shader_recompiler: Inline constant buffer impl

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -288,6 +288,10 @@ void EmitGetVcc(EmitContext& ctx) {
     UNREACHABLE_MSG("Unreachable instruction");
 }
 
+void EmitGetSccLo(EmitContext& ctx) {
+    UNREACHABLE_MSG("Unreachable instruction");
+}
+
 void EmitGetVccLo(EmitContext& ctx) {
     UNREACHABLE_MSG("Unreachable instruction");
 }
@@ -308,6 +312,10 @@ void EmitSetVcc(EmitContext& ctx) {
     UNREACHABLE_MSG("Unreachable instruction");
 }
 
+void EmitSetSccLo(EmitContext& ctx) {
+    UNREACHABLE_MSG("Unreachable instruction");
+}
+
 void EmitSetVccLo(EmitContext& ctx) {
     UNREACHABLE_MSG("Unreachable instruction");
 }

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -33,11 +33,13 @@ void EmitDeviceMemoryBarrier(EmitContext& ctx);
 void EmitGetScc(EmitContext& ctx);
 void EmitGetExec(EmitContext& ctx);
 void EmitGetVcc(EmitContext& ctx);
+void EmitGetSccLo(EmitContext& ctx);
 void EmitGetVccLo(EmitContext& ctx);
 void EmitGetVccHi(EmitContext& ctx);
 void EmitSetScc(EmitContext& ctx);
 void EmitSetExec(EmitContext& ctx);
 void EmitSetVcc(EmitContext& ctx);
+void EmitSetSccLo(EmitContext& ctx);
 void EmitSetVccLo(EmitContext& ctx);
 void EmitSetVccHi(EmitContext& ctx);
 void EmitPrologue(EmitContext& ctx);
@@ -245,6 +247,7 @@ Id EmitFPIsInf32(EmitContext& ctx, Id value);
 Id EmitFPIsInf64(EmitContext& ctx, Id value);
 Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
 Id EmitIAdd64(EmitContext& ctx, Id a, Id b);
+Id EmitIAddCary32(EmitContext& ctx, Id a, Id b);
 Id EmitISub32(EmitContext& ctx, Id a, Id b);
 Id EmitISub64(EmitContext& ctx, Id a, Id b);
 Id EmitSMulExt(EmitContext& ctx, Id a, Id b);

src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp

@@ -60,6 +60,10 @@ Id EmitIAdd64(EmitContext& ctx, Id a, Id b) {
     return ctx.OpIAdd(ctx.U64, a, b);
 }
 
+Id EmitIAddCary32(EmitContext& ctx, Id a, Id b) {
+    return ctx.OpIAddCarry(ctx.full_result_u32x2, a, b);
+}
+
 Id EmitISub32(EmitContext& ctx, Id a, Id b) {
     return ctx.OpISub(ctx.U32[1], a, b);
 }

src/shader_recompiler/frontend/structured_control_flow.cpp

@@ -633,7 +633,7 @@ private:
                 if (!stmt.block->is_dummy) {
                     const u32 start = stmt.block->begin_index;
                     const u32 size = stmt.block->end_index - start + 1;
-                    Translate(current_block, inst_list.subspan(start, size), info);
+                    Translate(current_block, stmt.block->begin, inst_list.subspan(start, size), info);
                 }
                 break;
             }

src/shader_recompiler/frontend/translate/scalar_alu.cpp

@@ -318,4 +318,16 @@ void Translator::S_SUB_U32(const GcnInst& inst) {
     ir.SetScc(ir.Imm1(false));
 }
 
+void Translator::S_GETPC_B64(u32 pc, const GcnInst& inst) {
+    // This only really exists to let resource tracking pass know
+    // there is an inline cbuf.
+    SetDst(inst.dst[0], ir.Imm32(pc));
+}
+
+void Translator::S_ADDC_U32(const GcnInst& inst) {
+    const IR::U32 src0{GetSrc(inst.src[0])};
+    const IR::U32 src1{GetSrc(inst.src[1])};
+    SetDst(inst.dst[0], ir.IAdd(ir.IAdd(src0, src1), ir.GetSccLo()));
+}
+
 } // namespace Shader::Gcn

