From e81739493a0cacc1efe3295f9d287d5d31b1a989 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 5 Feb 2021 05:58:02 -0300
Subject: shader: Constant propagation and global memory to storage buffer

---
 .../ir_opt/constant_propagation_pass.cpp           | 146 +++++++++++++++++++++
 1 file changed, 146 insertions(+)
 create mode 100644 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
new file mode 100644
index 000000000..02f5b653d
--- /dev/null
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -0,0 +1,146 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <type_traits>
+
+#include "common/bit_util.h"
+#include "shader_recompiler/exception.h"
+#include "shader_recompiler/frontend/ir/microinstruction.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+namespace {
+[[nodiscard]] u32 BitFieldUExtract(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 BitFieldUExtract({}, {}, {})", base, shift, count);
+    }
+    return (base >> shift) & ((1U << count) - 1);
+}
+
+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, u64>) {
+        return value.U64();
+    }
+}
+
+template <typename ImmFn>
+bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
+    const auto arg = [](const IR::Value& value) {
+        if constexpr (std::is_invocable_r_v<bool, ImmFn, bool, bool>) {
+            return value.U1();
+        } else if constexpr (std::is_invocable_r_v<u32, ImmFn, u32, u32>) {
+            return value.U32();
+        } else if constexpr (std::is_invocable_r_v<u64, ImmFn, u64, u64>) {
+            return value.U64();
+        }
+    };
+    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(lhs), arg(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(lhs), arg(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(rhs), arg(lhs_inst->Arg(1)))};
+            inst.SetArg(0, lhs_inst->Arg(0));
+            inst.SetArg(1, IR::Value{combined});
+        }
+    }
+    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});
+    }
+}
+
+template <typename T>
+void FoldAdd(IR::Inst& inst) {
+    if (inst.HasAssociatedPseudoOperation()) {
+        return;
+    }
+    if (!FoldCommutative(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));
+    }
+}
+
+void FoldLogicalAnd(IR::Inst& inst) {
+    if (!FoldCommutative(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 ConstantPropagation(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>(inst);
+    case IR::Opcode::IAdd64:
+        return FoldAdd<u64>(inst);
+    case IR::Opcode::BitFieldUExtract:
+        if (inst.AreAllArgsImmediates() && !inst.HasAssociatedPseudoOperation()) {
+            inst.ReplaceUsesWith(IR::Value{
+                BitFieldUExtract(inst.Arg(0).U32(), inst.Arg(1).U32(), inst.Arg(2).U32())});
+        }
+        break;
+    case IR::Opcode::LogicalAnd:
+        return FoldLogicalAnd(inst);
+    default:
+        break;
+    }
+}
+} // Anonymous namespace
+
+void ConstantPropagationPass(IR::Block& block) {
+    std::ranges::for_each(block, ConstantPropagation);
+}
+
+} // namespace Shader::Optimization
-- 
cgit v1.2.3


From be94ee88d227d0d3dbeabe9ade98bacd910c7a7e Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 5 Feb 2021 19:19:36 -0300
Subject: shader: Make typed IR

---
 .../ir_opt/constant_propagation_pass.cpp             | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 02f5b653d..7fb3192d8 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <type_traits>
 
+#include "common/bit_cast.h"
 #include "common/bit_util.h"
 #include "shader_recompiler/exception.h"
 #include "shader_recompiler/frontend/ir/microinstruction.h"
@@ -25,6 +26,8 @@ template <typename T>
         return value.U1();
     } else if constexpr (std::is_same_v<T, u32>) {
         return 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();
     }
@@ -115,6 +118,19 @@ void FoldLogicalAnd(IR::Inst& inst) {
     }
 }
 
+template <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 (value.InstRecursive()->Opcode() == reverse) {
+        inst.ReplaceUsesWith(arg_inst->Arg(0));
+    }
+}
+
 void ConstantPropagation(IR::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -123,6 +139,10 @@ void ConstantPropagation(IR::Inst& inst) {
         return FoldGetPred(inst);
     case IR::Opcode::IAdd32:
         return FoldAdd<u32>(inst);
+    case IR::Opcode::BitCastF32U32:
+        return FoldBitCast<f32, u32>(inst, IR::Opcode::BitCastU32F32);
+    case IR::Opcode::BitCastU32F32:
+        return FoldBitCast<u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(inst);
     case IR::Opcode::BitFieldUExtract:
-- 
cgit v1.2.3


From 6dafb08f52ac78119669a698c4b9a39bffd48f8f Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Sat, 6 Feb 2021 04:47:53 -0300
Subject: shader: Better constant folding

---
 .../ir_opt/constant_propagation_pass.cpp           | 59 +++++++++++++++++-----
 1 file changed, 47 insertions(+), 12 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 7fb3192d8..f1170c61e 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <tuple>
 #include <type_traits>
 
 #include "common/bit_cast.h"
@@ -13,12 +14,17 @@
 
 namespace Shader::Optimization {
 namespace {
-[[nodiscard]] u32 BitFieldUExtract(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 BitFieldUExtract({}, {}, {})", base, shift, count);
-    }
-    return (base >> shift) & ((1U << count) - 1);
-}
+// 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) {
@@ -104,6 +110,14 @@ void FoldAdd(IR::Inst& inst) {
     }
 }
 
+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(inst, [](bool a, bool b) { return a && b; })) {
         return;
@@ -131,6 +145,21 @@ void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
     }
 }
 
+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))...)};
+}
+
+template <typename Func>
+void FoldWhenAllImmediates(IR::Inst& inst, Func&& func) {
+    if (!inst.AreAllArgsImmediates() || inst.HasAssociatedPseudoOperation()) {
+        return;
+    }
+    using Indices = std::make_index_sequence<LambdaTraits<decltype(func)>::NUM_ARGS>;
+    inst.ReplaceUsesWith(EvalImmediates(inst, func, Indices{}));
+}
+
 void ConstantPropagation(IR::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -145,14 +174,20 @@ void ConstantPropagation(IR::Inst& inst) {
         return FoldBitCast<u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(inst);
-    case IR::Opcode::BitFieldUExtract:
-        if (inst.AreAllArgsImmediates() && !inst.HasAssociatedPseudoOperation()) {
-            inst.ReplaceUsesWith(IR::Value{
-                BitFieldUExtract(inst.Arg(0).U32(), inst.Arg(1).U32(), inst.Arg(2).U32())});
-        }
-        break;
+    case IR::Opcode::Select32:
+        return FoldSelect<u32>(inst);
     case IR::Opcode::LogicalAnd:
         return FoldLogicalAnd(inst);
+    case IR::Opcode::ULessThan:
+        return FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
+    case IR::Opcode::BitFieldUExtract:
+        return 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);
+        });
     default:
         break;
     }
-- 
cgit v1.2.3


From 9170200a11715d131645d1ffb92e86e6ef0d7e88 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Thu, 11 Feb 2021 16:39:06 -0300
Subject: shader: Initial implementation of an AST

---
 .../ir_opt/constant_propagation_pass.cpp           | 50 ++++++++++++++++++++++
 1 file changed, 50 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index f1170c61e..9fba6ac23 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -132,6 +132,32 @@ void FoldLogicalAnd(IR::Inst& inst) {
     }
 }
 
+void FoldLogicalOr(IR::Inst& inst) {
+    if (!FoldCommutative(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 <typename Dest, typename Source>
 void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
     const IR::Value value{inst.Arg(0)};
@@ -160,6 +186,24 @@ void FoldWhenAllImmediates(IR::Inst& inst, Func&& func) {
     inst.ReplaceUsesWith(EvalImmediates(inst, func, Indices{}));
 }
 
