ShuriZma
/
suyu
mirror of https://git.suyu.dev/suyu/suyu
1
0
Fork 1
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #1892 from Tinob/master 

Improve Zero flag implementation
This commit is contained in:
bunnei 2018-12-27 11:06:59 -05:00 committed by GitHub
parent 8e23e40a72 33056dd833
commit fa9acc26d9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 122 additions and 113 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

235
src/video_core/renderer_opengl/gl_shader_decompiler.cpp
View File

@ -347,6 +347,15 @@ public:
BuildInputList(); BuildInputList();
} }
void SetConditionalCodesFromExpression(const std::string& expresion) {
SetInternalFlag(InternalFlag::ZeroFlag, "(" + expresion + ") == 0");
LOG_WARNING(HW_GPU, "Condition codes implementation is incomplete.");
}
void SetConditionalCodesFromRegister(const Register& reg, u64 dest_elem = 0) {
SetConditionalCodesFromExpression(GetRegister(reg, static_cast<u32>(dest_elem)));
}
/** /**
* Returns code that does an integer size conversion for the specified size. * Returns code that does an integer size conversion for the specified size.
* @param value Value to perform integer size conversion on. * @param value Value to perform integer size conversion on.
@ -401,14 +410,24 @@ public:
* @param dest_num_components Number of components in the destination. * @param dest_num_components Number of components in the destination.
* @param value_num_components Number of components in the value. * @param value_num_components Number of components in the value.
* @param is_saturated Optional, when True, saturates the provided value. * @param is_saturated Optional, when True, saturates the provided value.
* @param sets_cc Optional, when True, sets the corresponding values to the implemented
* condition flags.
* @param dest_elem Optional, the destination element to use for the operation. * @param dest_elem Optional, the destination element to use for the operation.
*/ */
void SetRegisterToFloat(const Register& reg, u64 elem, const std::string& value, void SetRegisterToFloat(const Register& reg, u64 elem, const std::string& value,
u64 dest_num_components, u64 value_num_components, u64 dest_num_components, u64 value_num_components,
bool is_saturated = false, u64 dest_elem = 0, bool precise = false) { bool is_saturated = false, bool sets_cc = false, u64 dest_elem = 0,
bool precise = false) {
SetRegister(reg, elem, is_saturated ? "clamp(" + value + ", 0.0, 1.0)" : value, const std::string clamped_value = is_saturated ? "clamp(" + value + ", 0.0, 1.0)" : value;
dest_num_components, value_num_components, dest_elem, precise); SetRegister(reg, elem, clamped_value, dest_num_components, value_num_components, dest_elem,
precise);
if (sets_cc) {
if (reg == Register::ZeroIndex) {
SetConditionalCodesFromExpression(clamped_value);
} else {
SetConditionalCodesFromRegister(reg, dest_elem);
}
}
} }
/** /**
@ -419,25 +438,29 @@ public:
* @param dest_num_components Number of components in the destination. * @param dest_num_components Number of components in the destination.
* @param value_num_components Number of components in the value. * @param value_num_components Number of components in the value.
* @param is_saturated Optional, when True, saturates the provided value. * @param is_saturated Optional, when True, saturates the provided value.
* @param sets_cc Optional, when True, sets the corresponding values to the implemented
* condition flags.
* @param dest_elem Optional, the destination element to use for the operation. * @param dest_elem Optional, the destination element to use for the operation.
* @param size Register size to use for conversion instructions. * @param size Register size to use for conversion instructions.
