shader: Move LaneId to the warp emission file and fix AMD

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -66,7 +66,6 @@ void EmitSetOFlag(EmitContext& ctx);
 Id EmitWorkgroupId(EmitContext& ctx);
 Id EmitLocalInvocationId(EmitContext& ctx);
 Id EmitLoadLocal(EmitContext& ctx, Id word_offset);
-Id EmitLaneId(EmitContext& ctx);
 void EmitWriteLocal(EmitContext& ctx, Id word_offset, Id value);
 Id EmitUndefU1(EmitContext& ctx);
 Id EmitUndefU8(EmitContext& ctx);
@@ -403,6 +402,7 @@ Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I
                      Id derivates, Id offset, Id lod_clamp);
 Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
 void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id color);
+Id EmitLaneId(EmitContext& ctx);
 Id EmitVoteAll(EmitContext& ctx, Id pred);
 Id EmitVoteAny(EmitContext& ctx, Id pred);
 Id EmitVoteEqual(EmitContext& ctx, Id pred);

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -274,10 +274,6 @@ Id EmitLocalInvocationId(EmitContext& ctx) {
     return ctx.OpLoad(ctx.U32[3], ctx.local_invocation_id);
 }
 
-Id EmitLaneId(EmitContext& ctx) {
-    return ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id);
-}
-
 Id EmitLoadLocal(EmitContext& ctx, Id word_offset) {
     const Id pointer{ctx.OpAccessChain(ctx.private_u32, ctx.local_memory, word_offset)};
     return ctx.OpLoad(ctx.U32[1], pointer);

src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp

@@ -49,6 +49,14 @@ Id SelectValue(EmitContext& ctx, Id in_range, Id value, Id src_thread_id) {
 }
 } // Anonymous namespace
 
+Id EmitLaneId(EmitContext& ctx) {
+    const Id id{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
+    if (!ctx.profile.warp_size_potentially_larger_than_guest) {
+        return id;
+    }
+    return ctx.OpBitwiseAnd(ctx.U32[1], id, ctx.Constant(ctx.U32[1], 31U));
+}
+
 Id EmitVoteAll(EmitContext& ctx, Id pred) {
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return ctx.OpSubgroupAllKHR(ctx.U1, pred);

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -58,7 +58,6 @@ OPCODE(SetCFlag,                                            Void,           U1,
 OPCODE(SetOFlag,                                            Void,           U1,                                                                             )
 OPCODE(WorkgroupId,                                         U32x3,                                                                                          )
 OPCODE(LocalInvocationId,                                   U32x3,                                                                                          )
-OPCODE(LaneId,                                              U32,                                                                                            )
 
 // Undefined
 OPCODE(UndefU1,                                             U1,                                                                                             )
@@ -419,6 +418,7 @@ OPCODE(ImageRead,                                           U32x4,          U32,
 OPCODE(ImageWrite,                                          Void,           U32,            Opaque,         U32x4,                                          )
 
 // Warp operations
+OPCODE(LaneId,                                              U32,                                                                                            )
 OPCODE(VoteAll,                                             U1,             U1,                                                                             )
 OPCODE(VoteAny,                                             U1,             U1,                                                                             )
 OPCODE(VoteEqual,                                           U1,             U1,                                                                             )

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -343,11 +343,11 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::LocalInvocationId:
         info.uses_local_invocation_id = true;
         break;
+    case IR::Opcode::LaneId:
     case IR::Opcode::ShuffleIndex:
     case IR::Opcode::ShuffleUp:
     case IR::Opcode::ShuffleDown:
     case IR::Opcode::ShuffleButterfly:
-    case IR::Opcode::LaneId:
         info.uses_subgroup_invocation_id = true;
         break;
     case IR::Opcode::GetCbufU8: