diff --git a/src/shader_recompiler/CMakeLists.txt b/src/shader_recompiler/CMakeLists.txt
index 27fc79e21b..e1f4276a15 100644
--- a/src/shader_recompiler/CMakeLists.txt
+++ b/src/shader_recompiler/CMakeLists.txt
@@ -32,6 +32,8 @@ add_executable(shader_recompiler
frontend/ir/opcodes.cpp
frontend/ir/opcodes.h
frontend/ir/opcodes.inc
+ frontend/ir/post_order.cpp
+ frontend/ir/post_order.h
frontend/ir/pred.h
frontend/ir/program.cpp
frontend/ir/program.h
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.h b/src/shader_recompiler/backend/spirv/emit_spirv.h
index 46ec7a1bb8..6b09757d12 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv.h
+++ b/src/shader_recompiler/backend/spirv/emit_spirv.h
@@ -159,10 +159,10 @@ private:
Id EmitWorkgroupId(EmitContext& ctx);
Id EmitLocalInvocationId(EmitContext& ctx);
Id EmitUndefU1(EmitContext& ctx);
- void EmitUndefU8(EmitContext& ctx);
- void EmitUndefU16(EmitContext& ctx);
- void EmitUndefU32(EmitContext& ctx);
- void EmitUndefU64(EmitContext& ctx);
+ Id EmitUndefU8(EmitContext& ctx);
+ Id EmitUndefU16(EmitContext& ctx);
+ Id EmitUndefU32(EmitContext& ctx);
+ Id EmitUndefU64(EmitContext& ctx);
void EmitLoadGlobalU8(EmitContext& ctx);
void EmitLoadGlobalS8(EmitContext& ctx);
void EmitLoadGlobalU16(EmitContext& ctx);
@@ -297,12 +297,12 @@ private:
void EmitBitFieldInsert(EmitContext& ctx);
void EmitBitFieldSExtract(EmitContext& ctx);
Id EmitBitFieldUExtract(EmitContext& ctx, Id base, Id offset, Id count);
- void EmitSLessThan(EmitContext& ctx);
+ Id EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs);
void EmitULessThan(EmitContext& ctx);
void EmitIEqual(EmitContext& ctx);
void EmitSLessThanEqual(EmitContext& ctx);
void EmitULessThanEqual(EmitContext& ctx);
- void EmitSGreaterThan(EmitContext& ctx);
+ Id EmitSGreaterThan(EmitContext& ctx, Id lhs, Id rhs);
void EmitUGreaterThan(EmitContext& ctx);
void EmitINotEqual(EmitContext& ctx);
void EmitSGreaterThanEqual(EmitContext& ctx);
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp
index 3ef4f3d787..e811a63ab3 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp
@@ -73,8 +73,8 @@ Id EmitSPIRV::EmitBitFieldUExtract(EmitContext& ctx, Id base, Id offset, Id coun
return ctx.OpBitFieldUExtract(ctx.u32[1], base, offset, count);
}
-void EmitSPIRV::EmitSLessThan(EmitContext&) {
- throw NotImplementedException("SPIR-V Instruction");
+Id EmitSPIRV::EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs) {
+ return ctx.OpSLessThan(ctx.u1, lhs, rhs);
}
void EmitSPIRV::EmitULessThan(EmitContext&) {
@@ -93,8 +93,8 @@ void EmitSPIRV::EmitULessThanEqual(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
-void EmitSPIRV::EmitSGreaterThan(EmitContext&) {
- throw NotImplementedException("SPIR-V Instruction");
+Id EmitSPIRV::EmitSGreaterThan(EmitContext& ctx, Id lhs, Id rhs) {
+ return ctx.OpSGreaterThan(ctx.u1, lhs, rhs);
}
void EmitSPIRV::EmitUGreaterThan(EmitContext&) {
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_undefined.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_undefined.cpp
index 859b60a95d..a6f542360c 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_undefined.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_undefined.cpp
@@ -10,19 +10,19 @@ Id EmitSPIRV::EmitUndefU1(EmitContext& ctx) {
return ctx.OpUndef(ctx.u1);
}
-void EmitSPIRV::EmitUndefU8(EmitContext&) {
+Id EmitSPIRV::EmitUndefU8(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
-void EmitSPIRV::EmitUndefU16(EmitContext&) {
+Id EmitSPIRV::EmitUndefU16(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
-void EmitSPIRV::EmitUndefU32(EmitContext&) {
- throw NotImplementedException("SPIR-V Instruction");
+Id EmitSPIRV::EmitUndefU32(EmitContext& ctx) {
+ return ctx.OpUndef(ctx.u32[1]);
}
-void EmitSPIRV::EmitUndefU64(EmitContext&) {
+Id EmitSPIRV::EmitUndefU64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
diff --git a/src/shader_recompiler/frontend/ir/function.h b/src/shader_recompiler/frontend/ir/function.h
index fd7d564191..d1f0611467 100644
--- a/src/shader_recompiler/frontend/ir/function.h
