shader: Partial implementation of LDC

src/shader_recompiler/CMakeLists.txt

@@ -88,6 +88,7 @@ add_library(shader_recompiler STATIC
     frontend/maxwell/translate/impl/integer_shift_right.cpp
     frontend/maxwell/translate/impl/integer_short_multiply_add.cpp
     frontend/maxwell/translate/impl/integer_to_integer_conversion.cpp
+    frontend/maxwell/translate/impl/load_constant.cpp
     frontend/maxwell/translate/impl/load_effective_address.cpp
     frontend/maxwell/translate/impl/load_store_attribute.cpp
     frontend/maxwell/translate/impl/load_store_memory.cpp

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -104,15 +104,23 @@ void EmitContext::DefineCommonTypes(const Info& info) {
 
     U1 = Name(TypeBool(), "u1");
 
-    // TODO: Conditionally define these
-    AddCapability(spv::Capability::Int16);
-    AddCapability(spv::Capability::Int64);
-    U16 = Name(TypeInt(16, false), "u16");
-    U64 = Name(TypeInt(64, false), "u64");
-
     F32.Define(*this, TypeFloat(32), "f32");
     U32.Define(*this, TypeInt(32, false), "u32");
 
+    if (info.uses_int8) {
+        AddCapability(spv::Capability::Int8);
+        U8 = Name(TypeInt(8, false), "u8");
+        S8 = Name(TypeInt(8, true), "s8");
+    }
+    if (info.uses_int16) {
+        AddCapability(spv::Capability::Int16);
+        U16 = Name(TypeInt(16, false), "u16");
+        S16 = Name(TypeInt(16, true), "s16");
+    }
+    if (info.uses_int64) {
+        AddCapability(spv::Capability::Int64);
+        U64 = Name(TypeInt(64, false), "u64");
+    }
     if (info.uses_fp16) {
         AddCapability(spv::Capability::Float16);
         F16.Define(*this, TypeFloat(16), "f16");
@@ -151,26 +159,51 @@ void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     if (info.constant_buffer_descriptors.empty()) {
         return;
     }
-    const Id array_type{TypeArray(U32[1], Constant(U32[1], 4096))};
+    if (True(info.used_constant_buffer_types & IR::Type::U8)) {
-    Decorate(array_type, spv::Decoration::ArrayStride, 4U);
+        DefineConstantBuffers(info, &UniformDefinitions::U8, binding, U8, 'u', sizeof(u8));
+        DefineConstantBuffers(info, &UniformDefinitions::S8, binding, S8, 's', sizeof(s8));
+    }
+    if (True(info.used_constant_buffer_types & IR::Type::U16)) {
+        DefineConstantBuffers(info, &UniformDefinitions::U16, binding, U16, 'u', sizeof(u16));
+        DefineConstantBuffers(info, &UniformDefinitions::S16, binding, S16, 's', sizeof(s16));
+    }
+    if (True(info.used_constant_buffer_types & IR::Type::U32)) {
+        DefineConstantBuffers(info, &UniformDefinitions::U32, binding, U32[1], 'u', sizeof(u32));
+    }
+    if (True(info.used_constant_buffer_types & IR::Type::F32)) {
+        DefineConstantBuffers(info, &UniformDefinitions::F32, binding, F32[1], 'f', sizeof(f32));
+    }
+    if (True(info.used_constant_buffer_types & IR::Type::U64)) {
+        DefineConstantBuffers(info, &UniformDefinitions::U64, binding, U64, 'u', sizeof(u64));
+    }
+    for (const ConstantBufferDescriptor& desc : info.constant_buffer_descriptors) {
+        binding += desc.count;
+    }
+}
+
+void EmitContext::DefineConstantBuffers(const Info& info, Id UniformDefinitions::*member_type,
+                                        u32 binding, Id type, char type_char, u32 element_size) {
+    const Id array_type{TypeArray(type, Constant(U32[1], 65536U / element_size))};
+    Decorate(array_type, spv::Decoration::ArrayStride, element_size);
 
     const Id struct_type{TypeStruct(array_type)};
-    Name(struct_type, "cbuf_block");
+    Name(struct_type, fmt::format("cbuf_block_{}{}", type_char, element_size * CHAR_BIT));
     Decorate(struct_type, spv::Decoration::Block);
     MemberName(struct_type, 0, "data");
     MemberDecorate(struct_type, 0, spv::Decoration::Offset, 0U);
 
