IR_SwizzleVF>(form, color, temp2, src2, 0b00010010);
// - Multiply - Result in `dest`
env->emit_ir<IR_VFMath3Asm>(form, color, temp1, temp1, temp2, IR_VFMath3Asm::Kind::MUL);
// - Subtract - (ACC - Result Above)
env->emit_ir<IR_VFMath3Asm>(form, color, temp1, acc, temp1, IR_VFMath3Asm::Kind::SUB);
// - Move it into 'dest' safely (avoid mutating `w`)
env->emit_ir<IR_BlendVF>(form, color, dest, temp1, temp1, 0b0111);
return get_none();
}
Reference in a new issue View git blame Copy permalink