*/ */
void SetRegisterToInteger(const Register& reg, bool is_signed, u64 elem, void SetRegisterToInteger(const Register& reg, bool is_signed, u64 elem,
const std::string& value, u64 dest_num_components, const std::string& value, u64 dest_num_components,
u64 value_num_components, bool is_saturated = false, u64 value_num_components, bool is_saturated = false,
u64 dest_elem = 0, Register::Size size = Register::Size::Word, bool sets_cc = false, u64 dest_elem = 0,
bool sets_cc = false) { Register::Size size = Register::Size::Word) {
UNIMPLEMENTED_IF(is_saturated); UNIMPLEMENTED_IF(is_saturated);
const std::string final_value = ConvertIntegerSize(value, size);
const std::string func{is_signed ? "intBitsToFloat" : "uintBitsToFloat"}; const std::string func{is_signed ? "intBitsToFloat" : "uintBitsToFloat"};
SetRegister(reg, elem, func + '(' + ConvertIntegerSize(value, size) + ')', SetRegister(reg, elem, func + '(' + final_value + ')', dest_num_components,
dest_num_components, value_num_components, dest_elem, false); value_num_components, dest_elem, false);
if (sets_cc) { if (sets_cc) {
const std::string zero_condition = "( " + ConvertIntegerSize(value, size) + " == 0 )"; if (reg == Register::ZeroIndex) {
SetInternalFlag(InternalFlag::ZeroFlag, zero_condition); SetConditionalCodesFromExpression(final_value);
LOG_WARNING(HW_GPU, "Condition codes implementation is incomplete."); } else {
SetConditionalCodesFromRegister(reg, dest_elem);
}
} }
} }
@ -1275,7 +1298,7 @@ private:
void WriteLogicOperation(Register dest, LogicOperation logic_op, const std::string& op_a, void WriteLogicOperation(Register dest, LogicOperation logic_op, const std::string& op_a,
const std::string& op_b, const std::string& op_b,
Tegra::Shader::PredicateResultMode predicate_mode, Tegra::Shader::PredicateResultMode predicate_mode,
Tegra::Shader::Pred predicate) { Tegra::Shader::Pred predicate, const bool set_cc) {
std::string result{}; std::string result{};
switch (logic_op) { switch (logic_op) {
case LogicOperation::And: { case LogicOperation::And: {
@ -1299,7 +1322,7 @@ private:
} }
if (dest != Tegra::Shader::Register::ZeroIndex) { if (dest != Tegra::Shader::Register::ZeroIndex) {
regs.SetRegisterToInteger(dest, true, 0, result, 1, 1); regs.SetRegisterToInteger(dest, true, 0, result, 1, 1, false, set_cc);
} }
using Tegra::Shader::PredicateResultMode; using Tegra::Shader::PredicateResultMode;
@ -1319,7 +1342,8 @@ private:
} }
void WriteLop3Instruction(Register dest, const std::string& op_a, const std::string& op_b, void WriteLop3Instruction(Register dest, const std::string& op_a, const std::string& op_b,
const std::string& op_c, const std::string& imm_lut) { const std::string& op_c, const std::string& imm_lut,
const bool set_cc) {
if (dest == Tegra::Shader::Register::ZeroIndex) { if (dest == Tegra::Shader::Register::ZeroIndex) {
return; return;
} }
@ -1342,7 +1366,7 @@ private:
result += ')'; result += ')';
regs.SetRegisterToInteger(dest, true, 0, result, 1, 1); regs.SetRegisterToInteger(dest, true, 0, result, 1, 1, false, set_cc);
} }
void WriteTexsInstructionFloat(const Instruction& instr, const std::string& texture) { void WriteTexsInstructionFloat(const Instruction& instr, const std::string& texture) {
@ -1357,12 +1381,12 @@ private:
if (written_components < 2) { if (written_components < 2) {
// Write the first two swizzle components to gpr0 and gpr0+1 // Write the first two swizzle components to gpr0 and gpr0+1
regs.SetRegisterToFloat(instr.gpr0, component, texture, 1, 4, false, regs.SetRegisterToFloat(instr.gpr0, component, texture, 1, 4, false, false,
written_components % 2); written_components % 2);
} else { } else {
ASSERT(instr.texs.HasTwoDestinations()); ASSERT(instr.texs.HasTwoDestinations());
// Write the rest of the swizzle components to gpr28 and gpr28+1 // Write the rest of the swizzle components to gpr28 and gpr28+1
regs.SetRegisterToFloat(instr.gpr28, component, texture, 1, 4, false, regs.SetRegisterToFloat(instr.gpr28, component, texture, 1, 4, false, false,