+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::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -178,6 +222,10 @@ void ConstantPropagation(IR::Inst& inst) {
         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::ULessThan:
         return FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
     case IR::Opcode::BitFieldUExtract:
@@ -188,6 +236,8 @@ void ConstantPropagation(IR::Inst& inst) {
             }
             return (base >> shift) & ((1U << count) - 1);
         });
+    case IR::Opcode::BranchConditional:
+        return FoldBranchConditional(inst);
     default:
         break;
     }
-- 
cgit v1.2.3


From 8af9297f0972d0aaa8306369c5d04926b886a89e Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Sun, 14 Feb 2021 01:24:32 -0300
Subject: shader: Misc fixes

---
 .../ir_opt/constant_propagation_pass.cpp           | 27 +++++++++-------------
 1 file changed, 11 insertions(+), 16 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 9fba6ac23..cbde65b9b 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -32,6 +32,8 @@ template <typename T>
         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>) {
@@ -39,17 +41,8 @@ template <typename T>
     }
 }
 
-template <typename ImmFn>
+template <typename T, typename ImmFn>
 bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
-    const auto arg = [](const IR::Value& value) {
-        if constexpr (std::is_invocable_r_v<bool, ImmFn, bool, bool>) {
-            return value.U1();
-        } else if constexpr (std::is_invocable_r_v<u32, ImmFn, u32, u32>) {
-            return value.U32();
-        } else if constexpr (std::is_invocable_r_v<u64, ImmFn, u64, u64>) {
-            return value.U64();
-        }
-    };
     const IR::Value lhs{inst.Arg(0)};
     const IR::Value rhs{inst.Arg(1)};
 
@@ -57,14 +50,14 @@ bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
     const bool is_rhs_immediate{rhs.IsImmediate()};
 
     if (is_lhs_immediate && is_rhs_immediate) {
-        const auto result{imm_fn(arg(lhs), arg(rhs))};
+        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(lhs), arg(rhs_inst->Arg(1)))};
+            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 {
@@ -76,7 +69,7 @@ bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
     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(rhs), arg(lhs_inst->Arg(1)))};
+            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});
         }
@@ -101,7 +94,7 @@ void FoldAdd(IR::Inst& inst) {
     if (inst.HasAssociatedPseudoOperation()) {
         return;
     }
-    if (!FoldCommutative(inst, [](T a, T b) { return a + b; })) {
+    if (!FoldCommutative<T>(inst, [](T a, T b) { return a + b; })) {
         return;
     }
     const IR::Value rhs{inst.Arg(1)};
@@ -119,7 +112,7 @@ void FoldSelect(IR::Inst& inst) {
 }
 
 void FoldLogicalAnd(IR::Inst& inst) {
-    if (!FoldCommutative(inst, [](bool a, bool b) { return a && b; })) {
+    if (!FoldCommutative<bool>(inst, [](bool a, bool b) { return a && b; })) {
         return;
     }
     const IR::Value rhs{inst.Arg(1)};
@@ -133,7 +126,7 @@ void FoldLogicalAnd(IR::Inst& inst) {
 }
 
 void FoldLogicalOr(IR::Inst& inst) {
-    if (!FoldCommutative(inst, [](bool a, bool b) { return a || b; })) {
+    if (!FoldCommutative<bool>(inst, [](bool a, bool b) { return a || b; })) {
         return;
     }
     const IR::Value rhs{inst.Arg(1)};
@@ -226,6 +219,8 @@ void ConstantPropagation(IR::Inst& inst) {
         return FoldLogicalOr(inst);
     case IR::Opcode::LogicalNot:
         return FoldLogicalNot(inst);
+    case IR::Opcode::SLessThan:
+        return FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a < b; });
     case IR::Opcode::ULessThan:
         return FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
     case IR::Opcode::BitFieldUExtract:
-- 
cgit v1.2.3


From b5d7279d878211654b4abb165d94af763a365f47 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Tue, 16 Feb 2021 04:10:22 -0300
Subject: spirv: Initial bindings support

---
 .../ir_opt/constant_propagation_pass.cpp           | 76 ++++++++++++++++++----
 1 file changed, 64 insertions(+), 12 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index cbde65b9b..f1ad16d60 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -77,6 +77,16 @@ bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
     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}});
@@ -103,6 +113,52 @@ void FoldAdd(IR::Inst& inst) {
     }
 }
 
+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::GetCbuf && b->Opcode() == IR::Opcode::GetCbuf &&
+               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::GetCbuf) {
+        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)};
@@ -170,15 +226,6 @@ IR::Value EvalImmediates(const IR::Inst& inst, Func&& func, std::index_sequence<
     return IR::Value{func(Arg<Traits::ArgType<I>>(inst.Arg(I))...)};
 }
 
-template <typename Func>
-void FoldWhenAllImmediates(IR::Inst& inst, Func&& func) {
-    if (!inst.AreAllArgsImmediates() || inst.HasAssociatedPseudoOperation()) {
-        return;
-    }
-    using Indices = std::make_index_sequence<LambdaTraits<decltype(func)>::NUM_ARGS>;
-    inst.ReplaceUsesWith(EvalImmediates(inst, func, Indices{}));
-}
-
 void FoldBranchConditional(IR::Inst& inst) {
     const IR::U1 cond{inst.Arg(0)};
     if (cond.IsImmediate()) {
@@ -205,6 +252,8 @@ void ConstantPropagation(IR::Inst& inst) {
         return FoldGetPred(inst);
     case IR::Opcode::IAdd32:
         return FoldAdd<u32>(inst);
+    case IR::Opcode::ISub32:
+        return FoldISub32(inst);
     case IR::Opcode::BitCastF32U32:
         return FoldBitCast<f32, u32>(inst, IR::Opcode::BitCastU32F32);
     case IR::Opcode::BitCastU32F32:
@@ -220,17 +269,20 @@ void ConstantPropagation(IR::Inst& inst) {
     case IR::Opcode::LogicalNot:
         return FoldLogicalNot(inst);
     case IR::Opcode::SLessThan:
-        return FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a < b; });
+        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a < b; });
+        return;
     case IR::Opcode::ULessThan:
-        return FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
+        return;
     case IR::Opcode::BitFieldUExtract:
-        return FoldWhenAllImmediates(inst, [](u32 base, u32 shift, u32 count) {
+        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:
-- 
cgit v1.2.3


From 58914796c06662f4f901a4f195057ee1327cf055 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Tue, 16 Feb 2021 19:50:23 -0300
Subject: shader: Add XMAD multiplication folding optimization

---
 .../ir_opt/constant_propagation_pass.cpp           | 82 ++++++++++++++++++++--
 1 file changed, 77 insertions(+), 5 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index f1ad16d60..9eb61b54c 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -9,6 +9,7 @@
 #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"
 
@@ -99,8 +100,71 @@ void FoldGetPred(IR::Inst& inst) {
     }
 }
 
+/// 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 (uses: 1)
+     *   %rhs_mul = IMul32 %rhs_bfe, %factor_b (uses: 1)
+     *   %lhs_bfe = BitFieldUExtract %factor_a, #16, #16 (uses: 1)
+     *   %rhs_mul = IMul32 %lhs_bfe, %factor_b (uses: 1)
+     *   %lhs_shl = ShiftLeftLogical32 %rhs_mul, #16 (uses: 1)
+     *   %result  = IAdd32 %lhs_shl, %rhs_mul (uses: 10)
+     *
+     * 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::Inst& inst) {
+void FoldAdd(IR::Block& block, IR::Inst& inst) {
     if (inst.HasAssociatedPseudoOperation()) {
         return;
     }
@@ -110,6 +174,12 @@ void FoldAdd(IR::Inst& inst) {
     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;
+        }
     }
 }
 