-    const Id uniform_type{TypePointer(spv::StorageClass::Uniform, struct_type)};
+    const Id struct_pointer_type{TypePointer(spv::StorageClass::Uniform, struct_type)};
-    uniform_u32 = TypePointer(spv::StorageClass::Uniform, U32[1]);
+    const Id uniform_type{TypePointer(spv::StorageClass::Uniform, type)};
+    uniform_types.*member_type = uniform_type;
 
-    u32 index{};
     for (const ConstantBufferDescriptor& desc : info.constant_buffer_descriptors) {
-        const Id id{AddGlobalVariable(uniform_type, spv::StorageClass::Uniform)};
+        const Id id{AddGlobalVariable(struct_pointer_type, spv::StorageClass::Uniform)};
         Decorate(id, spv::Decoration::Binding, binding);
         Decorate(id, spv::Decoration::DescriptorSet, 0U);
         Name(id, fmt::format("c{}", desc.index));
-        std::fill_n(cbufs.data() + desc.index, desc.count, id);
+        for (size_t i = 0; i < desc.count; ++i) {
-        index += desc.count;
+            cbufs[desc.index + i].*member_type = id;
+        }
         binding += desc.count;
     }
 }

src/shader_recompiler/backend/spirv/emit_context.h

@@ -10,8 +10,8 @@
 #include <sirit/sirit.h>
 
 #include "shader_recompiler/frontend/ir/program.h"
-#include "shader_recompiler/shader_info.h"
 #include "shader_recompiler/profile.h"
+#include "shader_recompiler/shader_info.h"
 
 namespace Shader::Backend::SPIRV {
 
@@ -34,6 +34,16 @@ struct TextureDefinition {
     Id type;
 };
 
+struct UniformDefinitions {
+    Id U8{};
+    Id S8{};
+    Id U16{};
+    Id S16{};
+    Id U32{};
+    Id F32{};
+    Id U64{};
+};
+
 class EmitContext final : public Sirit::Module {
 public:
     explicit EmitContext(const Profile& profile, IR::Program& program);
@@ -45,7 +55,10 @@ public:
 
     Id void_id{};
     Id U1{};
+    Id U8{};
+    Id S8{};
     Id U16{};
+    Id S16{};
     Id U64{};
     VectorTypes F32;
     VectorTypes U32;
@@ -56,10 +69,11 @@ public:
     Id false_value{};
     Id u32_zero_value{};
 
-    Id uniform_u32{};
+    UniformDefinitions uniform_types;
+
     Id storage_u32{};
 
-    std::array<Id, Info::MAX_CBUFS> cbufs{};
+    std::array<UniformDefinitions, Info::MAX_CBUFS> cbufs{};
     std::array<Id, Info::MAX_SSBOS> ssbos{};
     std::vector<TextureDefinition> textures;
 
@@ -71,6 +85,8 @@ private:
     void DefineCommonConstants();
     void DefineSpecialVariables(const Info& info);
     void DefineConstantBuffers(const Info& info, u32& binding);
+    void DefineConstantBuffers(const Info& info, Id UniformDefinitions::*member_type, u32 binding,
+                               Id type, char type_char, u32 element_size);
     void DefineStorageBuffers(const Info& info, u32& binding);
     void DefineTextures(const Info& info, u32& binding);
     void DefineLabels(IR::Program& program);

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -34,7 +34,13 @@ void EmitGetPred(EmitContext& ctx);
 void EmitSetPred(EmitContext& ctx);
 void EmitSetGotoVariable(EmitContext& ctx);
 void EmitGetGotoVariable(EmitContext& ctx);
-Id EmitGetCbuf(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
+Id EmitGetCbufU64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
 void EmitGetAttribute(EmitContext& ctx);
 void EmitSetAttribute(EmitContext& ctx);
 void EmitGetAttributeIndexed(EmitContext& ctx);