written_components % 2); written_components % 2);
} }
@ -1871,8 +1895,6 @@ private:
instr.fmul.tab5c68_0 != 1, "FMUL tab5cb8_0({}) is not implemented", instr.fmul.tab5c68_0 != 1, "FMUL tab5cb8_0({}) is not implemented",
instr.fmul.tab5c68_0 instr.fmul.tab5c68_0
.Value()); // SMO typical sends 1 here which seems to be the default .Value()); // SMO typical sends 1 here which seems to be the default
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in FMUL is not implemented");
op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b); op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b);
@ -1896,20 +1918,17 @@ private:
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b + postfactor_op, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b + postfactor_op, 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, instr.generates_cc, 0, true);
break; break;
} }
case OpCode::Id::FADD_C: case OpCode::Id::FADD_C:
case OpCode::Id::FADD_R: case OpCode::Id::FADD_R:
case OpCode::Id::FADD_IMM: { case OpCode::Id::FADD_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in FADD is not implemented");
op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a); op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a);
op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b); op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b);
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, instr.generates_cc, 0, true);
break; break;
} }
case OpCode::Id::MUFU: { case OpCode::Id::MUFU: {
@ -1917,31 +1936,31 @@ private:
switch (instr.sub_op) { switch (instr.sub_op) {
case SubOp::Cos: case SubOp::Cos:
regs.SetRegisterToFloat(instr.gpr0, 0, "cos(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "cos(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Sin: case SubOp::Sin:
regs.SetRegisterToFloat(instr.gpr0, 0, "sin(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "sin(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Ex2: case SubOp::Ex2:
regs.SetRegisterToFloat(instr.gpr0, 0, "exp2(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "exp2(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Lg2: case SubOp::Lg2:
regs.SetRegisterToFloat(instr.gpr0, 0, "log2(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "log2(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Rcp: case SubOp::Rcp:
regs.SetRegisterToFloat(instr.gpr0, 0, "1.0 / " + op_a, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "1.0 / " + op_a, 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Rsq: case SubOp::Rsq:
regs.SetRegisterToFloat(instr.gpr0, 0, "inversesqrt(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "inversesqrt(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
case SubOp::Sqrt: case SubOp::Sqrt:
regs.SetRegisterToFloat(instr.gpr0, 0, "sqrt(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "sqrt(" + op_a + ')', 1, 1,
instr.alu.saturate_d, 0, true); instr.alu.saturate_d, false, 0, true);
break; break;
default: default:
UNIMPLEMENTED_MSG("Unhandled MUFU sub op={0:x}", UNIMPLEMENTED_MSG("Unhandled MUFU sub op={0:x}",
@ -1952,8 +1971,9 @@ private:
case OpCode::Id::FMNMX_C: case OpCode::Id::FMNMX_C:
case OpCode::Id::FMNMX_R: case OpCode::Id::FMNMX_R:
case OpCode::Id::FMNMX_IMM: { case OpCode::Id::FMNMX_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in FMNMX is not implemented"); instr.generates_cc,
"Condition codes generation in FMNMX is partially implemented");
op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a); op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a);
op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b); op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b);
@ -1964,7 +1984,7 @@ private:
regs.SetRegisterToFloat(instr.gpr0, 0, regs.SetRegisterToFloat(instr.gpr0, 0,
'(' + condition + ") ? min(" + parameters + ") : max(" + '(' + condition + ") ? min(" + parameters + ") : max(" +
parameters + ')', parameters + ')',