src/shader_recompiler/frontend/translate/scalar_memory.cpp

@@ -30,7 +30,8 @@ void Translator::S_BUFFER_LOAD_DWORD(int num_dwords, const GcnInst& inst) {
         }
         return ir.ShiftRightLogical(ir.GetScalarReg(IR::ScalarReg(smrd.offset)), ir.Imm32(2));
     }();
-    const IR::Value vsharp = ir.GetScalarReg(sbase);
+    const IR::Value vsharp = ir.CompositeConstruct(ir.GetScalarReg(sbase), ir.GetScalarReg(sbase + 1),
+                                                   ir.GetScalarReg(sbase + 2), ir.GetScalarReg(sbase + 3));
     IR::ScalarReg dst_reg{inst.dst[0].code};
     for (u32 i = 0; i < num_dwords; i++) {
         const IR::U32 index = ir.IAdd(dword_offset, ir.Imm32(i));

src/shader_recompiler/frontend/translate/translate.cpp

@@ -236,7 +236,7 @@ void Translator::EmitFetch(const GcnInst& inst) {
     }
 }
 
-void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info) {
+void Translate(IR::Block* block, u32 block_base, std::span<const GcnInst> inst_list, Info& info) {
     if (inst_list.empty()) {
         return;
     }
@@ -833,6 +833,9 @@ void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info)
         case Opcode::S_ADD_U32:
             translator.S_ADD_U32(inst);
             break;
+        case Opcode::S_ADDC_U32:
+            translator.S_ADDC_U32(inst);
+            break;
         case Opcode::S_SUB_U32:
         case Opcode::S_SUB_I32:
             translator.S_SUB_U32(inst);
@@ -878,6 +881,9 @@ void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info)
         case Opcode::V_READFIRSTLANE_B32:
             translator.V_READFIRSTLANE_B32(inst);
             break;
+        case Opcode::S_GETPC_B64:
+            translator.S_GETPC_B64(block_base, inst);
+            break;
         case Opcode::S_NOP:
         case Opcode::S_CBRANCH_EXECZ:
         case Opcode::S_CBRANCH_SCC0:
@@ -895,6 +901,7 @@ void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info)
                       magic_enum::enum_name(inst.opcode), opcode);
             info.translation_failed = true;
         }
+        block_base += inst.length;
     }
 }
 

src/shader_recompiler/frontend/translate/translate.h

@@ -80,6 +80,8 @@ public:
     void S_BREV_B32(const GcnInst& inst);
     void S_ADD_U32(const GcnInst& inst);
     void S_SUB_U32(const GcnInst& inst);
+    void S_GETPC_B64(u32 pc, const GcnInst& inst);
+    void S_ADDC_U32(const GcnInst& inst);
 
     // Scalar Memory
     void S_LOAD_DWORD(int num_dwords, const GcnInst& inst);
@@ -192,6 +194,6 @@ private:
     static std::array<bool, IR::NumScalarRegs> exec_contexts;
 };
 
-void Translate(IR::Block* block, std::span<const GcnInst> inst_list, Info& info);
+void Translate(IR::Block* block, u32 block_base, std::span<const GcnInst> inst_list, Info& info);
 
 } // namespace Shader::Gcn

src/shader_recompiler/frontend/translate/vector_memory.cpp

@@ -250,7 +250,9 @@ void Translator::BUFFER_LOAD_FORMAT(u32 num_dwords, bool is_typed, const GcnInst
         info.nfmt.Assign(static_cast<AmdGpu::NumberFormat>(mtbuf.nfmt));
     }
 
-    const IR::Value value = ir.LoadBuffer(num_dwords, ir.GetScalarReg(sharp), address, info);
+    const IR::Value handle = ir.CompositeConstruct(ir.GetScalarReg(sharp), ir.GetScalarReg(sharp + 1),
+                                                   ir.GetScalarReg(sharp + 2), ir.GetScalarReg(sharp + 3));
+    const IR::Value value = ir.LoadBuffer(num_dwords, handle, address, info);
     const IR::VectorReg dst_reg{inst.src[1].code};
     if (num_dwords == 1) {
         ir.SetVectorReg(dst_reg, IR::F32{value});
@@ -309,7 +311,9 @@ void Translator::BUFFER_STORE_FORMAT(u32 num_dwords, bool is_typed, const GcnIns
                                       ir.GetVectorReg<Shader::IR::F32>(src_reg + 3));
         break;
     }
-    ir.StoreBuffer(num_dwords, ir.GetScalarReg(sharp), address, value, info);
+    const IR::Value handle = ir.CompositeConstruct(ir.GetScalarReg(sharp), ir.GetScalarReg(sharp + 1),
+                                                   ir.GetScalarReg(sharp + 2), ir.GetScalarReg(sharp + 3));
+    ir.StoreBuffer(num_dwords, handle, address, value, info);
 }
 