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -30,17 +30,61 @@ void EmitGetGotoVariable(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-Id EmitGetCbuf(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+static Id GetCbuf(EmitContext& ctx, Id result_type, Id UniformDefinitions::*member_ptr,
+                  u32 element_size, const IR::Value& binding, const IR::Value& offset) {
     if (!binding.IsImmediate()) {
         throw NotImplementedException("Constant buffer indexing");
     }
+    const Id cbuf{ctx.cbufs[binding.U32()].*member_ptr};
+    const Id uniform_type{ctx.uniform_types.*member_ptr};
     if (!offset.IsImmediate()) {
-        throw NotImplementedException("Variable constant buffer offset");
+        Id index{ctx.Def(offset)};
+        if (element_size > 1) {
+            const u32 log2_element_size{static_cast<u32>(std::countr_zero(element_size))};
+            const Id shift{ctx.Constant(ctx.U32[1], log2_element_size)};
+            index = ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), shift);
         }
-    const Id imm_offset{ctx.Constant(ctx.U32[1], offset.U32() / 4)};
+        const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, index)};
-    const Id cbuf{ctx.cbufs[binding.U32()]};
+        return ctx.OpLoad(result_type, access_chain);
-    const Id access_chain{ctx.OpAccessChain(ctx.uniform_u32, cbuf, ctx.u32_zero_value, imm_offset)};
+    }
-    return ctx.OpLoad(ctx.U32[1], access_chain);
+    if (offset.U32() % element_size != 0) {
+        throw NotImplementedException("Unaligned immediate constant buffer load");
+    }
+    const Id imm_offset{ctx.Constant(ctx.U32[1], offset.U32() / element_size)};
+    const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, imm_offset)};
+    return ctx.OpLoad(result_type, access_chain);
+}
+
+Id EmitGetCbufU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    const Id load{GetCbuf(ctx, ctx.U8, &UniformDefinitions::U8, sizeof(u8), binding, offset)};
+    return ctx.OpUConvert(ctx.U32[1], load);
+}
+
+Id EmitGetCbufS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    const Id load{GetCbuf(ctx, ctx.S8, &UniformDefinitions::S8, sizeof(s8), binding, offset)};
+    return ctx.OpSConvert(ctx.U32[1], load);
+}
+
+Id EmitGetCbufU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    const Id load{GetCbuf(ctx, ctx.U16, &UniformDefinitions::U16, sizeof(u16), binding, offset)};
+    return ctx.OpUConvert(ctx.U32[1], load);
+}
+
+Id EmitGetCbufS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    const Id load{GetCbuf(ctx, ctx.S16, &UniformDefinitions::S16, sizeof(s16), binding, offset)};
+    return ctx.OpSConvert(ctx.U32[1], load);
+}
+
+Id EmitGetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    return GetCbuf(ctx, ctx.U32[1], &UniformDefinitions::U32, sizeof(u32), binding, offset);
+}
+
+Id EmitGetCbufF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    return GetCbuf(ctx, ctx.F32[1], &UniformDefinitions::F32, sizeof(f32), binding, offset);
+}
+
+Id EmitGetCbufU64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    return GetCbuf(ctx, ctx.U64, &UniformDefinitions::U64, sizeof(u64), binding, offset);
 }
 
 void EmitGetAttribute(EmitContext&) {

src/shader_recompiler/frontend/ir/ir_emitter.cpp

@@ -112,7 +112,27 @@ void IREmitter::SetPred(IR::Pred pred, const U1& value) {
 }
 
 U32 IREmitter::GetCbuf(const U32& binding, const U32& byte_offset) {
-    return Inst<U32>(Opcode::GetCbuf, binding, byte_offset);
+    return Inst<U32>(Opcode::GetCbufU32, binding, byte_offset);
+}
+
+UAny IREmitter::GetCbuf(const U32& binding, const U32& byte_offset, size_t bitsize,
+                        bool is_signed) {
+    switch (bitsize) {
+    case 8:
+        return Inst<U32>(is_signed ? Opcode::GetCbufS8 : Opcode::GetCbufU8, binding, byte_offset);
+    case 16:
+        return Inst<U32>(is_signed ? Opcode::GetCbufS16 : Opcode::GetCbufU16, binding, byte_offset);
+    case 32:
+        return Inst<U32>(Opcode::GetCbufU32, binding, byte_offset);
+    case 64:
+        return Inst<U64>(Opcode::GetCbufU64, binding, byte_offset);
+    default:
+        throw InvalidArgument("Invalid bit size {}", bitsize);
+    }
+}
+
+F32 IREmitter::GetFloatCbuf(const U32& binding, const U32& byte_offset) {
+    return Inst<F32>(Opcode::GetCbufF32, binding, byte_offset);
 }
 