1, 1, false, 0, true); 1, 1, false, instr.generates_cc, 0, true);
break; break;
} }
case OpCode::Id::RRO_C: case OpCode::Id::RRO_C:
@ -1989,18 +2009,16 @@ private:
break; break;
} }
case OpCode::Id::FMUL32_IMM: { case OpCode::Id::FMUL32_IMM: {
UNIMPLEMENTED_IF_MSG(instr.op_32.generates_cc, regs.SetRegisterToFloat(
"Condition codes generation in FMUL32 is not implemented"); instr.gpr0, 0,
regs.GetRegisterAsFloat(instr.gpr8) + " * " + GetImmediate32(instr), 1, 1,
regs.SetRegisterToFloat(instr.gpr0, 0, instr.fmul32.saturate, instr.op_32.generates_cc, 0, true);
regs.GetRegisterAsFloat(instr.gpr8) + " * " +
GetImmediate32(instr),
1, 1, instr.fmul32.saturate, 0, true);
break; break;
} }
case OpCode::Id::FADD32I: { case OpCode::Id::FADD32I: {
UNIMPLEMENTED_IF_MSG(instr.op_32.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in FADD32I is not implemented"); instr.op_32.generates_cc,
"Condition codes generation in FADD32I is partially implemented");
std::string op_a = regs.GetRegisterAsFloat(instr.gpr8); std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
std::string op_b = GetImmediate32(instr); std::string op_b = GetImmediate32(instr);
@ -2021,7 +2039,8 @@ private:
op_b = "-(" + op_b + ')'; op_b = "-(" + op_b + ')';
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, false, 0, true); regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, false,
instr.op_32.generates_cc, 0, true);
break; break;
} }
} }
@ -2035,16 +2054,14 @@ private:
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::BFE_IMM: { case OpCode::Id::BFE_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in BFE is not implemented");
std::string inner_shift = std::string inner_shift =
'(' + op_a + " << " + std::to_string(instr.bfe.GetLeftShiftValue()) + ')'; '(' + op_a + " << " + std::to_string(instr.bfe.GetLeftShiftValue()) + ')';
std::string outer_shift = std::string outer_shift =
'(' + inner_shift + " >> " + '(' + inner_shift + " >> " +
std::to_string(instr.bfe.GetLeftShiftValue() + instr.bfe.shift_position) + ')'; std::to_string(instr.bfe.GetLeftShiftValue() + instr.bfe.shift_position) + ')';
regs.SetRegisterToInteger(instr.gpr0, true, 0, outer_shift, 1, 1); regs.SetRegisterToInteger(instr.gpr0, true, 0, outer_shift, 1, 1, false,
instr.generates_cc);
break; break;
} }
default: { default: {
@ -2055,8 +2072,6 @@ private:
break; break;
} }
case OpCode::Type::Bfi: { case OpCode::Type::Bfi: {
UNIMPLEMENTED_IF(instr.generates_cc);
const auto [base, packed_shift] = [&]() -> std::tuple<std::string, std::string> { const auto [base, packed_shift] = [&]() -> std::tuple<std::string, std::string> {
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::BFI_IMM_R: case OpCode::Id::BFI_IMM_R:
@ -2071,9 +2086,10 @@ private:
const std::string offset = '(' + packed_shift + " & 0xff)"; const std::string offset = '(' + packed_shift + " & 0xff)";
const std::string bits = "((" + packed_shift + " >> 8) & 0xff)"; const std::string bits = "((" + packed_shift + " >> 8) & 0xff)";
const std::string insert = regs.GetRegisterAsInteger(instr.gpr8, 0, false); const std::string insert = regs.GetRegisterAsInteger(instr.gpr8, 0, false);
regs.SetRegisterToInteger( regs.SetRegisterToInteger(instr.gpr0, false, 0,
instr.gpr0, false, 0, "bitfieldInsert(" + base + ", " + insert + ", " + offset +
"bitfieldInsert(" + base + ", " + insert + ", " + offset + ", " + bits + ')', 1, 1); ", " + bits + ')',
1, 1, false, instr.generates_cc);
break; break;
} }
case OpCode::Type::Shift: { case OpCode::Type::Shift: {
@ -2095,9 +2111,6 @@ private:
case OpCode::Id::SHR_C: case OpCode::Id::SHR_C:
case OpCode::Id::SHR_R: case OpCode::Id::SHR_R:
case OpCode::Id::SHR_IMM: { case OpCode::Id::SHR_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in SHR is not implemented");
if (!instr.shift.is_signed) { if (!instr.shift.is_signed) {
// Logical shift right // Logical shift right
op_a = "uint(" + op_a + ')'; op_a = "uint(" + op_a + ')';
@ -2105,7 +2118,7 @@ private:
// Cast to int is superfluous for arithmetic shift, it's only for a logical shift // Cast to int is superfluous for arithmetic shift, it's only for a logical shift
regs.SetRegisterToInteger(instr.gpr0, true, 0, "int(" + op_a + " >> " + op_b + ')', regs.SetRegisterToInteger(instr.gpr0, true, 0, "int(" + op_a + " >> " + op_b + ')',
1, 1); 1, 1, false, instr.generates_cc);
break; break;
} }
case OpCode::Id::SHL_C: case OpCode::Id::SHL_C:
@ -2113,7 +2126,8 @@ private:
case OpCode::Id::SHL_IMM: case OpCode::Id::SHL_IMM:
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in SHL is not implemented"); "Condition codes generation in SHL is not implemented");
regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " << " + op_b, 1, 1); regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " << " + op_b, 1, 1, false,
instr.generates_cc);
break; break;
default: { default: {
UNIMPLEMENTED_MSG("Unhandled shift instruction: {}", opcode->get().GetName()); UNIMPLEMENTED_MSG("Unhandled shift instruction: {}", opcode->get().GetName());
@ -2127,18 +2141,17 @@ private:
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::IADD32I: case OpCode::Id::IADD32I:
UNIMPLEMENTED_IF_MSG(instr.op_32.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in IADD32I is not implemented"); instr.op_32.generates_cc,
"Condition codes generation in IADD32I is partially implemented");
if (instr.iadd32i.negate_a) if (instr.iadd32i.negate_a)
op_a = "-(" + op_a + ')'; op_a = "-(" + op_a + ')';
regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " + " + op_b, 1, 1, regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " + " + op_b, 1, 1,
instr.iadd32i.saturate != 0); instr.iadd32i.saturate, instr.op_32.generates_cc);
break; break;
case OpCode::Id::LOP32I: { case OpCode::Id::LOP32I: {
UNIMPLEMENTED_IF_MSG(instr.op_32.generates_cc,
"Condition codes generation in LOP32I is not implemented");
if (instr.alu.lop32i.invert_a) if (instr.alu.lop32i.invert_a)
op_a = "~(" + op_a + ')'; op_a = "~(" + op_a + ')';
@ -2148,7 +2161,7 @@ private:
WriteLogicOperation(instr.gpr0, instr.alu.lop32i.operation, op_a, op_b, WriteLogicOperation(instr.gpr0, instr.alu.lop32i.operation, op_a, op_b,
Tegra::Shader::PredicateResultMode::None, Tegra::Shader::PredicateResultMode::None,
Tegra::Shader::Pred::UnusedIndex); Tegra::Shader::Pred::UnusedIndex, instr.op_32.generates_cc);
break; break;
} }
default: { default: {
@ -2177,7 +2190,7 @@ private:
case OpCode::Id::IADD_R: case OpCode::Id::IADD_R:
case OpCode::Id::IADD_IMM: { case OpCode::Id::IADD_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in IADD is not implemented"); "Condition codes generation in IADD is partially implemented");
if (instr.alu_integer.negate_a) if (instr.alu_integer.negate_a)
op_a = "-(" + op_a + ')'; op_a = "-(" + op_a + ')';
@ -2186,14 +2199,15 @@ private:
op_b = "-(" + op_b + ')'; op_b = "-(" + op_b + ')';
regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " + " + op_b, 1, 1, regs.SetRegisterToInteger(instr.gpr0, true, 0, op_a + " + " + op_b, 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, instr.generates_cc);
break; break;
} }
case OpCode::Id::IADD3_C: case OpCode::Id::IADD3_C:
case OpCode::Id::IADD3_R: case OpCode::Id::IADD3_R:
case OpCode::Id::IADD3_IMM: { case OpCode::Id::IADD3_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in IADD3 is not implemented"); instr.generates_cc,
"Condition codes generation in IADD3 is partially implemented");
std::string op_c = regs.GetRegisterAsInteger(instr.gpr39); std::string op_c = regs.GetRegisterAsInteger(instr.gpr39);
@ -2249,14 +2263,16 @@ private:
result = '(' + op_a + " + " + op_b + " + " + op_c + ')'; result = '(' + op_a + " + " + op_b + " + " + op_c + ')';
} }
regs.SetRegisterToInteger(instr.gpr0, true, 0, result, 1, 1); regs.SetRegisterToInteger(instr.gpr0, true, 0, result, 1, 1, false,
instr.generates_cc);
break; break;
} }
case OpCode::Id::ISCADD_C: case OpCode::Id::ISCADD_C:
case OpCode::Id::ISCADD_R: case OpCode::Id::ISCADD_R:
case OpCode::Id::ISCADD_IMM: { case OpCode::Id::ISCADD_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in ISCADD is not implemented"); instr.generates_cc,
"Condition codes generation in ISCADD is partially implemented");
if (instr.alu_integer.negate_a) if (instr.alu_integer.negate_a)
op_a = "-(" + op_a + ')'; op_a = "-(" + op_a + ')';
@ -2267,7 +2283,8 @@ private:
const std::string shift = std::to_string(instr.alu_integer.shift_amount.Value()); const std::string shift = std::to_string(instr.alu_integer.shift_amount.Value());
regs.SetRegisterToInteger(instr.gpr0, true, 0, regs.SetRegisterToInteger(instr.gpr0, true, 0,
"((" + op_a + " << " + shift + ") + " + op_b + ')', 1, 1); "((" + op_a + " << " + shift + ") + " + op_b + ')', 1, 1,
false, instr.generates_cc);
break; break;
} }
case OpCode::Id::POPC_C: case OpCode::Id::POPC_C:
@ -2291,8 +2308,6 @@ private:
case OpCode::Id::LOP_C: case OpCode::Id::LOP_C:
case OpCode::Id::LOP_R: case OpCode::Id::LOP_R:
case OpCode::Id::LOP_IMM: { case OpCode::Id::LOP_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in LOP is not implemented");
if (instr.alu.lop.invert_a) if (instr.alu.lop.invert_a)
op_a = "~(" + op_a + ')'; op_a = "~(" + op_a + ')';
@ -2301,15 +2316,13 @@ private:
op_b = "~(" + op_b + ')'; op_b = "~(" + op_b + ')';
WriteLogicOperation(instr.gpr0, instr.alu.lop.operation, op_a, op_b, WriteLogicOperation(instr.gpr0, instr.alu.lop.operation, op_a, op_b,
instr.alu.lop.pred_result_mode, instr.alu.lop.pred48); instr.alu.lop.pred_result_mode, instr.alu.lop.pred48,
instr.generates_cc);
break; break;
} }
case OpCode::Id::LOP3_C: case OpCode::Id::LOP3_C:
case OpCode::Id::LOP3_R: case OpCode::Id::LOP3_R:
case OpCode::Id::LOP3_IMM: { case OpCode::Id::LOP3_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in LOP3 is not implemented");
const std::string op_c = regs.GetRegisterAsInteger(instr.gpr39); const std::string op_c = regs.GetRegisterAsInteger(instr.gpr39);
std::string lut; std::string lut;
@ -2319,15 +2332,16 @@ private:
lut = '(' + std::to_string(instr.alu.lop3.GetImmLut48()) + ')'; lut = '(' + std::to_string(instr.alu.lop3.GetImmLut48()) + ')';
} }
WriteLop3Instruction(instr.gpr0, op_a, op_b, op_c, lut); WriteLop3Instruction(instr.gpr0, op_a, op_b, op_c, lut, instr.generates_cc);
break; break;
} }
case OpCode::Id::IMNMX_C: case OpCode::Id::IMNMX_C:
case OpCode::Id::IMNMX_R: case OpCode::Id::IMNMX_R:
case OpCode::Id::IMNMX_IMM: { case OpCode::Id::IMNMX_IMM: {
UNIMPLEMENTED_IF(instr.imnmx.exchange != Tegra::Shader::IMinMaxExchange::None); UNIMPLEMENTED_IF(instr.imnmx.exchange != Tegra::Shader::IMinMaxExchange::None);
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(
"Condition codes generation in IMNMX is not implemented"); instr.generates_cc,
"Condition codes generation in IMNMX is partially implemented");
const std::string condition = const std::string condition =
GetPredicateCondition(instr.imnmx.pred, instr.imnmx.negate_pred != 0); GetPredicateCondition(instr.imnmx.pred, instr.imnmx.negate_pred != 0);
@ -2335,7 +2349,7 @@ private:
regs.SetRegisterToInteger(instr.gpr0, instr.imnmx.is_signed, 0, regs.SetRegisterToInteger(instr.gpr0, instr.imnmx.is_signed, 0,
'(' + condition + ") ? min(" + parameters + ") : max(" + '(' + condition + ") ? min(" + parameters + ") : max(" +
parameters + ')', parameters + ')',
1, 1); 1, 1, false, instr.generates_cc);
break; break;
} }
case OpCode::Id::LEA_R2: case OpCode::Id::LEA_R2:
@ -2396,7 +2410,8 @@ private:
UNIMPLEMENTED_IF_MSG(instr.lea.pred48 != static_cast<u64>(Pred::UnusedIndex), UNIMPLEMENTED_IF_MSG(instr.lea.pred48 != static_cast<u64>(Pred::UnusedIndex),
"Unhandled LEA Predicate"); "Unhandled LEA Predicate");
const std::string value = '(' + op_a + " + (" + op_b + "*(1 << " + op_c + ")))"; const std::string value = '(' + op_a + " + (" + op_b + "*(1 << " + op_c + ")))";
regs.SetRegisterToInteger(instr.gpr0, true, 0, value, 1, 1); regs.SetRegisterToInteger(instr.gpr0, true, 0, value, 1, 1, false,