 void Translator::IMAGE_GET_LOD(const GcnInst& inst) {

src/shader_recompiler/ir/ir_emitter.cpp

@@ -212,6 +212,10 @@ U1 IREmitter::GetVcc() {
     return Inst<U1>(Opcode::GetVcc);
 }
 
+U32 IREmitter::GetSccLo() {
+    return Inst<U32>(Opcode::GetSccLo);
+}
+
 U32 IREmitter::GetVccLo() {
     return Inst<U32>(Opcode::GetVccLo);
 }
@@ -232,6 +236,10 @@ void IREmitter::SetVcc(const U1& value) {
     Inst(Opcode::SetVcc, value);
 }
 
+void IREmitter::SetSccLo(const U32& value) {
+    Inst(Opcode::SetSccLo, value);
+}
+
 void IREmitter::SetVccLo(const U32& value) {
     Inst(Opcode::SetVccLo, value);
 }
@@ -898,6 +906,18 @@ U32U64 IREmitter::IAdd(const U32U64& a, const U32U64& b) {
     }
 }
 
+Value IREmitter::IAddCary(const U32& a, const U32& b) {
+    if (a.Type() != b.Type()) {
+        UNREACHABLE_MSG("Mismatching types {} and {}", a.Type(), b.Type());
+    }
+    switch (a.Type()) {
+    case Type::U32:
+        return Inst<U32>(Opcode::IAddCary32, a, b);
+    default:
+        ThrowInvalidType(a.Type());
+    }
+}
+
 U32U64 IREmitter::ISub(const U32U64& a, const U32U64& b) {
     if (a.Type() != b.Type()) {
         UNREACHABLE_MSG("Mismatching types {} and {}", a.Type(), b.Type());

src/shader_recompiler/ir/ir_emitter.h

@@ -64,11 +64,13 @@ public:
     [[nodiscard]] U1 GetScc();
     [[nodiscard]] U1 GetExec();
     [[nodiscard]] U1 GetVcc();
+    [[nodiscard]] U32 GetSccLo();
     [[nodiscard]] U32 GetVccLo();
     [[nodiscard]] U32 GetVccHi();
     void SetScc(const U1& value);
     void SetExec(const U1& value);
     void SetVcc(const U1& value);
+    void SetSccLo(const U32& value);
     void SetVccLo(const U32& value);
     void SetVccHi(const U32& value);
 
@@ -151,8 +153,9 @@ public:
     [[nodiscard]] F32F64 FPMin(const F32F64& lhs, const F32F64& rhs);
 
     [[nodiscard]] U32U64 IAdd(const U32U64& a, const U32U64& b);
+    [[nodiscard]] Value IAddCary(const U32& a, const U32& b);
     [[nodiscard]] U32U64 ISub(const U32U64& a, const U32U64& b);
-    [[nodiscard]] IR::Value IMulExt(const U32& a, const U32& b, bool is_signed = false);
+    [[nodiscard]] Value IMulExt(const U32& a, const U32& b, bool is_signed = false);
     [[nodiscard]] U32 IMul(const U32& a, const U32& b);
     [[nodiscard]] U32 IDiv(const U32& a, const U32& b, bool is_signed = false);
     [[nodiscard]] U32U64 INeg(const U32U64& value);

src/shader_recompiler/ir/opcodes.inc

@@ -56,11 +56,13 @@ OPCODE(SetAttribute,                                        Void,           Attr
 OPCODE(GetScc,                                             U1,             Void,                                                                            )
 OPCODE(GetExec,                                            U1,             Void,                                                                            )
 OPCODE(GetVcc,                                             U1,             Void,                                                                            )
+OPCODE(GetSccLo,                                           U32,            Void,                                                                            )
 OPCODE(GetVccLo,                                           U32,            Void,                                                                            )
 OPCODE(GetVccHi,                                           U32,            Void,                                                                            )
 OPCODE(SetScc,                                             Void,           U1,                                                                              )
 OPCODE(SetExec,                                            Void,           U1,                                                                              )
 OPCODE(SetVcc,                                             Void,           U1,                                                                              )
+OPCODE(SetSccLo,                                           Void,           U32,                                                                             )
 OPCODE(SetVccLo,                                           Void,           U32,                                                                             )
 OPCODE(SetVccHi,                                           Void,           U32,                                                                             )
 
