shader: Support SSA loops on IR

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

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);

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&) {

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");
 }
 

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

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

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

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);
     }

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();
         }

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

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);
     }
 }
 

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};
 }