summaryrefslogtreecommitdiffstats
path: root/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
blob: a39db2bf12a7d3454fa58cd4ee07b83d9608c7b7 (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.

#include <algorithm>
#include <tuple>
#include <type_traits>

#include "common/bit_cast.h"
#include "common/bit_util.h"
#include "shader_recompiler/exception.h"
#include "shader_recompiler/frontend/ir/ir_emitter.h"
#include "shader_recompiler/frontend/ir/microinstruction.h"
#include "shader_recompiler/ir_opt/passes.h"

namespace Shader::Optimization {
namespace {
// Metaprogramming stuff to get arguments information out of a lambda
template <typename Func>
struct LambdaTraits : LambdaTraits<decltype(&std::remove_reference_t<Func>::operator())> {};

template <typename ReturnType, typename LambdaType, typename... Args>
struct LambdaTraits<ReturnType (LambdaType::*)(Args...) const> {
    template <size_t I>
    using ArgType = std::tuple_element_t<I, std::tuple<Args...>>;

    static constexpr size_t NUM_ARGS{sizeof...(Args)};
};

template <typename T>
[[nodiscard]] T Arg(const IR::Value& value) {
    if constexpr (std::is_same_v<T, bool>) {
        return value.U1();
    } else if constexpr (std::is_same_v<T, u32>) {
        return value.U32();
    } else if constexpr (std::is_same_v<T, s32>) {
        return static_cast<s32>(value.U32());
    } else if constexpr (std::is_same_v<T, f32>) {
        return value.F32();
    } else if constexpr (std::is_same_v<T, u64>) {
        return value.U64();
    }
}

template <typename T, typename ImmFn>
bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
    const IR::Value lhs{inst.Arg(0)};
    const IR::Value rhs{inst.Arg(1)};

    const bool is_lhs_immediate{lhs.IsImmediate()};
    const bool is_rhs_immediate{rhs.IsImmediate()};

    if (is_lhs_immediate && is_rhs_immediate) {
        const auto result{imm_fn(Arg<T>(lhs), Arg<T>(rhs))};
        inst.ReplaceUsesWith(IR::Value{result});
        return false;
    }
    if (is_lhs_immediate && !is_rhs_immediate) {
        IR::Inst* const rhs_inst{rhs.InstRecursive()};
        if (rhs_inst->Opcode() == inst.Opcode() && rhs_inst->Arg(1).IsImmediate()) {
            const auto combined{imm_fn(Arg<T>(lhs), Arg<T>(rhs_inst->Arg(1)))};
            inst.SetArg(0, rhs_inst->Arg(0));
            inst.SetArg(1, IR::Value{combined});
        } else {
            // Normalize
            inst.SetArg(0, rhs);
            inst.SetArg(1, lhs);
        }
    }
    if (!is_lhs_immediate && is_rhs_immediate) {
        const IR::Inst* const lhs_inst{lhs.InstRecursive()};
        if (lhs_inst->Opcode() == inst.Opcode() && lhs_inst->Arg(1).IsImmediate()) {
            const auto combined{imm_fn(Arg<T>(rhs), Arg<T>(lhs_inst->Arg(1)))};
            inst.SetArg(0, lhs_inst->Arg(0));
            inst.SetArg(1, IR::Value{combined});
        }
    }
    return true;
}

template <typename Func>
bool FoldWhenAllImmediates(IR::Inst& inst, Func&& func) {
    if (!inst.AreAllArgsImmediates() || inst.HasAssociatedPseudoOperation()) {
        return false;
    }
    using Indices = std::make_index_sequence<LambdaTraits<decltype(func)>::NUM_ARGS>;
    inst.ReplaceUsesWith(EvalImmediates(inst, func, Indices{}));
    return true;
}

void FoldGetRegister(IR::Inst& inst) {
    if (inst.Arg(0).Reg() == IR::Reg::RZ) {
        inst.ReplaceUsesWith(IR::Value{u32{0}});
    }
}

void FoldGetPred(IR::Inst& inst) {
    if (inst.Arg(0).Pred() == IR::Pred::PT) {
        inst.ReplaceUsesWith(IR::Value{true});
    }
}