instr.generates_cc);
break; break;
} }
@ -2501,7 +2516,7 @@ private:
UNIMPLEMENTED_IF_MSG(instr.ffma.tab5980_1 != 0, "FFMA tab5980_1({}) not implemented", UNIMPLEMENTED_IF_MSG(instr.ffma.tab5980_1 != 0, "FFMA tab5980_1({}) not implemented",
instr.ffma.tab5980_1.Value()); instr.ffma.tab5980_1.Value());
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in FFMA is not implemented"); "Condition codes generation in FFMA is partially implemented");
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::FFMA_CR: { case OpCode::Id::FFMA_CR: {
@ -2532,7 +2547,7 @@ private:
} }
regs.SetRegisterToFloat(instr.gpr0, 0, "fma(" + op_a + ", " + op_b + ", " + op_c + ')', regs.SetRegisterToFloat(instr.gpr0, 0, "fma(" + op_a + ", " + op_b + ", " + op_c + ')',
1, 1, instr.alu.saturate_d, 0, true); 1, 1, instr.alu.saturate_d, instr.generates_cc, 0, true);
break; break;
} }
case OpCode::Type::Hfma2: { case OpCode::Type::Hfma2: {
@ -2603,16 +2618,14 @@ private:
} }
regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_output_signed, 0, op_a, 1, regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_output_signed, 0, op_a, 1,
1, instr.alu.saturate_d, 0, instr.conversion.dest_size, 1, instr.alu.saturate_d, instr.generates_cc, 0,
instr.generates_cc.Value() != 0); instr.conversion.dest_size);
break; break;
} }
case OpCode::Id::I2F_R: case OpCode::Id::I2F_R:
case OpCode::Id::I2F_C: { case OpCode::Id::I2F_C: {
UNIMPLEMENTED_IF(instr.conversion.dest_size != Register::Size::Word); UNIMPLEMENTED_IF(instr.conversion.dest_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.selector); UNIMPLEMENTED_IF(instr.conversion.selector);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in I2F is not implemented");
std::string op_a; std::string op_a;
if (instr.is_b_gpr) { if (instr.is_b_gpr) {
@ -2635,14 +2648,12 @@ private:
op_a = "-(" + op_a + ')'; op_a = "-(" + op_a + ')';
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1); regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1, false, instr.generates_cc);
break; break;
} }
case OpCode::Id::F2F_R: { case OpCode::Id::F2F_R: {
UNIMPLEMENTED_IF(instr.conversion.dest_size != Register::Size::Word); UNIMPLEMENTED_IF(instr.conversion.dest_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.src_size != Register::Size::Word); UNIMPLEMENTED_IF(instr.conversion.src_size != Register::Size::Word);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in F2F is not implemented");
std::string op_a = regs.GetRegisterAsFloat(instr.gpr20); std::string op_a = regs.GetRegisterAsFloat(instr.gpr20);
if (instr.conversion.abs_a) { if (instr.conversion.abs_a) {
@ -2674,14 +2685,13 @@ private:
break; break;
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1, instr.alu.saturate_d); regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1, instr.alu.saturate_d,
instr.generates_cc);
break; break;
} }
case OpCode::Id::F2I_R: case OpCode::Id::F2I_R:
case OpCode::Id::F2I_C: { case OpCode::Id::F2I_C: {
UNIMPLEMENTED_IF(instr.conversion.src_size != Register::Size::Word); UNIMPLEMENTED_IF(instr.conversion.src_size != Register::Size::Word);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in F2I is not implemented");
std::string op_a{}; std::string op_a{};
if (instr.is_b_gpr) { if (instr.is_b_gpr) {
@ -2724,7 +2734,8 @@ private:
} }
regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_output_signed, 0, op_a, 1, regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_output_signed, 0, op_a, 1,
1, false, 0, instr.conversion.dest_size); 1, false, instr.generates_cc, 0,
instr.conversion.dest_size);
break; break;
} }
default: { default: {
@ -2887,7 +2898,7 @@ private:
shader.AddLine(coord); shader.AddLine(coord);
if (depth_compare) { if (depth_compare) {
regs.SetRegisterToFloat(instr.gpr0, 0, texture, 1, 1, false); regs.SetRegisterToFloat(instr.gpr0, 0, texture, 1, 1);
} else { } else {
