shader: Implement SHFL

5 files changed, 53 insertions, 2 deletions

diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index 6280c08f6..418b7f5ac 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -374,6 +374,10 @@ U1 IREmitter::GetSparseFromOp(const Value& op) {
     return Inst<U1>(Opcode::GetSparseFromOp, op);
 }
 
+U1 IREmitter::GetInBoundsFromOp(const Value& op) {
+    return Inst<U1>(Opcode::GetInBoundsFromOp, op);
+}
+
 F16F32F64 IREmitter::FPAdd(const F16F32F64& a, const F16F32F64& b, FpControl control) {
     if (a.Type() != b.Type()) {
         throw InvalidArgument("Mismatching types {} and {}", a.Type(), b.Type());
@@ -1486,4 +1490,23 @@ U32 IREmitter::SubgroupBallot(const U1& value) {
     return Inst<U32>(Opcode::SubgroupBallot, value);
 }
 
+U32 IREmitter::ShuffleIndex(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                            const IR::U32& seg_mask) {
+    return Inst<U32>(Opcode::ShuffleIndex, value, index, clamp, seg_mask);
+}
+
+U32 IREmitter::ShuffleUp(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                         const IR::U32& seg_mask) {
+    return Inst<U32>(Opcode::ShuffleUp, value, index, clamp, seg_mask);
+}
+
+U32 IREmitter::ShuffleDown(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                           const IR::U32& seg_mask) {
+    return Inst<U32>(Opcode::ShuffleDown, value, index, clamp, seg_mask);
+}
+
+U32 IREmitter::ShuffleButterfly(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                                const IR::U32& seg_mask) {
+    return Inst<U32>(Opcode::ShuffleButterfly, value, index, clamp, seg_mask);
+}
 } // namespace Shader::IR
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h
index ebbda78a9..64738735e 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.h
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.h
@@ -104,6 +104,7 @@ public:
     [[nodiscard]] U1 GetCarryFromOp(const Value& op);
     [[nodiscard]] U1 GetOverflowFromOp(const Value& op);
     [[nodiscard]] U1 GetSparseFromOp(const Value& op);
+    [[nodiscard]] U1 GetInBoundsFromOp(const Value& op);
 
     [[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2);
     [[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2, const Value& e3);
@@ -147,7 +148,8 @@ public:
     [[nodiscard]] F32F64 FPRecipSqrt(const F32F64& value);
     [[nodiscard]] F32 FPSqrt(const F32& value);
     [[nodiscard]] F16F32F64 FPSaturate(const F16F32F64& value);
-    [[nodiscard]] F16F32F64 FPClamp(const F16F32F64& value, const F16F32F64& min_value, const F16F32F64& max_value);
+    [[nodiscard]] F16F32F64 FPClamp(const F16F32F64& value, const F16F32F64& min_value,
+                                    const F16F32F64& max_value);
     [[nodiscard]] F16F32F64 FPRoundEven(const F16F32F64& value, FpControl control = {});
     [[nodiscard]] F16F32F64 FPFloor(const F16F32F64& value, FpControl control = {});
     [[nodiscard]] F16F32F64 FPCeil(const F16F32F64& value, FpControl control = {});
@@ -242,6 +244,14 @@ public:
     [[nodiscard]] U1 VoteAny(const U1& value);
     [[nodiscard]] U1 VoteEqual(const U1& value);
     [[nodiscard]] U32 SubgroupBallot(const U1& value);
+    [[nodiscard]] U32 ShuffleIndex(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                                   const IR::U32& seg_mask);
+    [[nodiscard]] U32 ShuffleUp(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                                const IR::U32& seg_mask);
+    [[nodiscard]] U32 ShuffleDown(const IR::U32& value, const IR::U32& index, const IR::U32& clamp,
+                                  const IR::U32& seg_mask);
+    [[nodiscard]] U32 ShuffleButterfly(const IR::U32& value, const IR::U32& index,
+                                       const IR::U32& clamp, const IR::U32& seg_mask);
 