 U1 IREmitter::GetZFlag() {

src/shader_recompiler/frontend/ir/ir_emitter.h

@@ -47,6 +47,9 @@ public:
     void SetGotoVariable(u32 id, const U1& value);
 
     [[nodiscard]] U32 GetCbuf(const U32& binding, const U32& byte_offset);
+    [[nodiscard]] UAny GetCbuf(const U32& binding, const U32& byte_offset, size_t bitsize,
+                               bool is_signed);
+    [[nodiscard]] F32 GetFloatCbuf(const U32& binding, const U32& byte_offset);
 
     [[nodiscard]] U1 GetZFlag();
     [[nodiscard]] U1 GetSFlag();

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -21,7 +21,13 @@ OPCODE(GetPred,                                             U1,             Pred
 OPCODE(SetPred,                                             Void,           Pred,           U1,                                                             )
 OPCODE(GetGotoVariable,                                     U1,             U32,                                                                            )
 OPCODE(SetGotoVariable,                                     Void,           U32,            U1,                                                             )
-OPCODE(GetCbuf,                                             U32,            U32,            U32,                                                            )
+OPCODE(GetCbufU8,                                           U32,            U32,            U32,                                                            )
+OPCODE(GetCbufS8,                                           U32,            U32,            U32,                                                            )
+OPCODE(GetCbufU16,                                          U32,            U32,            U32,                                                            )
+OPCODE(GetCbufS16,                                          U32,            U32,            U32,                                                            )
+OPCODE(GetCbufU32,                                          U32,            U32,            U32,                                                            )
+OPCODE(GetCbufF32,                                          F32,            U32,            U32,                                                            )
+OPCODE(GetCbufU64,                                          U64,            U32,            U32,                                                            )
 OPCODE(GetAttribute,                                        U32,            Attribute,                                                                      )
 OPCODE(SetAttribute,                                        Void,           Attribute,      U32,                                                            )
 OPCODE(GetAttributeIndexed,                                 U32,            U32,                                                                            )

src/shader_recompiler/frontend/maxwell/translate/impl/impl.cpp

@@ -56,25 +56,32 @@ IR::F32 TranslatorVisitor::GetFloatReg39(u64 insn) {
     return ir.BitCast<IR::F32>(GetReg39(insn));
 }
 
-IR::U32 TranslatorVisitor::GetCbuf(u64 insn) {
+static std::pair<IR::U32, IR::U32> CbufAddr(u64 insn) {
     union {
         u64 raw;
         BitField<20, 14, s64> offset;
         BitField<34, 5, u64> binding;
     } const cbuf{insn};
+
     if (cbuf.binding >= 18) {
         throw NotImplementedException("Out of bounds constant buffer binding {}", cbuf.binding);
     }
     if (cbuf.offset >= 0x10'000 || cbuf.offset < 0) {
         throw NotImplementedException("Out of bounds constant buffer offset {}", cbuf.offset);
     }
-    const IR::U32 binding{ir.Imm32(static_cast<u32>(cbuf.binding))};
+    const IR::Value binding{static_cast<u32>(cbuf.binding)};
-    const IR::U32 byte_offset{ir.Imm32(static_cast<u32>(cbuf.offset) * 4)};
+    const IR::Value byte_offset{static_cast<u32>(cbuf.offset) * 4};
+    return {IR::U32{binding}, IR::U32{byte_offset}};
+}
+
+IR::U32 TranslatorVisitor::GetCbuf(u64 insn) {
+    const auto[binding, byte_offset]{CbufAddr(insn)};
     return ir.GetCbuf(binding, byte_offset);
 }
 
 IR::F32 TranslatorVisitor::GetFloatCbuf(u64 insn) {
-    return ir.BitCast<IR::F32>(GetCbuf(insn));
+    const auto[binding, byte_offset]{CbufAddr(insn)};
+    return ir.GetFloatCbuf(binding, byte_offset);
 }
 