@@ -244,14 +314,14 @@ void FoldBranchConditional(IR::Inst& inst) {
     }
 }
 
-void ConstantPropagation(IR::Inst& inst) {
+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>(inst);
+        return FoldAdd<u32>(block, inst);
     case IR::Opcode::ISub32:
         return FoldISub32(inst);
     case IR::Opcode::BitCastF32U32:
@@ -259,7 +329,7 @@ void ConstantPropagation(IR::Inst& inst) {
     case IR::Opcode::BitCastU32F32:
         return FoldBitCast<u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
-        return FoldAdd<u64>(inst);
+        return FoldAdd<u64>(block, inst);
     case IR::Opcode::Select32:
         return FoldSelect<u32>(inst);
     case IR::Opcode::LogicalAnd:
@@ -292,7 +362,9 @@ void ConstantPropagation(IR::Inst& inst) {
 } // Anonymous namespace
 
 void ConstantPropagationPass(IR::Block& block) {
-    std::ranges::for_each(block, ConstantPropagation);
+    for (IR::Inst& inst : block) {
+        ConstantPropagation(block, inst);
+    }
 }
 
 } // namespace Shader::Optimization
-- 
cgit v1.2.3


From 6db69990da9f232e6d982cdcb69c2e27d93075cf Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 19 Feb 2021 18:10:18 -0300
Subject: spirv: Add lower fp16 to fp32 pass

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 9eb61b54c..4d4e88259 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -104,12 +104,12 @@ void FoldGetPred(IR::Inst& inst) {
 bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
     /*
      * We are looking for this pattern:
-     *   %rhs_bfe = BitFieldUExtract %factor_a, #0, #16 (uses: 1)
-     *   %rhs_mul = IMul32 %rhs_bfe, %factor_b (uses: 1)
-     *   %lhs_bfe = BitFieldUExtract %factor_a, #16, #16 (uses: 1)
-     *   %rhs_mul = IMul32 %lhs_bfe, %factor_b (uses: 1)
-     *   %lhs_shl = ShiftLeftLogical32 %rhs_mul, #16 (uses: 1)
-     *   %result  = IAdd32 %lhs_shl, %rhs_mul (uses: 10)
+     *   %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
-- 
cgit v1.2.3


From e44752ddc8804961eb84f8c225bb36d5b4c77bc1 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 22 Feb 2021 22:59:16 -0300
Subject: shader: FMUL, select, RRO, and MUFU fixes

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 4d4e88259..ae3d5a7d6 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -330,7 +330,7 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return FoldBitCast<u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(block, inst);
-    case IR::Opcode::Select32:
+    case IR::Opcode::SelectU32:
         return FoldSelect<u32>(inst);
     case IR::Opcode::LogicalAnd:
         return FoldLogicalAnd(inst);
-- 
cgit v1.2.3


From 3a63fa0477ea8297c80133d35494e1dfdc012f95 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Tue, 9 Mar 2021 17:14:57 -0300
Subject: shader: Partial implementation of LDC

---
 .../ir_opt/constant_propagation_pass.cpp           | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index ae3d5a7d6..7ba9ebe9b 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -193,7 +193,7 @@ void FoldISub32(IR::Inst& inst) {
     // 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::GetCbuf && b->Opcode() == IR::Opcode::GetCbuf &&
+        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()};
@@ -207,7 +207,7 @@ void FoldISub32(IR::Inst& inst) {
         // Canonicalize local variables to simplify the following logic
         std::swap(op_a, op_b);
     }
-    if (op_b->Opcode() != IR::Opcode::GetCbuf) {
+    if (op_b->Opcode() != IR::Opcode::GetCbufU32) {
         return;
     }
     IR::Inst* const inst_cbuf{op_b};
@@ -277,7 +277,7 @@ void FoldLogicalNot(IR::Inst& inst) {
     }
 }
 
-template <typename Dest, typename Source>
+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()) {
@@ -285,8 +285,18 @@ void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
         return;
     }
     IR::Inst* const arg_inst{value.InstRecursive()};
-    if (value.InstRecursive()->Opcode() == reverse) {
+    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;
+        }
     }
 }
 
@@ -325,9 +335,9 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::ISub32:
         return FoldISub32(inst);
     case IR::Opcode::BitCastF32U32:
-        return FoldBitCast<f32, u32>(inst, IR::Opcode::BitCastU32F32);
+        return FoldBitCast<IR::Opcode::BitCastF32U32, f32, u32>(inst, IR::Opcode::BitCastU32F32);
     case IR::Opcode::BitCastU32F32:
-        return FoldBitCast<u32, f32>(inst, IR::Opcode::BitCastF32U32);
+        return FoldBitCast<IR::Opcode::BitCastU32F32, u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(block, inst);
     case IR::Opcode::SelectU32:
-- 
cgit v1.2.3


From 71f96fa6366dc6dd306a953bca1b958fb32bc55a Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Sun, 14 Mar 2021 03:41:05 -0300
Subject: shader: Implement CAL inlining function calls

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 7ba9ebe9b..a39db2bf1 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -371,9 +371,11 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
 }
 } // Anonymous namespace
 
-void ConstantPropagationPass(IR::Block& block) {
-    for (IR::Inst& inst : block) {
-        ConstantPropagation(block, inst);
+void ConstantPropagationPass(IR::Program& program) {
+    for (IR::Block* const block : program.post_order_blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            ConstantPropagation(*block, inst);
+        }
     }
 }
 
-- 
cgit v1.2.3


From e4e1cc11b8f7649171fe922b2899e57120bfba53 Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Sun, 21 Mar 2021 19:28:37 -0400
Subject: shader: Implement DMNMX, DSET, DSETP

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index a39db2bf1..ef7766d22 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -229,7 +229,6 @@ void FoldISub32(IR::Inst& inst) {
     }
 }
 
-template <typename T>
 void FoldSelect(IR::Inst& inst) {
     const IR::Value cond{inst.Arg(0)};
     if (cond.IsImmediate()) {
@@ -340,8 +339,15 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return FoldBitCast<IR::Opcode::BitCastU32F32, u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(block, inst);
+    case IR::Opcode::SelectU1:
+    case IR::Opcode::SelectU8:
+    case IR::Opcode::SelectU16:
     case IR::Opcode::SelectU32:
-        return FoldSelect<u32>(inst);
+    case IR::Opcode::SelectU64:
+    case IR::Opcode::SelectF16:
+    case IR::Opcode::SelectF32:
+    case IR::Opcode::SelectF64:
+        return FoldSelect(inst);
     case IR::Opcode::LogicalAnd:
         return FoldLogicalAnd(inst);
     case IR::Opcode::LogicalOr:
-- 
cgit v1.2.3


From 2be5c7eff40647344da7951dc5e20a7deebf78aa Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Sun, 21 Mar 2021 21:56:27 -0300
Subject: shader: Fold interpolation multiplications

---
 .../ir_opt/constant_propagation_pass.cpp           | 34 ++++++++++++++++++++++
 1 file changed, 34 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index ef7766d22..3dab424f6 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -236,6 +236,38 @@ void FoldSelect(IR::Inst& inst) {
     }
 }
 