+++ b/src/shader_recompiler/frontend/ir/function.h
@@ -12,6 +12,7 @@ namespace Shader::IR {
struct Function {
BlockList blocks;
+ BlockList post_order_blocks;
};
} // namespace Shader::IR
diff --git a/src/shader_recompiler/frontend/ir/post_order.cpp b/src/shader_recompiler/frontend/ir/post_order.cpp
new file mode 100644
index 0000000000..a48b8dec5a
--- /dev/null
+++ b/src/shader_recompiler/frontend/ir/post_order.cpp
@@ -0,0 +1,48 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <boost/container/flat_set.hpp>
+#include <boost/container/small_vector.hpp>
+
+#include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/post_order.h"
+
+namespace Shader::IR {
+
+BlockList PostOrder(const BlockList& blocks) {
+ boost::container::small_vector<Block*, 16> block_stack;
+ boost::container::flat_set<Block*> visited;
+
+ BlockList post_order_blocks;
+ post_order_blocks.reserve(blocks.size());
+
+ Block* const first_block{blocks.front()};
+ visited.insert(first_block);
+ block_stack.push_back(first_block);
+
+ const auto visit_branch = [&](Block* block, Block* branch) {
+ if (!branch) {
+ return false;
+ }
+ if (!visited.insert(branch).second) {
+ return false;
+ }
+ // Calling push_back twice is faster than insert on msvc
+ block_stack.push_back(block);
+ block_stack.push_back(branch);
+ return true;
+ };
+ while (!block_stack.empty()) {
+ Block* const block{block_stack.back()};
+ block_stack.pop_back();
+
+ if (!visit_branch(block, block->TrueBranch()) &&
+ !visit_branch(block, block->FalseBranch())) {
+ post_order_blocks.push_back(block);
+ }
+ }
+ return post_order_blocks;
+}
+
+} // namespace Shader::IR
diff --git a/src/shader_recompiler/frontend/ir/post_order.h b/src/shader_recompiler/frontend/ir/post_order.h
new file mode 100644
index 0000000000..30137ff57a
--- /dev/null
+++ b/src/shader_recompiler/frontend/ir/post_order.h
@@ -0,0 +1,13 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "shader_recompiler/frontend/ir/basic_block.h"
+
+namespace Shader::IR {
+
+BlockList PostOrder(const BlockList& blocks);
+
+} // namespace Shader::IR
diff --git a/src/shader_recompiler/frontend/maxwell/program.cpp b/src/shader_recompiler/frontend/maxwell/program.cpp
index 9fa912ed8e..dab6d68c02 100644
--- a/src/shader_recompiler/frontend/maxwell/program.cpp
+++ b/src/shader_recompiler/frontend/maxwell/program.cpp
@@ -7,6 +7,7 @@
#include <vector>
#include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/post_order.h"
#include "shader_recompiler/frontend/ir/structured_control_flow.h"
#include "shader_recompiler/frontend/maxwell/program.h"
#include "shader_recompiler/frontend/maxwell/translate/translate.h"
@@ -56,11 +57,14 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
}
fmt::print(stdout, "No optimizations: {}", IR::DumpProgram(program));
- std::ranges::for_each(functions, Optimization::SsaRewritePass);
for (IR::Function& function : functions) {
- Optimization::Invoke(Optimization::GlobalMemoryToStorageBufferPass, function);
- Optimization::Invoke(Optimization::ConstantPropagationPass, function);
- Optimization::Invoke(Optimization::DeadCodeEliminationPass, function);
+ function.post_order_blocks = PostOrder(function.blocks);
+ Optimization::SsaRewritePass(function.post_order_blocks);
+ }
+ for (IR::Function& function : functions) {
+ Optimization::PostOrderInvoke(Optimization::GlobalMemoryToStorageBufferPass, function);
+ Optimization::PostOrderInvoke(Optimization::ConstantPropagationPass, function);
+ Optimization::PostOrderInvoke(Optimization::DeadCodeEliminationPass, function);
Optimization::IdentityRemovalPass(function);
Optimization::VerificationPass(function);
}
diff --git a/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp b/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp
index bbaa412f6e..132b2012a8 100644
--- a/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp
+++ b/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp
@@ -13,7 +13,7 @@ namespace Shader::Optimization {
void DeadCodeEliminationPass(IR::Block& block) {
// We iterate over the instructions in reverse order.
