diff options
Diffstat (limited to '')
8 files changed, 164 insertions, 59 deletions
diff --git a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp index bfd2ae650..16278faab 100644 --- a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp +++ b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp @@ -29,6 +29,41 @@ void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) { }); } +void AddRegisterIndexedLdc(Info& info) { + info.uses_cbuf_indirect = true; + + for (u32 i = 0; i < Info::MAX_INDIRECT_CBUFS; i++) { + AddConstantBufferDescriptor(info, i, 1); + + // The shader can use any possible access size + info.constant_buffer_used_sizes[i] = 0x10'000; + } +} + +u32 GetElementSize(IR::Type& used_type, Shader::IR::Opcode opcode) { + switch (opcode) { + case IR::Opcode::GetCbufU8: + case IR::Opcode::GetCbufS8: + used_type |= IR::Type::U8; + return 1; + case IR::Opcode::GetCbufU16: + case IR::Opcode::GetCbufS16: + used_type |= IR::Type::U16; + return 2; + case IR::Opcode::GetCbufU32: + used_type |= IR::Type::U32; + return 4; + case IR::Opcode::GetCbufF32: + used_type |= IR::Type::F32; + return 4; + case IR::Opcode::GetCbufU32x2: + used_type |= IR::Type::U32x2; + return 8; + default: + throw InvalidArgument("Invalid opcode {}", opcode); + } +} + void GetPatch(Info& info, IR::Patch patch) { if (!IR::IsGeneric(patch)) { throw NotImplementedException("Reading non-generic patch {}", patch); @@ -463,42 +498,18 @@ void VisitUsages(Info& info, IR::Inst& inst) { case IR::Opcode::GetCbufU32x2: { const IR::Value index{inst.Arg(0)}; const IR::Value offset{inst.Arg(1)}; - if (!index.IsImmediate()) { - throw NotImplementedException("Constant buffer with non-immediate index"); - } - AddConstantBufferDescriptor(info, index.U32(), 1); - u32 element_size{}; - switch (inst.GetOpcode()) { - case IR::Opcode::GetCbufU8: - case IR::Opcode::GetCbufS8: - info.used_constant_buffer_types |= IR::Type::U8; - element_size = 1; - break; - case IR::Opcode::GetCbufU16: - case IR::Opcode::GetCbufS16: - info.used_constant_buffer_types |= IR::Type::U16; - element_size = 2; - break; - case IR::Opcode::GetCbufU32: - info.used_constant_buffer_types |= IR::Type::U32; - element_size = 4; - break; - case IR::Opcode::GetCbufF32: - info.used_constant_buffer_types |= IR::Type::F32; - element_size = 4; - break; - case IR::Opcode::GetCbufU32x2: - info.used_constant_buffer_types |= IR::Type::U32x2; - element_size = 8; - break; - default: - break; - } - u32& size{info.constant_buffer_used_sizes[index.U32()]}; - if (offset.IsImmediate()) { - size = Common::AlignUp(std::max(size, offset.U32() + element_size), 16u); + if (index.IsImmediate()) { + AddConstantBufferDescriptor(info, index.U32(), 1); + u32 element_size = GetElementSize(info.used_constant_buffer_types, inst.GetOpcode()); + u32& size{info.constant_buffer_used_sizes[index.U32()]}; + if (offset.IsImmediate()) { + size = Common::AlignUp(std::max(size, offset.U32() + element_size), 16u); + } else { + size = 0x10'000; + } } else { - size = 0x10'000; + AddRegisterIndexedLdc(info); + GetElementSize(info.used_indirect_cbuf_types, inst.GetOpcode()); } break; } diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp index c134a12bc..2a14e7f12 100644 --- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp +++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp @@ -8,7 +8,6 @@ #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/value.h" 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 400836301..6697fde85 100644 --- a/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp +++ b/src/shader_recompiler/ir_opt/dead_code_elimination_pass.cpp @@ -2,24 +2,104 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. +#include <algorithm> + +#include <boost/container/small_vector.hpp> + #include "shader_recompiler/frontend/ir/basic_block.h" #include "shader_recompiler/frontend/ir/value.h" #include "shader_recompiler/ir_opt/passes.h" namespace Shader::Optimization { - -void DeadCodeEliminationPass(IR::Program& program) { +namespace { +template <bool TEST_USES> +void DeadInstElimination(IR::Block* const 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::Block* const block : program.post_order_blocks) { - auto it{block->end()}; - while (it != block->begin()) { - --it; - if (!it->HasUses() && !it->MayHaveSideEffects()) { - it->Invalidate(); - it = block->Instructions().erase(it); + auto it{block->end()}; + while (it != block->begin()) { + --it; + if constexpr (TEST_USES) { + if (it->HasUses() || it->MayHaveSideEffects()) { + continue; + } + } + it->Invalidate(); + it = block->Instructions().erase(it); + } +} + +void DeletedPhiArgElimination(IR::Program& program, std::span<const IR::Block*> dead_blocks) { + for (IR::Block* const block : program.blocks) { + for (IR::Inst& phi : *block) { + if (!IR::IsPhi(phi)) { + continue; + } + for (size_t i = 0; i < phi.NumArgs(); ++i) { + if (std::ranges::find(dead_blocks, phi.PhiBlock(i)) == dead_blocks.end()) { + continue; + } + // Phi operand at this index is an unreachable block + phi.ErasePhiOperand(i); + --i; + } + } + } +} + +void DeadBranchElimination(IR::Program& program) { + boost::container::small_vector<const IR::Block*, 3> dead_blocks; + const auto begin_it{program.syntax_list.begin()}; + for (auto node_it = begin_it; node_it != program.syntax_list.end(); ++node_it) { + if (node_it->type != IR::AbstractSyntaxNode::Type::If) { + continue; + } + IR::Inst* const cond_ref{node_it->data.if_node.cond.Inst()}; + const IR::U1 cond{cond_ref->Arg(0)}; + if (!cond.IsImmediate()) { + continue; + } + if (cond.U1()) { + continue; + } + // False immediate condition. Remove condition ref, erase the entire branch. + cond_ref->Invalidate(); + // Account for nested if-statements within the if(false) branch + u32 nested_ifs{1u}; + while (node_it->type != IR::AbstractSyntaxNode::Type::EndIf || nested_ifs > 0) { + node_it = program.syntax_list.erase(node_it); + switch (node_it->type) { + case IR::AbstractSyntaxNode::Type::If: + ++nested_ifs; + break; + case IR::AbstractSyntaxNode::Type::EndIf: + --nested_ifs; + break; + case IR::AbstractSyntaxNode::Type::Block: { + IR::Block* const block{node_it->data.block}; + DeadInstElimination<false>(block); + dead_blocks.push_back(block); + break; + } + default: + break; } } + // Erase EndIf node of the if(false) branch + node_it = program.syntax_list.erase(node_it); + // Account for loop increment + --node_it; + } + if (!dead_blocks.empty()) { + DeletedPhiArgElimination(program, std::span(dead_blocks.data(), dead_blocks.size())); + } +} +} // namespace + +void DeadCodeEliminationPass(IR::Program& program) { + DeadBranchElimination(program); + for (IR::Block* const block : program.post_order_blocks) { + DeadInstElimination<true>(block); } } diff --git a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp index 38592afd0..3cc1cc07a 100644 --- a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp +++ b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp @@ -2,10 +2,7 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. -#include <algorithm> -#include <compare> #include <optional> -#include <queue> #include <boost/container/flat_set.hpp> #include <boost/container/small_vector.hpp> @@ -334,7 +331,8 @@ std::optional<LowAddrInfo> TrackLowAddress(IR::Inst* inst) { /// Tries to track the storage buffer address used by a global memory instruction std::optional<StorageBufferAddr> Track(const IR::Value& value, const Bias* bias) { const auto pred{[bias](const IR::Inst* inst) -> std::optional<StorageBufferAddr> { - if (inst->GetOpcode() != IR::Opcode::GetCbufU32) { + if (inst->GetOpcode() != IR::Opcode::GetCbufU32 && + inst->GetOpcode() != IR::Opcode::GetCbufU32x2) { return std::nullopt; } const IR::Value index{inst->Arg(0)}; diff --git a/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp b/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp index 773e1f961..622f94fc7 100644 --- a/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp +++ b/src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp @@ -2,9 +2,6 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. -#include <algorithm> - -#include "shader_recompiler/frontend/ir/ir_emitter.h" #include "shader_recompiler/frontend/ir/value.h" #include "shader_recompiler/ir_opt/passes.h" diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h index f877c7ba0..16ea3d80a 100644 --- a/src/shader_recompiler/ir_opt/passes.h +++ b/src/shader_recompiler/ir_opt/passes.h @@ -4,10 +4,7 @@ #pragma once -#include <span> - #include "shader_recompiler/environment.h" -#include "shader_recompiler/frontend/ir/basic_block.h" #include "shader_recompiler/frontend/ir/program.h" namespace Shader::Optimization { diff --git a/src/shader_recompiler/ir_opt/rescaling_pass.cpp b/src/shader_recompiler/ir_opt/rescaling_pass.cpp index c28500dd1..75679c793 100644 --- a/src/shader_recompiler/ir_opt/rescaling_pass.cpp +++ b/src/shader_recompiler/ir_opt/rescaling_pass.cpp @@ -2,7 +2,6 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. -#include "common/alignment.h" #include "common/settings.h" #include "shader_recompiler/environment.h" #include "shader_recompiler/frontend/ir/ir_emitter.h" @@ -183,6 +182,31 @@ void ScaleIntegerComposite(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_s } } +void ScaleIntegerOffsetComposite(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_scaled, + size_t index) { + const IR::Value composite{inst.Arg(index)}; + if (composite.IsEmpty()) { + return; + } + const auto info{inst.Flags<IR::TextureInstInfo>()}; + const IR::U32 x{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(composite, 0)})}; + const IR::U32 y{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(composite, 1)})}; + switch (info.type) { + case TextureType::ColorArray2D: + case TextureType::Color2D: + inst.SetArg(index, ir.CompositeConstruct(x, y)); + break; + case TextureType::Color1D: + case TextureType::ColorArray1D: + case TextureType::Color3D: + case TextureType::ColorCube: + case TextureType::ColorArrayCube: + case TextureType::Buffer: + // Nothing to patch here + break; + } +} + void SubScaleCoord(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_scaled) { const auto info{inst.Flags<IR::TextureInstInfo>()}; const IR::Value coord{inst.Arg(1)}; @@ -220,7 +244,7 @@ void SubScaleImageFetch(IR::Block& block, IR::Inst& inst) { const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))}; SubScaleCoord(ir, inst, is_scaled); // Scale ImageFetch offset - ScaleIntegerComposite(ir, inst, is_scaled, 2); + ScaleIntegerOffsetComposite(ir, inst, is_scaled, 2); } void SubScaleImageRead(IR::Block& block, IR::Inst& inst) { @@ -242,7 +266,7 @@ void PatchImageFetch(IR::Block& block, IR::Inst& inst) { const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))}; ScaleIntegerComposite(ir, inst, is_scaled, 1); // Scale ImageFetch offset - ScaleIntegerComposite(ir, inst, is_scaled, 2); + ScaleIntegerOffsetComposite(ir, inst, is_scaled, 2); } void PatchImageRead(IR::Block& block, IR::Inst& inst) { diff --git a/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp b/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp index 87aa09358..928557acb 100644 --- a/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp +++ b/src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp @@ -20,7 +20,6 @@ #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/opcodes.h" |