+void FoldFPMul32(IR::Inst& inst) {
+    const auto control{inst.Flags<IR::FpControl>()};
+    if (control.no_contraction) {
+        return;
+    }
+    // Fold interpolation operations
+    const IR::Value lhs_value{inst.Arg(0)};
+    const IR::Value rhs_value{inst.Arg(1)};
+    if (lhs_value.IsImmediate() || rhs_value.IsImmediate()) {
+        return;
+    }
+    IR::Inst* const lhs_op{lhs_value.InstRecursive()};
+    IR::Inst* const rhs_op{rhs_value.InstRecursive()};
+    if (lhs_op->Opcode() != IR::Opcode::FPMul32 || rhs_op->Opcode() != IR::Opcode::FPRecip32) {
+        return;
+    }
+    const IR::Value recip_source{rhs_op->Arg(0)};
+    const IR::Value lhs_mul_source{lhs_op->Arg(1).Resolve()};
+    if (recip_source.IsImmediate() || lhs_mul_source.IsImmediate()) {
+        return;
+    }
+    IR::Inst* const attr_a{recip_source.InstRecursive()};
+    IR::Inst* const attr_b{lhs_mul_source.InstRecursive()};
+    if (attr_a->Opcode() != IR::Opcode::GetAttribute ||
+        attr_b->Opcode() != IR::Opcode::GetAttribute) {
+        return;
+    }
+    if (attr_a->Arg(0).Attribute() == attr_b->Arg(0).Attribute()) {
+        inst.ReplaceUsesWith(lhs_op->Arg(0));
+    }
+}
+
 void FoldLogicalAnd(IR::Inst& inst) {
     if (!FoldCommutative<bool>(inst, [](bool a, bool b) { return a && b; })) {
         return;
@@ -348,6 +380,8 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::SelectF32:
     case IR::Opcode::SelectF64:
         return FoldSelect(inst);
+    case IR::Opcode::FPMul32:
+        return FoldFPMul32(inst);
     case IR::Opcode::LogicalAnd:
         return FoldLogicalAnd(inst);
     case IR::Opcode::LogicalOr:
-- 
cgit v1.2.3


From c7c518e280d1ac04adb08d45145690fd06ac7b18 Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Wed, 24 Mar 2021 23:41:55 +0100
Subject: shader: Implement TLD4 and TLD4_B

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 12 ++++++++++++
 1 file changed, 12 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 3dab424f6..28060dccf 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -403,6 +403,18 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
             return (base >> shift) & ((1U << count) - 1);
         });
         return;
+    case IR::Opcode::BitFieldSExtract:
+        FoldWhenAllImmediates(inst, [](s32 base, u32 shift, u32 count) {
+            const size_t back_shift = static_cast<size_t>(shift) + static_cast<size_t>(count);
+            if (back_shift > Common::BitSize<s32>()) {
+                throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldSExtract,
+                                 base, shift, count);
+            }
+            const size_t left_shift = Common::BitSize<s32>() - back_shift;
+            return static_cast<u32>(static_cast<s32>(base << left_shift) >>
+                                    static_cast<size_t>(Common::BitSize<s32>() - count));
+        });
+        return;
     case IR::Opcode::BranchConditional:
         return FoldBranchConditional(inst);
     default:
-- 
cgit v1.2.3


From 742d11c2ad948c8630be15901514ec9e5e5fcd20 Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Fri, 26 Mar 2021 16:02:04 +0100
Subject: shader: Implement TLD4.PTP

---
 .../ir_opt/constant_propagation_pass.cpp               | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 28060dccf..12159e738 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -355,6 +355,17 @@ void FoldBranchConditional(IR::Inst& inst) {
     }
 }
 
+void FoldConstantComposite(IR::Inst& inst, size_t amount = 2) {
+    for (size_t i = 0; i < amount; i++) {
+        if (!inst.Arg(i).IsConstantContainer()) {
+            return;
+        }
+    }
+    auto info{inst.Flags<IR::CompositeDecoration>()};
+    info.is_constant = true;
+    inst.SetFlags(info);
+}
+
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -380,6 +391,13 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::SelectF32:
     case IR::Opcode::SelectF64:
         return FoldSelect(inst);
+    case IR::Opcode::CompositeConstructU32x2:
+    case IR::Opcode::CompositeConstructF16x2:
+    case IR::Opcode::CompositeConstructF32x2:
+    case IR::Opcode::CompositeConstructF64x2:
+        return FoldConstantComposite(inst, 2);
+    case IR::Opcode::CompositeConstructArrayU32x2:
+        return FoldConstantComposite(inst, 4);
     case IR::Opcode::FPMul32:
         return FoldFPMul32(inst);
     case IR::Opcode::LogicalAnd:
-- 
cgit v1.2.3


From d9c5bd9509e82fcde72c18663989931f97ed6518 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 26 Mar 2021 16:46:07 -0300
Subject: shader: Refactor PTP and other minor changes

---
 .../ir_opt/constant_propagation_pass.cpp           | 22 ++--------------------
 1 file changed, 2 insertions(+), 20 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 12159e738..052f1609b 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -355,17 +355,6 @@ void FoldBranchConditional(IR::Inst& inst) {
     }
 }
 
-void FoldConstantComposite(IR::Inst& inst, size_t amount = 2) {
-    for (size_t i = 0; i < amount; i++) {
-        if (!inst.Arg(i).IsConstantContainer()) {
-            return;
-        }
-    }
-    auto info{inst.Flags<IR::CompositeDecoration>()};
-    info.is_constant = true;
-    inst.SetFlags(info);
-}
-
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -391,13 +380,6 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::SelectF32:
     case IR::Opcode::SelectF64:
         return FoldSelect(inst);
-    case IR::Opcode::CompositeConstructU32x2:
-    case IR::Opcode::CompositeConstructF16x2:
-    case IR::Opcode::CompositeConstructF32x2:
-    case IR::Opcode::CompositeConstructF64x2:
-        return FoldConstantComposite(inst, 2);
-    case IR::Opcode::CompositeConstructArrayU32x2:
-        return FoldConstantComposite(inst, 4);
     case IR::Opcode::FPMul32:
         return FoldFPMul32(inst);
     case IR::Opcode::LogicalAnd:
@@ -423,12 +405,12 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return;
     case IR::Opcode::BitFieldSExtract:
         FoldWhenAllImmediates(inst, [](s32 base, u32 shift, u32 count) {
-            const size_t back_shift = static_cast<size_t>(shift) + static_cast<size_t>(count);
+            const size_t back_shift{static_cast<size_t>(shift) + static_cast<size_t>(count)};
             if (back_shift > Common::BitSize<s32>()) {
                 throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldSExtract,
                                  base, shift, count);
             }
-            const size_t left_shift = Common::BitSize<s32>() - back_shift;
+            const size_t left_shift{Common::BitSize<s32>() - back_shift};
             return static_cast<u32>(static_cast<s32>(base << left_shift) >>
                                     static_cast<size_t>(Common::BitSize<s32>() - count));
         });
-- 
cgit v1.2.3


From 5f22cd89e20cd39c1395b7bd3b8e667a40f53751 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Wed, 31 Mar 2021 19:46:38 -0300
Subject: shader: Fix constant propagation to use reverse post order

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 052f1609b..7da4d50ef 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <tuple>
 #include <type_traits>
+#include <ranges>
 
 #include "common/bit_cast.h"
 #include "common/bit_util.h"
@@ -424,7 +425,7 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
 } // Anonymous namespace
 
 void ConstantPropagationPass(IR::Program& program) {
-    for (IR::Block* const block : program.post_order_blocks) {
+    for (IR::Block* const block : program.post_order_blocks | std::views::reverse) {
         for (IR::Inst& inst : block->Instructions()) {
             ConstantPropagation(*block, inst);
         }
-- 
cgit v1.2.3


