diff options
Diffstat (limited to 'src/codegen')
| -rw-r--r-- | src/codegen/spirv.zig | 406 | ||||
| -rw-r--r-- | src/codegen/spirv/Module.zig | 90 |
2 files changed, 217 insertions, 279 deletions
diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig index e1baf20156..8c4b0478de 100644 --- a/src/codegen/spirv.zig +++ b/src/codegen/spirv.zig @@ -159,7 +159,7 @@ pub const Object = struct { uav_link: std.AutoHashMapUnmanaged(struct { InternPool.Index, StorageClass }, SpvModule.Decl.Index) = .empty, /// A map that maps AIR intern pool indices to SPIR-V result-ids. - intern_map: InternMap = .{}, + intern_map: InternMap = .empty, /// This map serves a dual purpose: /// - It keeps track of pointers that are currently being emitted, so that we can tell @@ -314,7 +314,7 @@ const NavGen = struct { next_arg_index: u32 = 0, /// A map keeping track of which instruction generated which result-id. - inst_results: InstMap = .{}, + inst_results: InstMap = .empty, /// A map that maps AIR intern pool indices to SPIR-V result-ids. /// See `Object.intern_map`. @@ -469,7 +469,7 @@ const NavGen = struct { const zcu = self.pt.zcu; const ty = Type.fromInterned(zcu.intern_pool.typeOf(val)); - const decl_ptr_ty_id = try self.ptrType(ty, .Generic); + const decl_ptr_ty_id = try self.ptrType(ty, .Generic, .indirect); const spv_decl_index = blk: { const entry = try self.object.uav_link.getOrPut(self.object.gpa, .{ val, .Function }); @@ -532,7 +532,7 @@ const NavGen = struct { try self.spv.debugNameFmt(initializer_id, "initializer of __anon_{d}", .{@intFromEnum(val)}); - const fn_decl_ptr_ty_id = try self.ptrType(ty, .Function); + const fn_decl_ptr_ty_id = try self.ptrType(ty, .Function, .indirect); try self.spv.sections.types_globals_constants.emit(self.spv.gpa, .OpExtInst, .{ .id_result_type = fn_decl_ptr_ty_id, .id_result = result_id, @@ -721,36 +721,16 @@ const NavGen = struct { /// Emits a bool constant in a particular representation. fn constBool(self: *NavGen, value: bool, repr: Repr) !IdRef { - // TODO: Cache? - - const section = &self.spv.sections.types_globals_constants; - switch (repr) { - .indirect => { - return try self.constInt(Type.u1, @intFromBool(value), .indirect); - }, - .direct => { - const result_ty_id = try self.resolveType(Type.bool, .direct); - const result_id = self.spv.allocId(); - switch (value) { - inline else => |val_ct| try section.emit( - self.spv.gpa, - if (val_ct) .OpConstantTrue else .OpConstantFalse, - .{ - .id_result_type = result_ty_id, - .id_result = result_id, - }, - ), - } - return result_id; - }, - } + return switch (repr) { + .indirect => self.constInt(Type.u1, @intFromBool(value)), + .direct => self.spv.constBool(value), + }; } /// Emits an integer constant. /// This function, unlike SpvModule.constInt, takes care to bitcast /// the value to an unsigned int first for Kernels. - fn constInt(self: *NavGen, ty: Type, value: anytype, repr: Repr) !IdRef { - // TODO: Cache? + fn constInt(self: *NavGen, ty: Type, value: anytype) !IdRef { const zcu = self.pt.zcu; const scalar_ty = ty.scalarType(zcu); const int_info = scalar_ty.intInfo(zcu); @@ -763,18 +743,18 @@ const NavGen = struct { else => unreachable, }; - const bits: u64 = switch (signedness) { + const value64: u64 = switch (signedness) { .signed => @bitCast(@as(i64, @intCast(value))), .unsigned => @as(u64, @intCast(value)), }; // Manually truncate the value to the right amount of bits. - const truncated_bits = if (backing_bits == 64) - bits + const truncated_value = if (backing_bits == 64) + value64 else - bits & (@as(u64, 1) << @intCast(backing_bits)) - 1; + value64 & (@as(u64, 1) << @intCast(backing_bits)) - 1; - const result_ty_id = try self.resolveType(scalar_ty, repr); + const result_ty_id = try self.resolveType(scalar_ty, .indirect); const result_id = self.spv.allocId(); const section = &self.spv.sections.types_globals_constants; @@ -783,100 +763,42 @@ const NavGen = struct { 1...32 => try section.emit(self.spv.gpa, .OpConstant, .