@@ -216,6 +218,7 @@ OPCODE(FPIsInf64,                                           U1,             F64,
 // Integer operations
 OPCODE(IAdd32,                                              U32,            U32,            U32,                                                            )
 OPCODE(IAdd64,                                              U64,            U64,            U64,                                                            )
+OPCODE(IAddCary32,                                          U32x2,          U32,            U32,                                                            )
 OPCODE(ISub32,                                              U32,            U32,            U32,                                                            )
 OPCODE(ISub64,                                              U64,            U64,            U64,                                                            )
 OPCODE(IMul32,                                              U32,            U32,            U32,                                                            )

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -138,7 +138,8 @@ public:
     u32 Add(const BufferResource& desc) {
         const u32 index{Add(buffer_resources, desc, [&desc](const auto& existing) {
             return desc.sgpr_base == existing.sgpr_base &&
-                   desc.dword_offset == existing.dword_offset;
+                   desc.dword_offset == existing.dword_offset &&
+                   desc.inline_cbuf == existing.inline_cbuf;
         })};
         auto& buffer = buffer_resources[index];
         ASSERT(buffer.stride == desc.stride && buffer.num_records == desc.num_records);
@@ -219,13 +220,55 @@ SharpLocation TrackSharp(const IR::Inst* inst) {
     };
 }
 
+static constexpr size_t MaxUboSize = 65536;
+
+s32 TryHandleInlineCbuf(IR::Inst& inst, Info& info, Descriptors& descriptors, AmdGpu::Buffer& cbuf) {
+    /**
+     * Assert for the following pattern
+     * s_getpc_b64     s[32:33]
+     * s_add_u32       s32, <const>, s32
+     * s_addc_u32      s33, 0, s33
+     * s_mov_b32       s35, <const>
+     * s_movk_i32      s34, <const>
+     * buffer_load_format_xyz v[8:10], v1, s[32:35], 0 ...
+     **/
+    IR::Inst* handle = inst.Arg(0).InstRecursive();
+    IR::Inst* p0 = handle->Arg(0).InstRecursive();
+    if (p0->GetOpcode() != IR::Opcode::IAdd32 || !p0->Arg(0).IsImmediate()) {
+        return -1;
+    }
+    IR::Inst* p1 = handle->Arg(1).InstRecursive();
+    if (p1->GetOpcode() != IR::Opcode::IAdd32) {
+        return -1;
+    }
+    if (!handle->Arg(3).IsImmediate() || !handle->Arg(2).IsImmediate()) {
+        return -1;
+    }
+    // We have found this pattern. Build the sharp and assign a binding to it.
+    cbuf.raw0 = info.pgm_base + p0->Arg(0).U32() + p0->Arg(1).U32();
+    cbuf.num_records = handle->Arg(2).U32();
+    cbuf.raw11 = handle->Arg(3).U32();
+    return descriptors.Add(BufferResource{
+        .sgpr_base = std::numeric_limits<u32>::max(),
+        .dword_offset = 0,
+        .stride = cbuf.GetStride(),
+        .num_records = u32(cbuf.num_records),
+        .used_types = BufferDataType(inst),
+        .inline_cbuf = cbuf,
+        .is_storage = IsBufferStore(inst) || cbuf.GetSize() > MaxUboSize,
+    });
+}
+
 void PatchBufferInstruction(IR::Block& block, IR::Inst& inst, Info& info,
                             Descriptors& descriptors) {
-    static constexpr size_t MaxUboSize = 65536;
+    s32 binding{};
-    IR::Inst* producer = inst.Arg(0).InstRecursive();
+    AmdGpu::Buffer buffer;
+    if (binding = TryHandleInlineCbuf(inst, info, descriptors, buffer); binding == -1) {
+        IR::Inst* handle = inst.Arg(0).InstRecursive();
+        IR::Inst* producer = handle->Arg(0).InstRecursive();
         const auto sharp = TrackSharp(producer);
-    const auto buffer = info.ReadUd<AmdGpu::Buffer>(sharp.sgpr_base, sharp.dword_offset);
+        buffer = info.ReadUd<AmdGpu::Buffer>(sharp.sgpr_base, sharp.dword_offset);
-    const u32 binding = descriptors.Add(BufferResource{
+        binding = descriptors.Add(BufferResource{
             .sgpr_base = sharp.sgpr_base,
             .dword_offset = sharp.dword_offset,
             .stride = buffer.GetStride(),
@@ -233,6 +276,8 @@ void PatchBufferInstruction(IR::Block& block, IR::Inst& inst, Info& info,
             .used_types = BufferDataType(inst),
             .is_storage = IsBufferStore(inst) || buffer.GetSize() > MaxUboSize,
         });
+    }
+
     const auto inst_info = inst.Flags<IR::BufferInstInfo>();
     IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
     // Replace handle with binding index in buffer resource list.
@@ -240,7 +285,8 @@ void PatchBufferInstruction(IR::Block& block, IR::Inst& inst, Info& info,
     ASSERT(!buffer.swizzle_enable && !buffer.add_tid_enable);
     if (inst_info.is_typed) {
         ASSERT(inst_info.nfmt == AmdGpu::NumberFormat::Float &&
-               inst_info.dmft == AmdGpu::DataFormat::Format32_32_32_32);
+               (inst_info.dmft == AmdGpu::DataFormat::Format32_32_32_32 ||
+                inst_info.dmft == AmdGpu::DataFormat::Format32_32_32));
     }
     if (inst.GetOpcode() == IR::Opcode::ReadConstBuffer ||
         inst.GetOpcode() == IR::Opcode::ReadConstBufferU32) {