From c3bace756f2b21057e89d104ad18a34b2ad9083c Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Thu, 1 Apr 2021 06:39:47 +0200
Subject: shader: Fold comparisons and Pack/Unpack16

---
 .../ir_opt/constant_propagation_pass.cpp           | 42 +++++++++++++++++++++-
 1 file changed, 41 insertions(+), 1 deletion(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 7da4d50ef..15e16956e 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -3,9 +3,9 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <ranges>
 #include <tuple>
 #include <type_traits>
-#include <ranges>
 
 #include "common/bit_cast.h"
 #include "common/bit_util.h"
@@ -332,6 +332,18 @@ void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
     }
 }
 
+void FoldInverseFunc(IR::Inst& inst, IR::Opcode reverse) {
+    const IR::Value value{inst.Arg(0)};
+    if (value.IsImmediate()) {
+        return;
+    }
+    IR::Inst* const arg_inst{value.InstRecursive()};
+    if (arg_inst->Opcode() == reverse) {
+        inst.ReplaceUsesWith(arg_inst->Arg(0));
+        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)>;
@@ -372,6 +384,10 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return FoldBitCast<IR::Opcode::BitCastU32F32, u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(block, inst);
+    case IR::Opcode::PackHalf2x16:
+        return FoldInverseFunc(inst, IR::Opcode::UnpackHalf2x16);
+    case IR::Opcode::UnpackHalf2x16:
+        return FoldInverseFunc(inst, IR::Opcode::PackHalf2x16);
     case IR::Opcode::SelectU1:
     case IR::Opcode::SelectU8:
     case IR::Opcode::SelectU16:
@@ -395,6 +411,30 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::ULessThan:
         FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a < b; });
         return;
+    case IR::Opcode::SLessThanEqual:
+        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a <= b; });
+        return;
+    case IR::Opcode::ULessThanEqual:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a <= b; });
+        return;
+    case IR::Opcode::SGreaterThan:
+        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a > b; });
+        return;
+    case IR::Opcode::UGreaterThan:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a > b; });
+        return;
+    case IR::Opcode::SGreaterThanEqual:
+        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return a >= b; });
+        return;
+    case IR::Opcode::UGreaterThanEqual:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a >= b; });
+        return;
+    case IR::Opcode::IEqual:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a == b; });
+        return;
+    case IR::Opcode::INotEqual:
+        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>()) {
-- 
cgit v1.2.3


From bee81887998070d213c984d4ae4e5ae35de6fd96 Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Thu, 1 Apr 2021 07:42:58 +0200
Subject: shader: Fold composite extract

---
 .../ir_opt/constant_propagation_pass.cpp           | 62 ++++++++++++++++++++++
 1 file changed, 62 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 15e16956e..8999c3a3d 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -368,6 +368,50 @@ void FoldBranchConditional(IR::Inst& inst) {
     }
 }
 
+std::optional<IR::Value> FoldCompositeExtractImpl(IR::Value inst_value, IR::Opcode insert,
+                                                  IR::Opcode construct, u32 first_index) {
+    IR::Inst* const inst{inst_value.InstRecursive()};
+    if (inst->Opcode() == construct) {
+        return inst->Arg(first_index);
+    }
+
+    if (inst->Opcode() != insert) {
+        return std::nullopt;
+    }
+
+    IR::Value value_index{inst->Arg(2)};
+    if (!value_index.IsImmediate()) {
+        return std::nullopt;
+    }
+
+    const u32 second_index = value_index.U32();
+    if (first_index != second_index) {
+        IR::Value value_composite{inst->Arg(0)};
+        if (value_composite.IsImmediate()) {
+            return std::nullopt;
+        }
+        return FoldCompositeExtractImpl(value_composite, insert, construct, first_index);
+    }
+    return inst->Arg(1);
+}
+
+void FoldCompositeExtract(IR::Inst& inst, IR::Opcode construct, IR::Opcode insert) {
+    const IR::Value value_1{inst.Arg(0)};
+    const IR::Value value_2{inst.Arg(1)};
+    if (value_1.IsImmediate()) {
+        return;
+    }
+    if (!value_2.IsImmediate()) {
+        return;
+    }
+    const u32 first_index = value_2.U32();
+    auto result = FoldCompositeExtractImpl(value_1, insert, construct, first_index);
+    if (!result) {
+        return;
+    }
+    inst.ReplaceUsesWith(*result);
+}
+
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     switch (inst.Opcode()) {
     case IR::Opcode::GetRegister:
@@ -458,6 +502,24 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return;
     case IR::Opcode::BranchConditional:
         return FoldBranchConditional(inst);
+    case IR::Opcode::CompositeExtractF32x2:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x2,
+                                    IR::Opcode::CompositeInsertF32x2);
+    case IR::Opcode::CompositeExtractF32x3:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x3,
+                                    IR::Opcode::CompositeInsertF32x3);
+    case IR::Opcode::CompositeExtractF32x4:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x4,
+                                    IR::Opcode::CompositeInsertF32x4);
+    case IR::Opcode::CompositeExtractF16x2:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x2,
+                                    IR::Opcode::CompositeInsertF16x2);
+    case IR::Opcode::CompositeExtractF16x3:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x3,
+                                    IR::Opcode::CompositeInsertF16x3);
+    case IR::Opcode::CompositeExtractF16x4:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x4,
+                                    IR::Opcode::CompositeInsertF16x4);
     default:
         break;
     }
-- 
cgit v1.2.3


From baec84247fe815199595d9e8077b71f3b5c8317e Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Sat, 3 Apr 2021 01:48:39 +0200
Subject: shader: Address Feedback

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 8999c3a3d..1720d7a09 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -374,17 +374,14 @@ std::optional<IR::Value> FoldCompositeExtractImpl(IR::Value inst_value, IR::Opco
     if (inst->Opcode() == construct) {
         return inst->Arg(first_index);
     }
-
     if (inst->Opcode() != insert) {
         return std::nullopt;
     }
-
     IR::Value value_index{inst->Arg(2)};
     if (!value_index.IsImmediate()) {
         return std::nullopt;
     }
-
-    const u32 second_index = value_index.U32();
+    const u32 second_index{value_index.U32()};
     if (first_index != second_index) {
         IR::Value value_composite{inst->Arg(0)};
         if (value_composite.IsImmediate()) {
@@ -404,8 +401,8 @@ void FoldCompositeExtract(IR::Inst& inst, IR::Opcode construct, IR::Opcode inser
     if (!value_2.IsImmediate()) {
         return;
     }
-    const u32 first_index = value_2.U32();
-    auto result = FoldCompositeExtractImpl(value_1, insert, construct, first_index);
+    const u32 first_index{value_2.U32()};
+    const std::optional result{FoldCompositeExtractImpl(value_1, insert, construct, first_index)};
     if (!result) {
         return;
     }
-- 
cgit v1.2.3


From 0bb85f6a753c769266c95c4ba146b25b9eaaaffd Mon Sep 17 00:00:00 2001
From: lat9nq <22451773+lat9nq@users.noreply.github.com>
Date: Mon, 5 Apr 2021 22:25:22 -0400
Subject: shader_recompiler,video_core: Cleanup some GCC and Clang errors

Mostly fixing unused *, implicit conversion, braced scalar init,
fpermissive, and some others.