{ .id_result_type = result_ty_id, .id_result = result_id, - .value = .{ .uint32 = @truncate(truncated_bits) }, + .value = .{ .uint32 = @truncate(truncated_value) }, }), 33...64 => try section.emit(self.spv.gpa, .OpConstant, .{ .id_result_type = result_ty_id, .id_result = result_id, - .value = .{ .uint64 = truncated_bits }, + .value = .{ .uint64 = truncated_value }, }), else => unreachable, // TODO: Large integer constants } - if (!ty.isVector(zcu)) { - return result_id; - } - - const n = ty.vectorLen(zcu); - const ids = try self.gpa.alloc(IdRef, n); - defer self.gpa.free(ids); - @memset(ids, result_id); - - const vec_ty_id = try self.resolveType(ty, repr); - const vec_result_id = self.spv.allocId(); - try self.func.body.emit(self.spv.gpa, .OpCompositeConstruct, .{ - .id_result_type = vec_ty_id, - .id_result = vec_result_id, - .constituents = ids, - }); - return vec_result_id; + if (!ty.isVector(zcu)) return result_id; + return self.constructCompositeSplat(ty, result_id); } - /// Construct a struct at runtime. - /// ty must be a struct type. - /// Constituents should be in `indirect` representation (as the elements of a struct should be). - /// Result is in `direct` representation. - fn constructStruct(self: *NavGen, ty: Type, types: []const Type, constituents: []const IdRef) !IdRef { - assert(types.len == constituents.len); - + pub fn constructComposite(self: *NavGen, result_ty_id: IdRef, constituents: []const IdRef) !IdRef { const result_id = self.spv.allocId(); - try self.func.body.emit(self.spv.gpa, .OpCompositeConstruct, .{ - .id_result_type = try self.resolveType(ty, .direct), - .id_result = result_id, - .constituents = constituents, - }); - return result_id; - } - - /// Construct a vector at runtime. - /// ty must be an vector type. - fn constructVector(self: *NavGen, ty: Type, constituents: []const IdRef) !IdRef { - const zcu = self.pt.zcu; - assert(ty.vectorLen(zcu) == constituents.len); - - // Note: older versions of the Khronos SPRIV-LLVM translator crash on this instruction - // because it cannot construct structs which' operands are not constant. - // See https://github.com/KhronosGroup/SPIRV-LLVM-Translator/issues/1349 - // Currently this is the case for Intel OpenCL CPU runtime (2023-WW46), but the - // alternatives dont work properly: - // - using temporaries/pointers doesn't work properly with vectors of bool, causes - // backends that use llvm to crash - // - using OpVectorInsertDynamic doesn't work for non-spirv-vectors of bool. - - const result_id = self.spv.allocId(); - try self.func.body.emit(self.spv.gpa, .OpCompositeConstruct, .{ - .id_result_type = try self.resolveType(ty, .direct), + try self.func.body.emit(self.gpa, .OpCompositeConstruct, .{ + .id_result_type = result_ty_id, .id_result = result_id, .constituents = constituents, }); return result_id; } - /// Construct a vector at runtime with all lanes set to the same value. - /// ty must be an vector type. - fn constructVectorSplat(self: *NavGen, ty: Type, constituent: IdRef) !IdRef { + /// Construct a composite at runtime with all lanes set to the same value. + /// ty must be an aggregate type. + fn constructCompositeSplat(self: *NavGen, ty: Type, constituent: IdRef) !IdRef { const zcu = self.pt.zcu; - const n = ty.vectorLen(zcu); + const n = ty.arrayLen(zcu); const constituents = try self.gpa.alloc(IdRef, n); defer self.gpa.free(constituents); @memset(constituents, constituent); - return try self.constructVector(ty, constituents); - } - - /// Construct an array at runtime. - /// ty must be an array type. - /// Constituents should be in `indirect` representation (as the elements of an array should be). - /// Result is in `direct` representation. - fn constructArray(self: *NavGen, ty: Type, constituents: []const IdRef) !IdRef { - const result_id = self.spv.allocId(); - try self.func.body.emit(self.spv.gpa, .OpCompositeConstruct, .