src/shader_recompiler/ir/passes/ssa_rewrite_pass.cpp

@@ -32,6 +32,7 @@ struct SccFlagTag : FlagTag {};
 struct ExecFlagTag : FlagTag {};
 struct VccFlagTag : FlagTag {};
 struct VccLoTag : FlagTag {};
+struct SccLoTag : FlagTag {};
 struct VccHiTag : FlagTag {};
 
 struct GotoVariable : FlagTag {
@@ -44,7 +45,7 @@ struct GotoVariable : FlagTag {
 };
 
 using Variant = std::variant<IR::ScalarReg, IR::VectorReg, GotoVariable, SccFlagTag, ExecFlagTag,
-                             VccFlagTag, VccLoTag, VccHiTag>;
+                             VccFlagTag, SccLoTag, VccLoTag, VccHiTag>;
 using ValueMap = std::unordered_map<IR::Block*, IR::Value>;
 
 struct DefTable {
@@ -83,6 +84,13 @@ struct DefTable {
         exec_flag.insert_or_assign(block, value);
     }
 
+    const IR::Value& Def(IR::Block* block, SccLoTag) {
+        return scc_lo_flag[block];
+    }
+    void SetDef(IR::Block* block, SccLoTag, const IR::Value& value) {
+        scc_lo_flag.insert_or_assign(block, value);
+    }
+
     const IR::Value& Def(IR::Block* block, VccLoTag) {
         return vcc_lo_flag[block];
     }
@@ -108,6 +116,7 @@ struct DefTable {
     ValueMap scc_flag;
     ValueMap exec_flag;
     ValueMap vcc_flag;
+    ValueMap scc_lo_flag;
     ValueMap vcc_lo_flag;
     ValueMap vcc_hi_flag;
 };
@@ -124,6 +133,10 @@ IR::Opcode UndefOpcode(const VccLoTag&) noexcept {
     return IR::Opcode::UndefU32;
 }
 