/// Replaces the pattern generated by two XMAD multiplications
bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
    /*
     * We are looking for this pattern:
     *   %rhs_bfe = BitFieldUExtract %factor_a, #0, #16
     *   %rhs_mul = IMul32 %rhs_bfe, %factor_b
     *   %lhs_bfe = BitFieldUExtract %factor_a, #16, #16
     *   %rhs_mul = IMul32 %lhs_bfe, %factor_b
     *   %lhs_shl = ShiftLeftLogical32 %rhs_mul, #16
     *   %result  = IAdd32 %lhs_shl, %rhs_mul
     *
     * And replacing it with
     *   %result  = IMul32 %factor_a, %factor_b
     *
     * This optimization has been proven safe by LLVM and MSVC.
     */
    const IR::Value lhs_arg{inst.Arg(0)};
    const IR::Value rhs_arg{inst.Arg(1)};
    if (lhs_arg.IsImmediate() || rhs_arg.IsImmediate()) {
        return false;
    }
    IR::Inst* const lhs_shl{lhs_arg.InstRecursive()};
    if (lhs_shl->Opcode() != IR::Opcode::ShiftLeftLogical32 || lhs_shl->Arg(1) != IR::Value{16U}) {
        return false;
    }
    if (lhs_shl->Arg(0).IsImmediate()) {
        return false;
    }
    IR::Inst* const lhs_mul{lhs_shl->Arg(0).InstRecursive()};
    IR::Inst* const rhs_mul{rhs_arg.InstRecursive()};
    if (lhs_mul->Opcode() != IR::Opcode::IMul32 || rhs_mul->Opcode() != IR::Opcode::IMul32) {
        return false;
    }
    if (lhs_mul->Arg(1).Resolve() != rhs_mul->Arg(1).Resolve()) {
        return false;
    }
    const IR::U32 factor_b{lhs_mul->Arg(1)};
    if (lhs_mul->Arg(0).IsImmediate() || rhs_mul->Arg(0).IsImmediate()) {
        return false;
    }
    IR::Inst* const lhs_bfe{lhs_mul->Arg(0).InstRecursive()};
    IR::Inst* const rhs_bfe{rhs_mul->Arg(0).InstRecursive()};
    if (lhs_bfe->Opcode() != IR::Opcode::BitFieldUExtract) {
        return false;
    }
    if (rhs_bfe->Opcode() != IR::Opcode::BitFieldUExtract) {
        return false;
    }
    if (lhs_bfe->Arg(1) != IR::Value{16U} || lhs_bfe->Arg(2) != IR::Value{16U}) {
        return false;
    }
    if (rhs_bfe->Arg(1) != IR::Value{0U} || rhs_bfe->Arg(2) != IR::Value{16U}) {
        return false;
    }
    if (lhs_bfe->Arg(0).Resolve() != rhs_bfe->Arg(0).Resolve()) {
        return false;
    }
    const IR::U32 factor_a{lhs_bfe->Arg(0)};
    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
    inst.ReplaceUsesWith(ir.IMul(factor_a, factor_b));
    return true;
}

template <typename T>
void FoldAdd(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        return;
    }
    if (!FoldCommutative<T>(inst, [](T a, T b) { return a + b; })) {
        return;
    }
    const IR::Value rhs{inst.Arg(1)};
    if (rhs.IsImmediate() && Arg<T>(rhs) == 0) {
        inst.ReplaceUsesWith(inst.Arg(0));
        return;
    }
    if constexpr (std::is_same_v<T, u32>) {
        if (FoldXmadMultiply(block, inst)) {
            return;
        }
    }
}

void FoldISub32(IR::Inst& inst) {
    if (FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a - b; })) {
        return;
    }
    if (inst.Arg(0).IsImmediate() || inst.Arg(1).IsImmediate()) {
        return;
    }
    // ISub32 is generally used to subtract two constant buffers, compare and replace this with
    // zero if they equal.
    const auto equal_cbuf{[](IR::Inst* a, IR::Inst* b) {
        return a->Opcode() == IR::Opcode::GetCbufU32 && b->Opcode() == IR::Opcode::GetCbufU32 &&
               a->Arg(0) == b->Arg(0) && a->Arg(1) == b->Arg(1);
    }};
    IR::Inst* op_a{inst.Arg(0).InstRecursive()};
    IR::Inst* op_b{inst.Arg(1).InstRecursive()};
    if (equal_cbuf(op_a, op_b)) {
        inst.ReplaceUsesWith(IR::Value{u32{0}});
        return;
    }
    // It's also possible a value is being added to a cbuf and then subtracted
    if (op_b->Opcode() == IR::Opcode::IAdd32) {
        // Canonicalize local variables to simplify the following logic
        std::swap(op_a, op_b);
    }
    if (op_b->Opcode() != IR::Opcode::GetCbufU32) {
        return;
    }
    IR::Inst* const inst_cbuf{op_b};
    if (op_a->Opcode() != IR::Opcode::IAdd32) {
        return;
    }
    IR::Value add_op_a{op_a->Arg(0)};
    IR::Value add_op_b{op_a->Arg(1)};
    if (add_op_b.IsImmediate()) {
        // Canonicalize
        std::swap(add_op_a, add_op_b);
    }
    if (add_op_b.IsImmediate()) {
        return;
    }
    IR::Inst* const add_cbuf{add_op_b.InstRecursive()};
    if (equal_cbuf(add_cbuf, inst_cbuf)) {
        inst.ReplaceUsesWith(add_op_a);
    }
}

template <typename T>
void FoldSelect(IR::Inst& inst) {
    const IR::Value cond{inst.Arg(0)};
    if (cond.IsImmediate()) {
        inst.ReplaceUsesWith(cond.U1() ? inst.Arg(1) : inst.Arg(2));
    }
}

void FoldLogicalAnd(IR::Inst& inst) {
    if (!FoldCommutative<bool>(inst, [](bool a, bool b) { return a && b; })) {
        return;
    }
    const IR::Value rhs{inst.Arg(1)};
    if (rhs.IsImmediate()) {
        if (rhs.U1()) {
            inst.ReplaceUsesWith(inst.Arg(0));
        } else {
            inst.ReplaceUsesWith(IR::Value{false});
        }
    }
}