{ - .id_result_type = try self.resolveType(ty, .direct), - .id_result = result_id, - .constituents = constituents, - }); - return result_id; + const result_ty_id = try self.resolveType(ty, .direct); + return self.constructComposite(result_ty_id, constituents); } /// This function generates a load for a constant in direct (ie, non-memory) representation. @@ -947,9 +869,9 @@ const NavGen = struct { }, .int => { if (ty.isSignedInt(zcu)) { - break :cache try self.constInt(ty, val.toSignedInt(zcu), repr); + break :cache try self.constInt(ty, val.toSignedInt(zcu)); } else { - break :cache try self.constInt(ty, val.toUnsignedInt(zcu), repr); + break :cache try self.constInt(ty, val.toUnsignedInt(zcu)); } }, .float => { @@ -970,7 +892,7 @@ const NavGen = struct { }, .err => |err| { const value = try pt.getErrorValue(err.name); - break :cache try self.constInt(ty, value, repr); + break :cache try self.constInt(ty, value); }, .error_union => |error_union| { // TODO: Error unions may be constructed with constant instructions if the payload type @@ -1011,7 +933,8 @@ const NavGen = struct { types = .{ payload_ty, err_ty }; } - return try self.constructStruct(ty, &types, &constituents); + const comp_ty_id = try self.resolveType(ty, .direct); + return try self.constructComposite(comp_ty_id, &constituents); }, .enum_tag => { const int_val = try val.intFromEnum(ty, pt); @@ -1020,14 +943,10 @@ const NavGen = struct { }, .ptr => return self.constantPtr(val), .slice => |slice| { - const ptr_ty = ty.slicePtrFieldType(zcu); const ptr_id = try self.constantPtr(Value.fromInterned(slice.ptr)); const len_id = try self.constant(Type.usize, Value.fromInterned(slice.len), .indirect); - return self.constructStruct( - ty, - &.{ ptr_ty, Type.usize }, - &.{ ptr_id, len_id }, - ); + const comp_ty_id = try self.resolveType(ty, .direct); + return try self.constructComposite(comp_ty_id, &.{ ptr_id, len_id }); }, .opt => { const payload_ty = ty.optionalChild(zcu); @@ -1053,11 +972,8 @@ const NavGen = struct { else try self.spv.constUndef(try self.resolveType(payload_ty, .indirect)); - return try self.constructStruct( - ty, - &.{ payload_ty, Type.bool }, - &.{ payload_id, has_pl_id }, - ); + const comp_ty_id = try self.resolveType(ty, .direct); + return try self.constructComposite(comp_ty_id, &.{ payload_id, has_pl_id }); }, .aggregate => |aggregate| switch (ip.indexToKey(ty.ip_index)) { inline .array_type, .vector_type => |array_type, tag| { @@ -1077,7 +993,7 @@ const NavGen = struct { // TODO: This is really space inefficient, perhaps there is a better // way to do it? for (constituents, bytes.toSlice(constituents.len, ip)) |*constituent, byte| { - constituent.* = try self.constInt(elem_ty, byte, child_repr); + constituent.* = try self.constInt(elem_ty, byte); } }, .elems => |elems| { @@ -1090,11 +1006,8 @@ const NavGen = struct { }, } - switch (tag) { - .array_type => return self.constructArray(ty, constituents), - .vector_type => return self.constructVector(ty, constituents), - else => unreachable, - } + const comp_ty_id = try self.resolveType(ty, .direct); + return self.constructComposite(comp_ty_id, constituents); }, .struct_type => { const struct_type = zcu.typeToStruct(ty).?; @@ -1124,7 +1037,8 @@ const NavGen = struct { try constituents.append(field_id); } - return try self.constructStruct(ty, types.items, constituents.items); + const comp_ty_id = try self.resolveType(ty, .direct); + return try self.constructComposite(comp_ty_id, constituents.items); }, .tuple_type => unreachable, // TODO else => unreachable, @@ -1149,8 +1063,6 @@ const NavGen = struct { } fn constantPtr(self: *NavGen, ptr_val: Value) Error!IdRef { - // TODO: Caching?? - const pt = self.pt; if (ptr_val.isUndef(pt.zcu)) { @@ -1201,7 +1113,7 @@ const NavGen = struct { .elem_ptr => |elem| { const parent_ptr_id = try self.derivePtr(elem.parent.