+IR::Opcode UndefOpcode(const SccLoTag&) noexcept {
+    return IR::Opcode::UndefU32;
+}
+
 IR::Opcode UndefOpcode(const VccHiTag&) noexcept {
     return IR::Opcode::UndefU32;
 }
@@ -321,6 +334,9 @@ void VisitInst(Pass& pass, IR::Block* block, IR::Inst& inst) {
     case IR::Opcode::SetVcc:
         pass.WriteVariable(VccFlagTag{}, block, inst.Arg(0));
         break;
+    case IR::Opcode::SetSccLo:
+        pass.WriteVariable(SccLoTag{}, block, inst.Arg(0));
+        break;
     case IR::Opcode::SetVccLo:
         pass.WriteVariable(VccLoTag{}, block, inst.Arg(0));
         break;
@@ -350,6 +366,9 @@ void VisitInst(Pass& pass, IR::Block* block, IR::Inst& inst) {
     case IR::Opcode::GetVcc:
         inst.ReplaceUsesWith(pass.ReadVariable(VccFlagTag{}, block));
         break;
+    case IR::Opcode::GetSccLo:
+        inst.ReplaceUsesWith(pass.ReadVariable(SccLoTag{}, block));
+        break;
     case IR::Opcode::GetVccLo:
         inst.ReplaceUsesWith(pass.ReadVariable(VccLoTag{}, block));
         break;

src/shader_recompiler/runtime_info.h

@@ -4,7 +4,6 @@
 #pragma once
 
 #include <span>
-#include <vector>
 #include <boost/container/static_vector.hpp>
 #include "common/assert.h"
 #include "common/types.h"
@@ -69,15 +68,18 @@ enum class VsOutput : u32 {
 };
 using VsOutputMap = std::array<VsOutput, 4>;
 
+struct Info;
+
 struct BufferResource {
     u32 sgpr_base;
     u32 dword_offset;
     u32 stride;
     u32 num_records;
     IR::Type used_types;
+    AmdGpu::Buffer inline_cbuf;
     bool is_storage;
 
-    auto operator<=>(const BufferResource&) const = default;
+    constexpr AmdGpu::Buffer GetVsharp(const Info& info) const noexcept;
 };
 using BufferResourceList = boost::container::static_vector<BufferResource, 16>;
 
@@ -162,6 +164,7 @@ struct Info {
     std::span<const u32> user_data;
     Stage stage;
 
+    uintptr_t pgm_base{};
     u32 shared_memory_size{};
     bool uses_group_quad{};
     bool uses_shared_u8{};
@@ -180,6 +183,10 @@ struct Info {
     }
 };
 
+constexpr AmdGpu::Buffer BufferResource::GetVsharp(const Info& info) const noexcept {
+    return inline_cbuf ? inline_cbuf : info.ReadUd<AmdGpu::Buffer>(sgpr_base, dword_offset);
+}
+
 } // namespace Shader
 
 template <>

src/video_core/amdgpu/liverpool.h

@@ -85,14 +85,14 @@ struct Liverpool {
         } settings;
         UserData user_data;
 
-        template <typename T = u8>
+        template <typename T = u8*>
-        const T* Address() const {
+        const T Address() const {
             const uintptr_t addr = uintptr_t(address_hi) << 40 | uintptr_t(address_lo) << 8;
-            return reinterpret_cast<const T*>(addr);
+            return reinterpret_cast<const T>(addr);
         }
 
         std::span<const u32> Code() const {
-            const u32* code = Address<u32>();
+            const u32* code = Address<u32*>();
             BinaryInfo bininfo;
             std::memcpy(&bininfo, code + (code[1] + 1) * 2, sizeof(bininfo));
             const u32 num_dwords = bininfo.length / sizeof(u32);
@@ -128,10 +128,10 @@ struct Liverpool {
         INSERT_PADDING_WORDS(0x2A);
         UserData user_data;
 
-        template <typename T = u8>
+        template <typename T = u8*>
-        const T* Address() const {
+        const T Address() const {
             const uintptr_t addr = uintptr_t(address_hi) << 40 | uintptr_t(address_lo) << 8;
-            return reinterpret_cast<const T*>(addr);
+            return reinterpret_cast<const T>(addr);
         }
 
         u32 SharedMemSize() const noexcept {
@@ -140,7 +140,7 @@ struct Liverpool {
         }
 
         std::span<const u32> Code() const {
-            const u32* code = Address<u32>();
+            const u32* code = Address<u32*>();
             BinaryInfo bininfo;
             std::memcpy(&bininfo, code + (code[1] + 1) * 2, sizeof(bininfo));
             const u32 num_dwords = bininfo.length / sizeof(u32);
@@ -150,7 +150,7 @@ struct Liverpool {
 
     template <typename Shader>
     static constexpr auto* GetBinaryInfo(const Shader& sh) {
-        const auto* code = sh.template Address<u32>();
+        const auto* code = sh.template Address<u32*>();
         const auto* bininfo = std::bit_cast<const BinaryInfo*>(code + (code[1] + 1) * 2);
         ASSERT_MSG(bininfo->Valid(), "Invalid shader binary header");
         return bininfo;