shader.AddLine("vec4 texture_tmp = " + texture + ';'); shader.AddLine("vec4 texture_tmp = " + texture + ';');
std::size_t dest_elem{}; std::size_t dest_elem{};
@ -2896,7 +2907,7 @@ private:
// Skip disabled components // Skip disabled components
continue; continue;
} }
regs.SetRegisterToFloat(instr.gpr0, elem, "texture_tmp", 1, 4, false, regs.SetRegisterToFloat(instr.gpr0, elem, "texture_tmp", 1, 4, false, false,
dest_elem); dest_elem);
++dest_elem; ++dest_elem;
} }
@ -2982,7 +2993,7 @@ private:
// Skip disabled components // Skip disabled components
continue; continue;
} }
regs.SetRegisterToFloat(instr.gpr0, elem, "texture_tmp", 1, 4, false, regs.SetRegisterToFloat(instr.gpr0, elem, "texture_tmp", 1, 4, false, false,
dest_elem); dest_elem);
++dest_elem; ++dest_elem;
} }
@ -3231,7 +3242,7 @@ private:
} }
case OpCode::Type::PredicateSetRegister: { case OpCode::Type::PredicateSetRegister: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in PSET is not implemented"); "Condition codes generation in PSET is partially implemented");
const std::string op_a = const std::string op_a =
GetPredicateCondition(instr.pset.pred12, instr.pset.neg_pred12 != 0); GetPredicateCondition(instr.pset.pred12, instr.pset.neg_pred12 != 0);
@ -3248,10 +3259,11 @@ private:
const std::string result = '(' + predicate + ") " + combiner + " (" + second_pred + ')'; const std::string result = '(' + predicate + ") " + combiner + " (" + second_pred + ')';
if (instr.pset.bf == 0) { if (instr.pset.bf == 0) {
const std::string value = '(' + result + ") ? 0xFFFFFFFF : 0"; const std::string value = '(' + result + ") ? 0xFFFFFFFF : 0";
regs.SetRegisterToInteger(instr.gpr0, false, 0, value, 1, 1); regs.SetRegisterToInteger(instr.gpr0, false, 0, value, 1, 1, false,
instr.generates_cc);
} else { } else {
const std::string value = '(' + result + ") ? 1.0 : 0.0"; const std::string value = '(' + result + ") ? 1.0 : 0.0";
regs.SetRegisterToFloat(instr.gpr0, 0, value, 1, 1); regs.SetRegisterToFloat(instr.gpr0, 0, value, 1, 1, false, instr.generates_cc);
} }
break; break;
} }
@ -3368,14 +3380,11 @@ private:
") " + combiner + " (" + second_pred + "))"; ") " + combiner + " (" + second_pred + "))";
if (instr.fset.bf) { if (instr.fset.bf) {
regs.SetRegisterToFloat(instr.gpr0, 0, predicate + " ? 1.0 : 0.0", 1, 1); regs.SetRegisterToFloat(instr.gpr0, 0, predicate + " ? 1.0 : 0.0", 1, 1, false,
instr.generates_cc);
} else { } else {
regs.SetRegisterToInteger(instr.gpr0, false, 0, predicate + " ? 0xFFFFFFFF : 0", 1, regs.SetRegisterToInteger(instr.gpr0, false, 0, predicate + " ? 0xFFFFFFFF : 0", 1,
1); 1, false, instr.generates_cc);
}
if (instr.generates_cc.Value() != 0) {
regs.SetInternalFlag(InternalFlag::ZeroFlag, predicate);
LOG_WARNING(HW_GPU, "FSET Condition Code is incomplete");
} }
break; break;
} }
@ -3462,7 +3471,7 @@ private:
UNIMPLEMENTED_IF(instr.xmad.sign_a); UNIMPLEMENTED_IF(instr.xmad.sign_a);
UNIMPLEMENTED_IF(instr.xmad.sign_b); UNIMPLEMENTED_IF(instr.xmad.sign_b);
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in XMAD is not implemented"); "Condition codes generation in XMAD is partially implemented");
std::string op_a{regs.GetRegisterAsInteger(instr.gpr8, 0, instr.xmad.sign_a)}; std::string op_a{regs.GetRegisterAsInteger(instr.gpr8, 0, instr.xmad.sign_a)};
std::string op_b; std::string op_b;
@ -3548,7 +3557,8 @@ private:
sum = "((" + sum + " & 0xFFFF) | (" + src2 + "<< 16))"; sum = "((" + sum + " & 0xFFFF) | (" + src2 + "<< 16))";
} }
regs.SetRegisterToInteger(instr.gpr0, is_signed, 0, sum, 1, 1); regs.SetRegisterToInteger(instr.gpr0, is_signed, 0, sum, 1, 1, false,
instr.generates_cc);
break; break;
} }
default: { default: {
@ -3752,8 +3762,7 @@ private:
} }
regs.SetRegisterToInteger(instr.gpr0, result_signed, 1, result, 1, 1, regs.SetRegisterToInteger(instr.gpr0, result_signed, 1, result, 1, 1,
instr.vmad.saturate == 1, 0, Register::Size::Word, instr.vmad.saturate, instr.vmad.cc);
instr.vmad.cc);
break; break;
} }
case OpCode::Id::VSETP: { case OpCode::Id::VSETP: {