void FoldLogicalOr(IR::Inst& inst) {
    if (!FoldCommutative<bool>(inst, [](bool a, bool b) { return a || b; })) {
        return;
    }
    const IR::Value rhs{inst.Arg(1)};
    if (rhs.IsImmediate()) {
        if (rhs.U1()) {
            inst.ReplaceUsesWith(IR::Value{true});
        } else {
            inst.ReplaceUsesWith(inst.Arg(0));
        }
    }
}

void FoldLogicalNot(IR::Inst& inst) {
    const IR::U1 value{inst.Arg(0)};
    if (value.IsImmediate()) {
        inst.ReplaceUsesWith(IR::Value{!value.U1()});
        return;
    }
    IR::Inst* const arg{value.InstRecursive()};
    if (arg->Opcode() == IR::Opcode::LogicalNot) {
        inst.ReplaceUsesWith(arg->Arg(0));
    }
}

template <IR::Opcode op, typename Dest, typename Source>
void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
    const IR::Value value{inst.Arg(0)};
    if (value.IsImmediate()) {
        inst.ReplaceUsesWith(IR::Value{Common::BitCast<Dest>(Arg<Source>(value))});
        return;
    }
    IR::Inst* const arg_inst{value.InstRecursive()};
    if (arg_inst->Opcode() == reverse) {
        inst.ReplaceUsesWith(arg_inst->Arg(0));
        return;
    }
    if constexpr (op == IR::Opcode::BitCastF32U32) {
        if (arg_inst->Opcode() == IR::Opcode::GetCbufU32) {
            // Replace the bitcast with a typed constant buffer read
            inst.ReplaceOpcode(IR::Opcode::GetCbufF32);
            inst.SetArg(0, arg_inst->Arg(0));
            inst.SetArg(1, arg_inst->Arg(1));
            return;
        }
    }
}

template <typename Func, size_t... I>
IR::Value EvalImmediates(const IR::Inst& inst, Func&& func, std::index_sequence<I...>) {
    using Traits = LambdaTraits<decltype(func)>;
    return IR::Value{func(Arg<Traits::ArgType<I>>(inst.Arg(I))...)};
}

void FoldBranchConditional(IR::Inst& inst) {
    const IR::U1 cond{inst.Arg(0)};
    if (cond.IsImmediate()) {
        // TODO: Convert to Branch
        return;
    }
    const IR::Inst* cond_inst{cond.InstRecursive()};
    if (cond_inst->Opcode() == IR::Opcode::LogicalNot) {
        const IR::Value true_label{inst.Arg(1)};
        const IR::Value false_label{inst.Arg(2)};
        // Remove negation on the conditional (take the parameter out of LogicalNot) and swap
        // the branches
        inst.SetArg(0, cond_inst->Arg(0));
        inst.SetArg(1, false_label);
        inst.SetArg(2, true_label);
    }
}

void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
    switch (inst.Opcode()) {
    case IR::Opcode::GetRegister:
        return FoldGetRegister(inst);
    case IR::Opcode::GetPred:
        return FoldGetPred(inst);
    case IR::Opcode::IAdd32:
        return FoldAdd<u32>(block, inst);
    case IR::Opcode::ISub32:
        return FoldISub32(inst);
    case IR::Opcode::BitCastF32U32:
        return FoldBitCast<IR::Opcode::BitCastF32U32, f32, u32>(inst, IR::Opcode::BitCastU32F32);
    case IR::Opcode::BitCastU32F32:
        return FoldBitCast<IR::Opcode::BitCastU32F32, u32, f32>(inst, IR::Opcode::BitCastF32U32);
    case IR::Opcode::IAdd64:
        return FoldAdd<u64>(block, inst);
    case IR::Opcode::SelectU32:
        return FoldSelect<u32>(inst);
    case IR::Opcode::LogicalAnd:
        return FoldLogicalAnd(inst);
    case IR::Opcode::LogicalOr:
        return FoldLogicalOr(inst);
    case IR::Opcode::LogicalNot:
        return FoldLogicalNot(inst);
    case IR::Opcode::SLessThan:
        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a < b; });
        return;
    case IR::Opcode::ULessThan:
        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
        return;
    case IR::Opcode::BitFieldUExtract:
        FoldWhenAllImmediates(inst, [](u32 base, u32 shift, u32 count) {
            if (static_cast<size_t>(shift) + static_cast<size_t>(count) > Common::BitSize<u32>()) {
                throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldUExtract,
                                 base, shift, count);
            }
            return (base >> shift) & ((1U << count) - 1);
        });
        return;
    case IR::Opcode::BranchConditional:
        return FoldBranchConditional(inst);
    default:
        break;
    }
}
} // Anonymous namespace

void ConstantPropagationPass(IR::Program& program) {
    for (IR::Block* const block : program.post_order_blocks) {
        for (IR::Inst& inst : block->Instructions()) {
            ConstantPropagation(*block, inst);
        }
    }
}

} // namespace Shader::Optimization