// This is because removing an instruction reduces the number of uses for earlier instructions.
- for (IR::Inst& inst : std::views::reverse(block)) {
+ for (IR::Inst& inst : block | std::views::reverse) {
if (!inst.HasUses() && !inst.MayHaveSideEffects()) {
inst.Invalidate();
}
diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h
index 578a24d89b..30eb315889 100644
--- a/src/shader_recompiler/ir_opt/passes.h
+++ b/src/shader_recompiler/ir_opt/passes.h
@@ -4,14 +4,16 @@
#pragma once
+#include <span>
+
#include "shader_recompiler/frontend/ir/basic_block.h"
#include "shader_recompiler/frontend/ir/function.h"
namespace Shader::Optimization {
template <typename Func>
-void Invoke(Func&& func, IR::Function& function) {
- for (const auto& block : function.blocks) {
+void PostOrderInvoke(Func&& func, IR::Function& function) {
+ for (const auto& block : function.post_order_blocks) {
func(*block);
}
}
@@ -20,7 +22,7 @@ void ConstantPropagationPass(IR::Block& block);
void DeadCodeEliminationPass(IR::Block& block);
void GlobalMemoryToStorageBufferPass(IR::Block& block);
void IdentityRemovalPass(IR::Function& function);
-void SsaRewritePass(IR::Function& function);
+void SsaRewritePass(std::span<IR::Block* const> post_order_blocks);
void VerificationPass(const IR::Function& function);
} // namespace Shader::Optimization
diff --git a/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp b/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp
index 7eaf719c4e..13f9c914a2 100644
--- a/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp
+++ b/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp
@@ -14,7 +14,13 @@
// https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6
//
+#include <ranges>
+#include <span>
+#include <variant>
+#include <vector>
+
#include <boost/container/flat_map.hpp>
+#include <boost/container/flat_set.hpp>
#include "shader_recompiler/frontend/ir/basic_block.h"
#include "shader_recompiler/frontend/ir/function.h"
@@ -26,9 +32,9 @@
namespace Shader::Optimization {
namespace {
-using ValueMap = boost::container::flat_map<IR::Block*, IR::Value, std::less<IR::Block*>>;
-
-struct FlagTag {};
+struct FlagTag {
+ auto operator<=>(const FlagTag&) const noexcept = default;
+};
struct ZeroFlagTag : FlagTag {};
struct SignFlagTag : FlagTag {};
struct CarryFlagTag : FlagTag {};
@@ -38,9 +44,15 @@ struct GotoVariable : FlagTag {
GotoVariable() = default;
explicit GotoVariable(u32 index_) : index{index_} {}
+ auto operator<=>(const GotoVariable&) const noexcept = default;
+
u32 index;
};
+using Variant = std::variant<IR::Reg, IR::Pred, ZeroFlagTag, SignFlagTag, CarryFlagTag,
+ OverflowFlagTag, GotoVariable>;
+using ValueMap = boost::container::flat_map<IR::Block*, IR::Value, std::less<IR::Block*>>;
+
struct DefTable {
[[nodiscard]] ValueMap& operator[](IR::Reg variable) noexcept {
return regs[IR::RegIndex(variable)];
@@ -102,19 +114,35 @@ public:
}
IR::Value ReadVariable(auto variable, IR::Block* block) {
- auto& def{current_def[variable]};
+ const ValueMap& def{current_def[variable]};
if (const auto it{def.find(block)}; it != def.end()) {
return it->second;
}
return ReadVariableRecursive(variable, block);
}
+ void SealBlock(IR::Block* block) {
+ const auto it{incomplete_phis.find(block)};
+ if (it != incomplete_phis.end()) {
+ for (auto& [variant, phi] : it->second) {
