shader: Implement TLD4 and TLD4_B

3 files changed, 29 insertions, 1 deletions

diff --git a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
index 61cc314c7..6fe06fda8 100644
--- a/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
+++ b/src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
@@ -352,14 +352,20 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::BindlessImageSampleExplicitLod:
     case IR::Opcode::BindlessImageSampleDrefImplicitLod:
     case IR::Opcode::BindlessImageSampleDrefExplicitLod:
+    case IR::Opcode::BindlessImageGather:
+    case IR::Opcode::BindlessImageGatherDref:
     case IR::Opcode::BoundImageSampleImplicitLod:
     case IR::Opcode::BoundImageSampleExplicitLod:
     case IR::Opcode::BoundImageSampleDrefImplicitLod:
     case IR::Opcode::BoundImageSampleDrefExplicitLod:
+    case IR::Opcode::BoundImageGather:
+    case IR::Opcode::BoundImageGatherDref:
     case IR::Opcode::ImageSampleImplicitLod:
     case IR::Opcode::ImageSampleExplicitLod:
     case IR::Opcode::ImageSampleDrefImplicitLod:
-    case IR::Opcode::ImageSampleDrefExplicitLod: {
+    case IR::Opcode::ImageSampleDrefExplicitLod:
+    case IR::Opcode::ImageGather:
+    case IR::Opcode::ImageGatherDref: {
         const TextureType type{inst.Flags<IR::TextureInstInfo>().type};
         info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D ||
                                 type == TextureType::Shadow1D || type == TextureType::ShadowArray1D;
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:
diff --git a/src/shader_recompiler/ir_opt/texture_pass.cpp b/src/shader_recompiler/ir_opt/texture_pass.cpp
index 2c8164b8a..454ac3e71 100644
--- a/src/shader_recompiler/ir_opt/texture_pass.cpp
+++ b/src/shader_recompiler/ir_opt/texture_pass.cpp
@@ -45,6 +45,12 @@ IR::Opcode IndexedInstruction(const IR::Inst& inst) {
     case IR::Opcode::BoundImageSampleDrefExplicitLod:
     case IR::Opcode::BindlessImageSampleDrefExplicitLod:
         return IR::Opcode::ImageSampleDrefExplicitLod;
+    case IR::Opcode::BindlessImageGather:
+    case IR::Opcode::BoundImageGather:
+        return IR::Opcode::ImageGather;
+    case IR::Opcode::BindlessImageGatherDref:
+    case IR::Opcode::BoundImageGatherDref:
+        return IR::Opcode::ImageGatherDref;
     default:
         return IR::Opcode::Void;
     }
@@ -56,11 +62,15 @@ bool IsBindless(const IR::Inst& inst) {
     case IR::Opcode::BindlessImageSampleExplicitLod:
     case IR::Opcode::BindlessImageSampleDrefImplicitLod:
     case IR::Opcode::BindlessImageSampleDrefExplicitLod:
+    case IR::Opcode::BindlessImageGather:
+    case IR::Opcode::BindlessImageGatherDref:
         return true;
     case IR::Opcode::BoundImageSampleImplicitLod:
     case IR::Opcode::BoundImageSampleExplicitLod:
     case IR::Opcode::BoundImageSampleDrefImplicitLod:
     case IR::Opcode::BoundImageSampleDrefExplicitLod:
+    case IR::Opcode::BoundImageGather:
+    case IR::Opcode::BoundImageGatherDref:
         return false;
     default:
         throw InvalidArgument("Invalid opcode {}", inst.Opcode());