shader: Partial implementation of LDC

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()) {