+ std::visit([&](auto& variable) { AddPhiOperands(variable, *phi, block); }, variant);
+ }
+ }
+ sealed_blocks.insert(block);
+ }
+
private:
IR::Value ReadVariableRecursive(auto variable, IR::Block* block) {
IR::Value val;
- if (const std::span preds{block->ImmediatePredecessors()}; preds.size() == 1) {
+ if (!sealed_blocks.contains(block)) {
+ // Incomplete CFG
+ IR::Inst* phi{&*block->PrependNewInst(block->begin(), IR::Opcode::Phi)};
+ incomplete_phis[block].insert_or_assign(variable, phi);
+ val = IR::Value{&*phi};
+ } else if (const std::span imm_preds{block->ImmediatePredecessors()};
+ imm_preds.size() == 1) {
// Optimize the common case of one predecessor: no phi needed
- val = ReadVariable(variable, preds.front());
+ val = ReadVariable(variable, imm_preds.front());
} else {
// Break potential cycles with operandless phi
IR::Inst& phi_inst{*block->PrependNewInst(block->begin(), IR::Opcode::Phi)};
@@ -127,8 +155,8 @@ private:
}
IR::Value AddPhiOperands(auto variable, IR::Inst& phi, IR::Block* block) {
- for (IR::Block* const pred : block->ImmediatePredecessors()) {
- phi.AddPhiOperand(pred, ReadVariable(variable, pred));
+ for (IR::Block* const imm_pred : block->ImmediatePredecessors()) {
+ phi.AddPhiOperand(imm_pred, ReadVariable(variable, imm_pred));
}
return TryRemoveTrivialPhi(phi, block, UndefOpcode(variable));
}
@@ -159,6 +187,9 @@ private:
return same;
}
+ boost::container::flat_set<IR::Block*> sealed_blocks;
+ boost::container::flat_map<IR::Block*, boost::container::flat_map<Variant, IR::Inst*>>
+ incomplete_phis;
DefTable current_def;
};
@@ -218,14 +249,19 @@ void VisitInst(Pass& pass, IR::Block* block, IR::Inst& inst) {
break;
}
}
+
+void VisitBlock(Pass& pass, IR::Block* block) {
+ for (IR::Inst& inst : block->Instructions()) {
+ VisitInst(pass, block, inst);
+ }
+ pass.SealBlock(block);
+}
} // Anonymous namespace
-void SsaRewritePass(IR::Function& function) {
+void SsaRewritePass(std::span<IR::Block* const> post_order_blocks) {
Pass pass;
- for (IR::Block* const block : function.blocks) {
- for (IR::Inst& inst : block->Instructions()) {
- VisitInst(pass, block, inst);
- }
+ for (IR::Block* const block : post_order_blocks | std::views::reverse) {
+ VisitBlock(pass, block);
}
}
diff --git a/src/shader_recompiler/main.cpp b/src/shader_recompiler/main.cpp
index 92358232c3..29f65966c3 100644
--- a/src/shader_recompiler/main.cpp
+++ b/src/shader_recompiler/main.cpp
@@ -69,14 +69,12 @@ int main() {
// FileEnvironment env{"D:\\Shaders\\Database\\Oninaki\\CS8F146B41DB6BD826.bin"};
FileEnvironment env{"D:\\Shaders\\shader.bin"};
- for (int i = 0; i < 1; ++i) {
- block_pool->ReleaseContents();
- inst_pool->ReleaseContents();
- flow_block_pool->ReleaseContents();
- Flow::CFG cfg{env, *flow_block_pool, 0};
- fmt::print(stdout, "{}\n", cfg.Dot());
- IR::Program program{TranslateProgram(*inst_pool, *block_pool, env, cfg)};
- fmt::print(stdout, "{}\n", IR::DumpProgram(program));
- Backend::SPIRV::EmitSPIRV spirv{program};
- }
+ block_pool->ReleaseContents();
+ inst_pool->ReleaseContents();
+ flow_block_pool->ReleaseContents();
+ Flow::CFG cfg{env, *flow_block_pool, 0};
+ fmt::print(stdout, "{}\n", cfg.Dot());
+ IR::Program program{TranslateProgram(*inst_pool, *block_pool, env, cfg)};
+ fmt::print(stdout, "{}\n", IR::DumpProgram(program));
+ // Backend::SPIRV::EmitSPIRV spirv{program};
}