Some Clang errors likely remain in video_core, and std::ranges is still
a pertinent issue in shader_recompiler

shader_recompiler: cmake: Force bracket depth to 1024 on Clang
Increases the maximum fold expression depth

thread_worker: Include condition_variable

Don't use list initializers in control flow

Co-authored-by: ReinUsesLisp <reinuseslisp@airmail.cc>
---
 .../ir_opt/constant_propagation_pass.cpp           | 49 ++++++++++++----------
 1 file changed, 26 insertions(+), 23 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 1720d7a09..61fbbe04c 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -58,7 +58,7 @@ bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
     }
     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()) {
+        if (rhs_inst->GetOpcode() == inst.GetOpcode() && 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});
@@ -70,7 +70,7 @@ bool FoldCommutative(IR::Inst& inst, ImmFn&& imm_fn) {
     }
     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()) {
+        if (lhs_inst->GetOpcode() == inst.GetOpcode() && 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});
@@ -123,7 +123,8 @@ bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
         return false;
     }
     IR::Inst* const lhs_shl{lhs_arg.InstRecursive()};
-    if (lhs_shl->Opcode() != IR::Opcode::ShiftLeftLogical32 || lhs_shl->Arg(1) != IR::Value{16U}) {
+    if (lhs_shl->GetOpcode() != IR::Opcode::ShiftLeftLogical32 ||
+        lhs_shl->Arg(1) != IR::Value{16U}) {
         return false;
     }
     if (lhs_shl->Arg(0).IsImmediate()) {
@@ -131,7 +132,7 @@ bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
     }
     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) {
+    if (lhs_mul->GetOpcode() != IR::Opcode::IMul32 || rhs_mul->GetOpcode() != IR::Opcode::IMul32) {
         return false;
     }
     if (lhs_mul->Arg(1).Resolve() != rhs_mul->Arg(1).Resolve()) {
@@ -143,10 +144,10 @@ bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
     }
     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) {
+    if (lhs_bfe->GetOpcode() != IR::Opcode::BitFieldUExtract) {
         return false;
     }
-    if (rhs_bfe->Opcode() != IR::Opcode::BitFieldUExtract) {
+    if (rhs_bfe->GetOpcode() != IR::Opcode::BitFieldUExtract) {
         return false;
     }
     if (lhs_bfe->Arg(1) != IR::Value{16U} || lhs_bfe->Arg(2) != IR::Value{16U}) {
@@ -194,8 +195,9 @@ void FoldISub32(IR::Inst& inst) {
     // 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);
+        return a->GetOpcode() == IR::Opcode::GetCbufU32 &&
+               b->GetOpcode() == 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()};
@@ -204,15 +206,15 @@ void FoldISub32(IR::Inst& inst) {
         return;
     }
     // It's also possible a value is being added to a cbuf and then subtracted
-    if (op_b->Opcode() == IR::Opcode::IAdd32) {
+    if (op_b->GetOpcode() == IR::Opcode::IAdd32) {
         // Canonicalize local variables to simplify the following logic
         std::swap(op_a, op_b);
     }
-    if (op_b->Opcode() != IR::Opcode::GetCbufU32) {
+    if (op_b->GetOpcode() != IR::Opcode::GetCbufU32) {
         return;
     }
     IR::Inst* const inst_cbuf{op_b};
-    if (op_a->Opcode() != IR::Opcode::IAdd32) {
+    if (op_a->GetOpcode() != IR::Opcode::IAdd32) {
         return;
     }
     IR::Value add_op_a{op_a->Arg(0)};
@@ -250,7 +252,8 @@ void FoldFPMul32(IR::Inst& inst) {
     }
     IR::Inst* const lhs_op{lhs_value.InstRecursive()};
     IR::Inst* const rhs_op{rhs_value.InstRecursive()};
-    if (lhs_op->Opcode() != IR::Opcode::FPMul32 || rhs_op->Opcode() != IR::Opcode::FPRecip32) {
+    if (lhs_op->GetOpcode() != IR::Opcode::FPMul32 ||
+        rhs_op->GetOpcode() != IR::Opcode::FPRecip32) {
         return;
     }
     const IR::Value recip_source{rhs_op->Arg(0)};
@@ -260,8 +263,8 @@ void FoldFPMul32(IR::Inst& inst) {
     }
     IR::Inst* const attr_a{recip_source.InstRecursive()};
     IR::Inst* const attr_b{lhs_mul_source.InstRecursive()};
-    if (attr_a->Opcode() != IR::Opcode::GetAttribute ||
-        attr_b->Opcode() != IR::Opcode::GetAttribute) {
+    if (attr_a->GetOpcode() != IR::Opcode::GetAttribute ||
+        attr_b->GetOpcode() != IR::Opcode::GetAttribute) {
         return;
     }
     if (attr_a->Arg(0).Attribute() == attr_b->Arg(0).Attribute()) {
@@ -304,7 +307,7 @@ void FoldLogicalNot(IR::Inst& inst) {
         return;
     }
     IR::Inst* const arg{value.InstRecursive()};
-    if (arg->Opcode() == IR::Opcode::LogicalNot) {
+    if (arg->GetOpcode() == IR::Opcode::LogicalNot) {
         inst.ReplaceUsesWith(arg->Arg(0));
     }
 }
@@ -317,12 +320,12 @@ void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
         return;
     }
     IR::Inst* const arg_inst{value.InstRecursive()};
-    if (arg_inst->Opcode() == reverse) {
+    if (arg_inst->GetOpcode() == reverse) {
         inst.ReplaceUsesWith(arg_inst->Arg(0));
         return;
     }
     if constexpr (op == IR::Opcode::BitCastF32U32) {
-        if (arg_inst->Opcode() == IR::Opcode::GetCbufU32) {
+        if (arg_inst->GetOpcode() == IR::Opcode::GetCbufU32) {
             // Replace the bitcast with a typed constant buffer read
             inst.ReplaceOpcode(IR::Opcode::GetCbufF32);
             inst.SetArg(0, arg_inst->Arg(0));
@@ -338,7 +341,7 @@ void FoldInverseFunc(IR::Inst& inst, IR::Opcode reverse) {
         return;
     }
     IR::Inst* const arg_inst{value.InstRecursive()};
-    if (arg_inst->Opcode() == reverse) {
+    if (arg_inst->GetOpcode() == reverse) {
         inst.ReplaceUsesWith(arg_inst->Arg(0));
         return;
     }
@@ -347,7 +350,7 @@ void FoldInverseFunc(IR::Inst& inst, IR::Opcode reverse) {
 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))...)};
+    return IR::Value{func(Arg<typename Traits::template ArgType<I>>(inst.Arg(I))...)};
 }
 
 void FoldBranchConditional(IR::Inst& inst) {
@@ -357,7 +360,7 @@ void FoldBranchConditional(IR::Inst& inst) {
         return;
     }
     const IR::Inst* cond_inst{cond.InstRecursive()};
-    if (cond_inst->Opcode() == IR::Opcode::LogicalNot) {
+    if (cond_inst->GetOpcode() == 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
@@ -371,10 +374,10 @@ void FoldBranchConditional(IR::Inst& inst) {
 std::optional<IR::Value> FoldCompositeExtractImpl(IR::Value inst_value, IR::Opcode insert,
                                                   IR::Opcode construct, u32 first_index) {
     IR::Inst* const inst{inst_value.InstRecursive()};
-    if (inst->Opcode() == construct) {
+    if (inst->GetOpcode() == construct) {
         return inst->Arg(first_index);
     }
-    if (inst->Opcode() != insert) {
+    if (inst->GetOpcode() != insert) {
         return std::nullopt;
     }
     IR::Value value_index{inst->Arg(2)};
@@ -410,7 +413,7 @@ void FoldCompositeExtract(IR::Inst& inst, IR::Opcode construct, IR::Opcode inser
 }
 