 IR::U32 TranslatorVisitor::GetImm20(u64 insn) {
@@ -83,6 +90,7 @@ IR::U32 TranslatorVisitor::GetImm20(u64 insn) {
         BitField<20, 19, u64> value;
         BitField<56, 1, u64> is_negative;
     } const imm{insn};
+
     if (imm.is_negative != 0) {
         const s64 raw{static_cast<s64>(imm.value)};
         return ir.Imm32(static_cast<s32>(-(1LL << 19) + raw));

src/shader_recompiler/frontend/maxwell/translate/impl/load_constant.cpp

@@ -0,0 +1,85 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+enum class Mode : u64 {
+    Default,
+    IL,
+    IS,
+    ISL,
+};
+
+enum class Size : u64 {
+    U8,
+    S8,
+    U16,
+    S16,
+    B32,
+    B64,
+};
+
+std::pair<IR::U32, IR::U32> Slot(IR::IREmitter& ir, Mode mode, const IR::U32& imm_index,
+                                 const IR::U32& reg, const IR::U32& imm) {
+    switch (mode) {
+    case Mode::Default:
+        return {imm_index, ir.IAdd(reg, imm)};
+    default:
+        break;
+    }
+    throw NotImplementedException("Mode {}", mode);
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::LDC(u64 insn) {
+    union {
+        u64 raw;
+        BitField<0, 8, IR::Reg> dest_reg;
+        BitField<8, 8, IR::Reg> src_reg;
+        BitField<20, 16, s64> offset;
+        BitField<36, 5, u64> index;
+        BitField<44, 2, Mode> mode;
+        BitField<48, 3, Size> size;
+    } const ldc{insn};
+
+    const IR::U32 imm_index{ir.Imm32(static_cast<u32>(ldc.index))};
+    const IR::U32 reg{X(ldc.src_reg)};
+    const IR::U32 imm{ir.Imm32(static_cast<s32>(ldc.offset))};
+    const auto [index, offset]{Slot(ir, ldc.mode, imm_index, reg, imm)};
+    switch (ldc.size) {
+    case Size::U8:
+        X(ldc.dest_reg, ir.GetCbuf(index, offset, 8, false));
+        break;
+    case Size::S8:
+        X(ldc.dest_reg, ir.GetCbuf(index, offset, 8, true));
+        break;
+    case Size::U16:
+        X(ldc.dest_reg, ir.GetCbuf(index, offset, 16, false));
+        break;
+    case Size::S16:
+        X(ldc.dest_reg, ir.GetCbuf(index, offset, 16, true));
+        break;
+    case Size::B32:
+        X(ldc.dest_reg, ir.GetCbuf(index, offset, 32, false));
+        break;
+    case Size::B64: {
+        if (!IR::IsAligned(ldc.dest_reg, 2)) {
+            throw NotImplementedException("Unaligned destination register");
+        }
+        const IR::Value vector{ir.UnpackUint2x32(ir.GetCbuf(index, offset, 64, false))};
+        for (int i = 0; i < 2; ++i) {
+            X(ldc.dest_reg + i, IR::U32{ir.CompositeExtract(vector, i)});
+        }
+        break;
+    }
+    default:
+        throw NotImplementedException("Invalid size {}", ldc.size.Value());
+    }
+}
+
+} // namespace Shader::Maxwell

src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp

@@ -425,10 +425,6 @@ void TranslatorVisitor::LD(u64) {
     ThrowNotImplemented(Opcode::LD);
 }
 
-void TranslatorVisitor::LDC(u64) {
-    ThrowNotImplemented(Opcode::LDC);
-}
-
 void TranslatorVisitor::LDL(u64) {
     ThrowNotImplemented(Opcode::LDL);
 }