src/video_core/amdgpu/resource.h

@@ -22,6 +22,7 @@ enum class CompSwizzle : u32 {
 // Table 8.5 Buffer Resource Descriptor [Sea Islands Series Instruction Set Architecture]
 struct Buffer {
     union {
+        u64 raw0;
         BitField<0, 44, u64> base_address;
         BitField<48, 14, u64> stride;
         BitField<62, 1, u64> cache_swizzle;
@@ -29,6 +30,7 @@ struct Buffer {
     };
     u32 num_records;
     union {
+        u32 raw11;
         BitField<0, 3, u32> dst_sel_x;
         BitField<3, 3, u32> dst_sel_y;
         BitField<6, 3, u32> dst_sel_z;
@@ -41,6 +43,14 @@ struct Buffer {
         BitField<23, 1, u32> add_tid_enable;
     };
 
+    operator bool() const noexcept {
+        return base_address != 0;
+    }
+
+    bool operator==(const Buffer& other) const noexcept {
+        return std::memcmp(this, &other, sizeof(Buffer)) == 0;
+    }
+
     CompSwizzle GetSwizzle(u32 comp) const noexcept {
         return static_cast<CompSwizzle>((dst_sel.Value() >> (comp * 3)) & 0x7);
     }

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -91,7 +91,7 @@ bool ComputePipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& s
     u32 binding{};
 
     for (const auto& buffer : info.buffers) {
-        const auto vsharp = info.ReadUd<AmdGpu::Buffer>(buffer.sgpr_base, buffer.dword_offset);
+        const auto vsharp = buffer.GetVsharp(info);
         const u32 size = vsharp.GetSize();
         const VAddr address = vsharp.base_address.Value();
         texture_cache.OnCpuWrite(address);

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -326,7 +326,7 @@ void GraphicsPipeline::BindResources(Core::MemoryManager* memory, StreamBuffer&
 
     for (const auto& stage : stages) {
         for (const auto& buffer : stage.buffers) {
-            const auto vsharp = stage.ReadUd<AmdGpu::Buffer>(buffer.sgpr_base, buffer.dword_offset);
+            const auto vsharp = buffer.GetVsharp(stage);
             const VAddr address = vsharp.base_address.Value();
             const u32 size = vsharp.GetSize();
             const u32 offset = staging.Copy(address, size,

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -198,7 +198,7 @@ void PipelineCache::RefreshGraphicsKey() {
 
     for (u32 i = 0; i < MaxShaderStages; i++) {
         auto* pgm = regs.ProgramForStage(i);
-        if (!pgm || !pgm->Address<u32>()) {
+        if (!pgm || !pgm->Address<u32*>()) {
             key.stage_hashes[i] = 0;
             continue;
         }
@@ -248,17 +248,14 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline() {
             DumpShader(code, hash, stage, "bin");
         }
 
-        if (hash == 0xcafe3773 || hash == 0xc6602df2) {
-            return nullptr;
-        }
-
         block_pool.ReleaseContents();
         inst_pool.ReleaseContents();
 
         // Recompile shader to IR.
         try {
             LOG_INFO(Render_Vulkan, "Compiling {} shader {:#x}", stage, hash);
-            const Shader::Info info = MakeShaderInfo(stage, pgm->user_data, regs);
+            Shader::Info info = MakeShaderInfo(stage, pgm->user_data, regs);
+            info.pgm_base = pgm->Address<uintptr_t>();
             programs[i] = Shader::TranslateProgram(inst_pool, block_pool, code, std::move(info));
 
             // Compile IR to SPIR-V
@@ -296,8 +293,9 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline() {
     // Recompile shader to IR.
     try {
         LOG_INFO(Render_Vulkan, "Compiling cs shader {:#x}", compute_key);
-        const Shader::Info info =
+        Shader::Info info =
             MakeShaderInfo(Shader::Stage::Compute, cs_pgm.user_data, liverpool->regs);
+        info.pgm_base = cs_pgm.Address<uintptr_t>();
         auto program = Shader::TranslateProgram(inst_pool, block_pool, code, std::move(info));
 
         // Compile IR to SPIR-V