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
-    switch (inst.Opcode()) {
+    switch (inst.GetOpcode()) {
     case IR::Opcode::GetRegister:
         return FoldGetRegister(inst);
     case IR::Opcode::GetPred:
-- 
cgit v1.2.3


From 2597cee85b74be40bfecf0dc9cda90263d6cce40 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 12 Apr 2021 19:41:53 -0300
Subject: shader: Add constant propagation for *&^| binary operations

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 12 ++++++++++++
 1 file changed, 12 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 61fbbe04c..ee73b5b60 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -422,6 +422,9 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return FoldAdd<u32>(block, inst);
     case IR::Opcode::ISub32:
         return FoldISub32(inst);
+    case IR::Opcode::IMul32:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a * b; });
+        return;
     case IR::Opcode::BitCastF32U32:
         return FoldBitCast<IR::Opcode::BitCastF32U32, f32, u32>(inst, IR::Opcode::BitCastU32F32);
     case IR::Opcode::BitCastU32F32:
@@ -479,6 +482,15 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::INotEqual:
         FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a != b; });
         return;
+    case IR::Opcode::BitwiseAnd32:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a & b; });
+        return;
+    case IR::Opcode::BitwiseOr32:
+        FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a | b; });
+        return;
+    case IR::Opcode::BitwiseXor32:
+        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>()) {
-- 
cgit v1.2.3


From 04c459fc8d99b41fa8a03c49523599e9bf797f9d Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Sat, 17 Apr 2021 11:56:45 +0200
Subject: shader: Implement fine derivates constant propagation

---
 .../ir_opt/constant_propagation_pass.cpp           | 67 ++++++++++++++++++++++
 1 file changed, 67 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index ee73b5b60..983fb20ab 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -412,6 +412,71 @@ void FoldCompositeExtract(IR::Inst& inst, IR::Opcode construct, IR::Opcode inser
     inst.ReplaceUsesWith(*result);
 }
 
+IR::Value GetThroughCast(IR::Value value, IR::Opcode expected_cast) {
+    if (value.IsImmediate()) {
+        return value;
+    }
+    IR::Inst* const inst{value.InstRecursive()};
+    if (inst->GetOpcode() == expected_cast) {
+        return inst->Arg(0).Resolve();
+    }
+    return value;
+}
+
+void FoldFSwizzleAdd(IR::Block& block, IR::Inst& inst) {
+    const IR::Value swizzle{inst.Arg(2)};
+    if (!swizzle.IsImmediate()) {
+        return;
+    }
+
+    const IR::Value value_1{GetThroughCast(inst.Arg(0).Resolve(), IR::Opcode::BitCastF32U32)};
+    const IR::Value value_2{GetThroughCast(inst.Arg(1).Resolve(), IR::Opcode::BitCastF32U32)};
+
+    if (value_1.IsImmediate()) {
+        return;
+    }
+
+    const u32 swizzle_value{swizzle.U32()};
+    if (swizzle_value != 0x99 && swizzle_value != 0xA5) {
+        return;
+    }
+
+    IR::Inst* const inst2{value_1.InstRecursive()};
+    if (inst2->GetOpcode() != IR::Opcode::ShuffleButterfly) {
+        return;
+    }
+    const IR::Value value_3{GetThroughCast(inst2->Arg(0).Resolve(), IR::Opcode::BitCastU32F32)};
+    if (value_2 != value_3) {
+        return;
+    }
+
+    const IR::Value index{inst2->Arg(1)};
+    const IR::Value clamp{inst2->Arg(2)};
+    const IR::Value segmentation_mask{inst2->Arg(3)};
+
+    if (!index.IsImmediate() || !clamp.IsImmediate() || !segmentation_mask.IsImmediate()) {
+        return;
+    }
+
+    if (clamp.U32() != 3 || segmentation_mask.U32() != 28) {
+        return;
+    }
+
+    if (swizzle_value == 0x99) {
+        // DPdxFine
+        if (index.U32() == 1) {
+            IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+            inst.ReplaceUsesWith(ir.DPdxFine(IR::F32{value_2}));
+        }
+    } else if (swizzle_value == 0xA5) {
+        // DPdyFine
+        if (index.U32() == 2) {
+            IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+            inst.ReplaceUsesWith(ir.DPdyFine(IR::F32{value_2}));
+        }
+    }
+}
+
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     switch (inst.GetOpcode()) {
     case IR::Opcode::GetRegister:
@@ -532,6 +597,8 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::CompositeExtractF16x4:
         return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x4,
                                     IR::Opcode::CompositeInsertF16x4);
+    case IR::Opcode::FSwizzleAdd:
+        return FoldFSwizzleAdd(block, inst);
     default:
         break;
     }
-- 
cgit v1.2.3


From 080857b60e78836901cf6e9601f48613812fcd04 Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Sat, 17 Apr 2021 12:51:43 +0200
Subject: shader: Add coarse derivatives

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 8 --------
 1 file changed, 8 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 983fb20ab..7e86f64a8 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -428,19 +428,15 @@ void FoldFSwizzleAdd(IR::Block& block, IR::Inst& inst) {
     if (!swizzle.IsImmediate()) {
         return;
     }
-
     const IR::Value value_1{GetThroughCast(inst.Arg(0).Resolve(), IR::Opcode::BitCastF32U32)};
     const IR::Value value_2{GetThroughCast(inst.Arg(1).Resolve(), IR::Opcode::BitCastF32U32)};
-
     if (value_1.IsImmediate()) {
         return;
     }
-
     const u32 swizzle_value{swizzle.U32()};
     if (swizzle_value != 0x99 && swizzle_value != 0xA5) {
         return;
     }
-
     IR::Inst* const inst2{value_1.InstRecursive()};
     if (inst2->GetOpcode() != IR::Opcode::ShuffleButterfly) {
         return;
@@ -449,19 +445,15 @@ void FoldFSwizzleAdd(IR::Block& block, IR::Inst& inst) {
     if (value_2 != value_3) {
         return;
     }
-
     const IR::Value index{inst2->Arg(1)};
     const IR::Value clamp{inst2->Arg(2)};
     const IR::Value segmentation_mask{inst2->Arg(3)};
-
     if (!index.IsImmediate() || !clamp.IsImmediate() || !segmentation_mask.IsImmediate()) {
         return;
     }
-
     if (clamp.U32() != 3 || segmentation_mask.U32() != 28) {
         return;
     }
-
     if (swizzle_value == 0x99) {
         // DPdxFine
         if (index.U32() == 1) {
-- 
cgit v1.2.3


From a8d46a5eae4e0d65c08f6574cffc48bc55d00940 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 19 Apr 2021 16:36:57 -0300
Subject: shader: Add constant propagation for arithmetic right shifts

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 3 +++
 1 file changed, 3 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 7e86f64a8..2206f93c2 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -482,6 +482,9 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::IMul32:
         FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a * b; });
         return;
+    case IR::Opcode::ShiftRightArithmetic32:
+        FoldWhenAllImmediates(inst, [](s32 a, s32 b) { return static_cast<u32>(a >> b); });
+        return;
     case IR::Opcode::BitCastF32U32:
         return FoldBitCast<IR::Opcode::BitCastF32U32, f32, u32>(inst, IR::Opcode::BitCastU32F32);
     case IR::Opcode::BitCastU32F32:
-- 
cgit v1.2.3