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -25,18 +25,13 @@ void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) {
 
 void VisitUsages(Info& info, IR::Inst& inst) {
     switch (inst.Opcode()) {
-    case IR::Opcode::WorkgroupId:
-        info.uses_workgroup_id = true;
-        break;
-    case IR::Opcode::LocalInvocationId:
-        info.uses_local_invocation_id = true;
-        break;
     case IR::Opcode::CompositeConstructF16x2:
     case IR::Opcode::CompositeConstructF16x3:
     case IR::Opcode::CompositeConstructF16x4:
     case IR::Opcode::CompositeExtractF16x2:
     case IR::Opcode::CompositeExtractF16x3:
     case IR::Opcode::CompositeExtractF16x4:
+    case IR::Opcode::SelectF16:
     case IR::Opcode::BitCastU16F16:
     case IR::Opcode::BitCastF16U16:
     case IR::Opcode::PackFloat2x16:
@@ -75,13 +70,139 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::FPTrunc64:
         info.uses_fp64 = true;
         break;
-    case IR::Opcode::GetCbuf:
+    default:
+        break;
+    }
+    switch (inst.Opcode()) {
+    case IR::Opcode::GetCbufU8:
+    case IR::Opcode::GetCbufS8:
+    case IR::Opcode::UndefU8:
+    case IR::Opcode::LoadGlobalU8:
+    case IR::Opcode::LoadGlobalS8:
+    case IR::Opcode::WriteGlobalU8:
+    case IR::Opcode::WriteGlobalS8:
+    case IR::Opcode::LoadStorageU8:
+    case IR::Opcode::LoadStorageS8:
+    case IR::Opcode::WriteStorageU8:
+    case IR::Opcode::WriteStorageS8:
+    case IR::Opcode::SelectU8:
+        info.uses_int8 = true;
+        break;
+    default:
+        break;
+    }
+    switch (inst.Opcode()) {
+    case IR::Opcode::GetCbufU16:
+    case IR::Opcode::GetCbufS16:
+    case IR::Opcode::UndefU16:
+    case IR::Opcode::LoadGlobalU16:
+    case IR::Opcode::LoadGlobalS16:
+    case IR::Opcode::WriteGlobalU16:
+    case IR::Opcode::WriteGlobalS16:
+    case IR::Opcode::LoadStorageU16:
+    case IR::Opcode::LoadStorageS16:
+    case IR::Opcode::WriteStorageU16:
+    case IR::Opcode::WriteStorageS16:
+    case IR::Opcode::SelectU16:
+    case IR::Opcode::BitCastU16F16:
+    case IR::Opcode::BitCastF16U16:
+    case IR::Opcode::ConvertS16F16:
+    case IR::Opcode::ConvertS16F32:
+    case IR::Opcode::ConvertS16F64:
+    case IR::Opcode::ConvertU16F16:
+    case IR::Opcode::ConvertU16F32:
+    case IR::Opcode::ConvertU16F64:
+        info.uses_int16 = true;
+        break;
+    default:
+        break;
+    }
+    switch (inst.Opcode()) {
+    case IR::Opcode::GetCbufU64:
+    case IR::Opcode::UndefU64:
+    case IR::Opcode::LoadGlobalU8:
+    case IR::Opcode::LoadGlobalS8:
+    case IR::Opcode::LoadGlobalU16:
+    case IR::Opcode::LoadGlobalS16:
+    case IR::Opcode::LoadGlobal32:
+    case IR::Opcode::LoadGlobal64:
+    case IR::Opcode::LoadGlobal128:
+    case IR::Opcode::WriteGlobalU8:
+    case IR::Opcode::WriteGlobalS8:
+    case IR::Opcode::WriteGlobalU16:
+    case IR::Opcode::WriteGlobalS16:
+    case IR::Opcode::WriteGlobal32:
+    case IR::Opcode::WriteGlobal64:
+    case IR::Opcode::WriteGlobal128:
+    case IR::Opcode::SelectU64:
+    case IR::Opcode::BitCastU64F64:
+    case IR::Opcode::BitCastF64U64:
+    case IR::Opcode::PackUint2x32:
+    case IR::Opcode::UnpackUint2x32:
+    case IR::Opcode::IAdd64:
+    case IR::Opcode::ISub64:
+    case IR::Opcode::INeg64:
+    case IR::Opcode::ShiftLeftLogical64:
+    case IR::Opcode::ShiftRightLogical64:
+    case IR::Opcode::ShiftRightArithmetic64:
+    case IR::Opcode::ConvertS64F16:
+    case IR::Opcode::ConvertS64F32:
+    case IR::Opcode::ConvertS64F64:
+    case IR::Opcode::ConvertU64F16:
+    case IR::Opcode::ConvertU64F32:
+    case IR::Opcode::ConvertU64F64:
+    case IR::Opcode::ConvertU64U32:
+    case IR::Opcode::ConvertU32U64:
+    case IR::Opcode::ConvertF16U64:
+    case IR::Opcode::ConvertF32U64:
+    case IR::Opcode::ConvertF64U64:
+        info.uses_int64 = true;
+        break;
+    default:
+        break;
+    }
+    switch (inst.Opcode()) {
+    case IR::Opcode::WorkgroupId:
+        info.uses_workgroup_id = true;
+        break;
+    case IR::Opcode::LocalInvocationId:
+        info.uses_local_invocation_id = true;
+        break;
+    case IR::Opcode::GetCbufU8:
+    case IR::Opcode::GetCbufS8:
+    case IR::Opcode::GetCbufU16:
+    case IR::Opcode::GetCbufS16:
+    case IR::Opcode::GetCbufU32:
+    case IR::Opcode::GetCbufF32:
+    case IR::Opcode::GetCbufU64: {
         if (const IR::Value index{inst.Arg(0)}; index.IsImmediate()) {
             AddConstantBufferDescriptor(info, index.U32(), 1);
         } else {
             throw NotImplementedException("Constant buffer with non-immediate index");
         }
+        switch (inst.Opcode()) {
+        case IR::Opcode::GetCbufU8:
+        case IR::Opcode::GetCbufS8:
+            info.used_constant_buffer_types |= IR::Type::U8;
             break;
+        case IR::Opcode::GetCbufU16:
+        case IR::Opcode::GetCbufS16:
+            info.used_constant_buffer_types |= IR::Type::U16;
+            break;
+        case IR::Opcode::GetCbufU32:
+            info.used_constant_buffer_types |= IR::Type::U32;
+            break;
+        case IR::Opcode::GetCbufF32:
+            info.used_constant_buffer_types |= IR::Type::F32;
+            break;
+        case IR::Opcode::GetCbufU64:
+            info.used_constant_buffer_types |= IR::Type::U64;
+            break;
+        default:
+            break;
+        }
+        break;
+    }
     case IR::Opcode::BindlessImageSampleImplicitLod:
     case IR::Opcode::BindlessImageSampleExplicitLod:
     case IR::Opcode::BindlessImageSampleDrefImplicitLod:

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:

src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp

@@ -203,7 +203,7 @@ std::optional<StorageBufferAddr> Track(IR::Block* block, const IR::Value& value,
         return std::nullopt;
     }
     const IR::Inst* const inst{value.InstRecursive()};
-    if (inst->Opcode() == IR::Opcode::GetCbuf) {
+    if (inst->Opcode() == IR::Opcode::GetCbufU32) {
         const IR::Value index{inst->Arg(0)};
         const IR::Value offset{inst->Arg(1)};
         if (!index.IsImmediate()) {

src/shader_recompiler/ir_opt/texture_pass.cpp

@@ -78,7 +78,7 @@ std::optional<ConstBufferAddr> Track(IR::Block* block, const IR::Value& value,
         return std::nullopt;
     }
     const IR::Inst* const inst{value.InstRecursive()};
-    if (inst->Opcode() == IR::Opcode::GetCbuf) {
+    if (inst->Opcode() == IR::Opcode::GetCbufU32) {
         const IR::Value index{inst->Arg(0)};
         const IR::Value offset{inst->Arg(1)};
         if (!index.IsImmediate()) {

src/shader_recompiler/shader_info.h

@@ -7,6 +7,7 @@
 #include <array>
 
 #include "common/common_types.h"
+#include "shader_recompiler/frontend/ir/type.h"
 
 #include <boost/container/small_vector.hpp>
 #include <boost/container/static_vector.hpp>
@@ -61,10 +62,15 @@ struct Info {
     bool uses_fp16_denorms_preserve{};
     bool uses_fp32_denorms_flush{};
     bool uses_fp32_denorms_preserve{};
+    bool uses_int8{};
+    bool uses_int16{};
+    bool uses_int64{};
     bool uses_image_1d{};
     bool uses_sampled_1d{};
     bool uses_sparse_residency{};
 
+    IR::Type used_constant_buffer_types{};
+
     u32 constant_buffer_mask{};
 
     boost::container::static_vector<ConstantBufferDescriptor, MAX_CBUFS>