 private:
     IR::Block::iterator insertion_point;
diff --git a/src/shader_recompiler/frontend/ir/microinstruction.cpp b/src/shader_recompiler/frontend/ir/microinstruction.cpp
index ba3968056..be8eb4d4c 100644
--- a/src/shader_recompiler/frontend/ir/microinstruction.cpp
+++ b/src/shader_recompiler/frontend/ir/microinstruction.cpp
@@ -89,6 +89,7 @@ bool Inst::IsPseudoInstruction() const noexcept {
     case Opcode::GetCarryFromOp:
     case Opcode::GetOverflowFromOp:
     case Opcode::GetSparseFromOp:
+    case Opcode::GetInBoundsFromOp:
         return true;
     default:
         return false;
@@ -123,6 +124,9 @@ Inst* Inst::GetAssociatedPseudoOperation(IR::Opcode opcode) {
     case Opcode::GetSparseFromOp:
         CheckPseudoInstruction(associated_insts->sparse_inst, Opcode::GetSparseFromOp);
         return associated_insts->sparse_inst;
+    case Opcode::GetInBoundsFromOp:
+        CheckPseudoInstruction(associated_insts->in_bounds_inst, Opcode::GetInBoundsFromOp);
+        return associated_insts->in_bounds_inst;
     default:
         throw InvalidArgument("{} is not a pseudo-instruction", opcode);
     }
@@ -262,6 +266,10 @@ void Inst::Use(const Value& value) {
         AllocAssociatedInsts(assoc_inst);
         SetPseudoInstruction(assoc_inst->sparse_inst, this);
         break;
+    case Opcode::GetInBoundsFromOp:
+        AllocAssociatedInsts(assoc_inst);
+        SetPseudoInstruction(assoc_inst->in_bounds_inst, this);
+        break;
     default:
         break;
     }
@@ -289,6 +297,10 @@ void Inst::UndoUse(const Value& value) {
         AllocAssociatedInsts(assoc_inst);
         RemovePseudoInstruction(assoc_inst->overflow_inst, Opcode::GetOverflowFromOp);
         break;
+    case Opcode::GetInBoundsFromOp:
+        AllocAssociatedInsts(assoc_inst);
+        RemovePseudoInstruction(assoc_inst->in_bounds_inst, Opcode::GetInBoundsFromOp);
+        break;
     default:
         break;
     }
diff --git a/src/shader_recompiler/frontend/ir/microinstruction.h b/src/shader_recompiler/frontend/ir/microinstruction.h
index d5336c438..770bbd550 100644
--- a/src/shader_recompiler/frontend/ir/microinstruction.h
+++ b/src/shader_recompiler/frontend/ir/microinstruction.h
@@ -134,6 +134,7 @@ static_assert(sizeof(Inst) <= 128, "Inst size unintentionally increased");
 
 struct AssociatedInsts {
     union {
+        Inst* in_bounds_inst;
         Inst* sparse_inst;
         Inst* zero_inst{};
     };
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index dd17212a1..a2479c46a 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -159,6 +159,7 @@ OPCODE(GetSignFromOp,                                       U1,             Opaq
 OPCODE(GetCarryFromOp,                                      U1,             Opaque,                                                                         )
 OPCODE(GetOverflowFromOp,                                   U1,             Opaque,                                                                         )
 OPCODE(GetSparseFromOp,                                     U1,             Opaque,                                                                         )
+OPCODE(GetInBoundsFromOp,                                   U1,             Opaque,                                                                         )
 
 // Floating-point operations
 OPCODE(FPAbs16,                                             F16,            F16,                                                                            )
@@ -363,8 +364,12 @@ OPCODE(ImageSampleExplicitLod,                              F32x4,          U32,
 OPCODE(ImageSampleDrefImplicitLod,                          F32,            U32,            Opaque,         F32,            Opaque,         Opaque,         )
 OPCODE(ImageSampleDrefExplicitLod,                          F32,            U32,            Opaque,         F32,            Opaque,         Opaque,         )
 
-// Vote operations
+// Warp operations
 OPCODE(VoteAll,                                             U1,             U1,                                                                             )
 OPCODE(VoteAny,                                             U1,             U1,                                                                             )
 OPCODE(VoteEqual,                                           U1,             U1,                                                                             )
 OPCODE(SubgroupBallot,                                      U32,            U1,                                                                             )
+OPCODE(ShuffleIndex,                                        U32,            U32,            U32,            U32,            U32,                            )
+OPCODE(ShuffleUp,                                           U32,            U32,            U32,            U32,            U32,                            )
+OPCODE(ShuffleDown,                                         U32,            U32,            U32,            U32,            U32,                            )
+OPCODE(ShuffleButterfly,                                    U32,            U32,            U32,            U32,            U32,                            )