From 050e81500c002f304d581f28700de549b828a2bc Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Wed, 21 Apr 2021 00:35:47 -0300
Subject: shader: Move microinstruction header to the value header

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 2206f93c2..770d3de61 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -11,7 +11,7 @@
 #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/frontend/ir/value.h"
 #include "shader_recompiler/ir_opt/passes.h"
 
 namespace Shader::Optimization {
-- 
cgit v1.2.3


From 50eb03382e8ac8eb4aeb7cdc488a7ee097fec39d Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 23 Apr 2021 08:17:53 -0300
Subject: shader: Fix error checking in bitfieldExtract and implement
 bitfieldInsert folding

---
 .../ir_opt/constant_propagation_pass.cpp              | 19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 770d3de61..f16c5e8f6 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -553,7 +553,7 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         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>()) {
+            if (static_cast<size_t>(shift) + static_cast<size_t>(count) > 32) {
                 throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldUExtract,
                                  base, shift, count);
             }
@@ -563,13 +563,22 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::BitFieldSExtract:
         FoldWhenAllImmediates(inst, [](s32 base, u32 shift, u32 count) {
             const size_t back_shift{static_cast<size_t>(shift) + static_cast<size_t>(count)};
-            if (back_shift > Common::BitSize<s32>()) {
+            const size_t left_shift{32 - back_shift};
+            const size_t right_shift{static_cast<size_t>(32 - count)};
+            if (back_shift >= 32 || left_shift >= 32 || right_shift >= 32) {
                 throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldSExtract,
                                  base, shift, count);
             }
-            const size_t left_shift{Common::BitSize<s32>() - back_shift};
-            return static_cast<u32>(static_cast<s32>(base << left_shift) >>
-                                    static_cast<size_t>(Common::BitSize<s32>() - count));
+            return static_cast<u32>((base << left_shift) >> right_shift);
+        });
+        return;
+    case IR::Opcode::BitFieldInsert:
+        FoldWhenAllImmediates(inst, [](u32 base, u32 insert, u32 offset, u32 bits) {
+            if (bits >= 32 || offset >= 32) {
+                throw LogicError("Undefined result in {}({}, {}, {}, {})",
+                                 IR::Opcode::BitFieldInsert, base, insert, offset, bits);
+            }
+            return (base & ~(~(~0u << bits) << offset)) | (insert << offset);
         });
         return;
     case IR::Opcode::BranchConditional:
-- 
cgit v1.2.3


From 20e86fd61512626e267824c1a5469084c2d36c7a Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Sat, 24 Apr 2021 00:49:14 -0400
Subject: shader: Fix BFE s32 undefined check

Our unit tests were hitting this exception.
---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index f16c5e8f6..b1c45d13a 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -565,7 +565,7 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
             const size_t back_shift{static_cast<size_t>(shift) + static_cast<size_t>(count)};
             const size_t left_shift{32 - back_shift};
             const size_t right_shift{static_cast<size_t>(32 - count)};
-            if (back_shift >= 32 || left_shift >= 32 || right_shift >= 32) {
+            if (back_shift > 32 || left_shift >= 32 || right_shift >= 32) {
                 throw LogicError("Undefined result in {}({}, {}, {})", IR::Opcode::BitFieldSExtract,
                                  base, shift, count);
             }
-- 
cgit v1.2.3


From d54d7de40e7295827b0e4e4026441b53d3fc9569 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Fri, 14 May 2021 00:40:54 -0300
Subject: glasm: Rework control flow introducing a syntax list

This commit regresses VertexA shaders, their transformation pass has to
be adapted to the new control flow.
---
 .../ir_opt/constant_propagation_pass.cpp             | 20 --------------------
 1 file changed, 20 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index b1c45d13a..66f1391db 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -353,24 +353,6 @@ IR::Value EvalImmediates(const IR::Inst& inst, Func&& func, std::index_sequence<
     return IR::Value{func(Arg<typename Traits::template 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->GetOpcode() == 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);
-    }
-}
-
 std::optional<IR::Value> FoldCompositeExtractImpl(IR::Value inst_value, IR::Opcode insert,
                                                   IR::Opcode construct, u32 first_index) {
     IR::Inst* const inst{inst_value.InstRecursive()};
@@ -581,8 +563,6 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
             return (base & ~(~(~0u << bits) << offset)) | (insert << offset);
         });
         return;
-    case IR::Opcode::BranchConditional:
-        return FoldBranchConditional(inst);
     case IR::Opcode::CompositeExtractF32x2:
         return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x2,
                                     IR::Opcode::CompositeInsertF32x2);
-- 
cgit v1.2.3


From fbf5cdcba0f3b03e8cd3019fb285a96037b05f26 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Thu, 27 May 2021 19:04:26 -0300
Subject: shader: Fix FSwizzleAdd folding when going through phi nodes

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 66f1391db..796b4122d 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -440,13 +440,13 @@ void FoldFSwizzleAdd(IR::Block& block, IR::Inst& inst) {
         // DPdxFine
         if (index.U32() == 1) {
             IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
-            inst.ReplaceUsesWith(ir.DPdxFine(IR::F32{value_2}));
+            inst.ReplaceUsesWith(ir.DPdxFine(IR::F32{inst.Arg(1)}));
         }
     } else if (swizzle_value == 0xA5) {
         // DPdyFine
         if (index.U32() == 2) {
             IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
-            inst.ReplaceUsesWith(ir.DPdyFine(IR::F32{value_2}));
+            inst.ReplaceUsesWith(ir.DPdyFine(IR::F32{inst.Arg(1)}));
         }
     }
 }
-- 
cgit v1.2.3


From 97e80dda55aec0ff791e4990f09c85e7a7067730 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Wed, 23 Jun 2021 01:32:09 -0300
Subject: shader: Add constant propagation to integer vectors

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 9 +++++++++
 1 file changed, 9 insertions(+)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 796b4122d..3c72203ad 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -563,6 +563,15 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
             return (base & ~(~(~0u << bits) << offset)) | (insert << offset);
         });
         return;
+    case IR::Opcode::CompositeExtractU32x2:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructU32x2,
+                                    IR::Opcode::CompositeInsertU32x2);
+    case IR::Opcode::CompositeExtractU32x3:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructU32x3,
+                                    IR::Opcode::CompositeInsertU32x3);
+    case IR::Opcode::CompositeExtractU32x4:
+        return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructU32x4,
+                                    IR::Opcode::CompositeInsertU32x4);
     case IR::Opcode::CompositeExtractF32x2:
         return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x2,
                                     IR::Opcode::CompositeInsertF32x2);
-- 
cgit v1.2.3


From bf2956d77ab0ad06c4b5505cc9906e51e5878274 Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 12 Jul 2021 05:22:01 -0300
Subject: shader: Avoid usage of C++20 ranges to build in clang

---
 src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

(limited to 'src/shader_recompiler/ir_opt/constant_propagation_pass.cpp')

diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 3c72203ad..8dd6d6c2c 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -3,7 +3,6 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
-#include <ranges>
 #include <tuple>
 #include <type_traits>
 
@@ -599,7 +598,9 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
 } // Anonymous namespace
 
 void ConstantPropagationPass(IR::Program& program) {
-    for (IR::Block* const block : program.post_order_blocks | std::views::reverse) {
+    const auto end{program.post_order_blocks.rend()};
+    for (auto it = program.post_order_blocks.rbegin(); it != end; ++it) {
+        IR::Block* const block{*it};
         for (IR::Inst& inst : block->Instructions()) {
             ConstantPropagation(*block, inst);
         }
-- 
cgit v1.2.3