*); const parent_ptr_ty = try elem.parent.ptrType(pt); - const index_id = try self.constInt(Type.usize, elem.elem_idx, .direct); + const index_id = try self.constInt(Type.usize, elem.elem_idx); return self.ptrElemPtr(parent_ptr_ty, parent_ptr_id, index_id); }, .offset_and_cast => |oac| { @@ -1255,7 +1167,7 @@ const NavGen = struct { // Uav refs are always generic. assert(ty.ptrAddressSpace(zcu) == .generic); - const decl_ptr_ty_id = try self.ptrType(uav_ty, .Generic); + const decl_ptr_ty_id = try self.ptrType(uav_ty, .Generic, .indirect); const ptr_id = try self.resolveUav(uav.val); if (decl_ptr_ty_id != ty_id) { @@ -1310,7 +1222,7 @@ const NavGen = struct { const storage_class = self.spvStorageClass(nav.getAddrspace()); try self.addFunctionDep(spv_decl_index, storage_class); - const decl_ptr_ty_id = try self.ptrType(nav_ty, storage_class); + const decl_ptr_ty_id = try self.ptrType(nav_ty, storage_class, .indirect); const ptr_id = switch (storage_class) { .Generic => try self.castToGeneric(decl_ptr_ty_id, decl_id), @@ -1359,23 +1271,11 @@ const NavGen = struct { } fn arrayType(self: *NavGen, len: u32, child_ty: IdRef) !IdRef { - // TODO: Cache?? - const len_id = try self.constInt(Type.u32, len, .direct); - const result_id = self.spv.allocId(); - - try self.spv.sections.types_globals_constants.emit(self.spv.gpa, .OpTypeArray, .{ - .id_result = result_id, - .element_type = child_ty, - .length = len_id, - }); - return result_id; - } - - fn ptrType(self: *NavGen, child_ty: Type, storage_class: StorageClass) !IdRef { - return try self.ptrType2(child_ty, storage_class, .indirect); + const len_id = try self.constInt(Type.u32, len); + return self.spv.arrayType(len_id, child_ty); } - fn ptrType2(self: *NavGen, child_ty: Type, storage_class: StorageClass, child_repr: Repr) !IdRef { + fn ptrType(self: *NavGen, child_ty: Type, storage_class: StorageClass, child_repr: Repr) !IdRef { const key = .{ child_ty.toIntern(), storage_class, child_repr }; const entry = try self.ptr_types.getOrPut(self.gpa, key); if (entry.found_existing) { @@ -1408,8 +1308,7 @@ const NavGen = struct { } fn functionType(self: *NavGen, return_ty: Type, param_types: []const Type) !IdRef { - // TODO: Cache?? - + const return_ty_id = try self.resolveFnReturnType(return_ty); const param_ids = try self.gpa.alloc(IdRef, param_types.len); defer self.gpa.free(param_ids); @@ -1417,14 +1316,7 @@ const NavGen = struct { param_id.* = try self.resolveType(param_ty, .direct); } - const ty_id = self.spv.allocId(); - try self.spv.sections.types_globals_constants.emit(self.spv.gpa, .OpTypeFunction, .{ - .id_result = ty_id, - .return_type = try self.resolveFnReturnType(return_ty), - .id_ref_2 = param_ids, - }); - - return ty_id; + return self.spv.functionType(return_ty_id, param_ids); } fn zigScalarOrVectorTypeLike(self: *NavGen, new_ty: Type, base_ty: Type) !Type { @@ -1702,13 +1594,7 @@ const NavGen = struct { const child_ty = Type.fromInterned(ptr_info.child); const storage_class = self.spvStorageClass(ptr_info.flags.address_space); - const ptr_ty_id = try self.ptrType(child_ty, storage_class); - - if (target.os.tag == .vulkan and ptr_info.flags.size == .many) { - try self.spv.decorate(ptr_ty_id, .{ .ArrayStride = .{ - .array_stride = @intCast(child_ty.abiSize(zcu)), - } }); - } + const ptr_ty_id = try self.ptrType(child_ty, storage_class, .indirect); if (ptr_info.flags.size != .slice) { return ptr_ty_id; @@ -1755,10 +1641,6 @@ const NavGen = struct { defer self.gpa.free(type_name); try self.spv.debugName(result_id, type_name); - if (target.os.tag == .vulkan) { - try self.spv.decorate(result_id, .Block); // Decorate all structs as block for now... - } - return result_id; }, .struct_type => ip.loadStructType(ty.toIntern()), @@ -1804,10 +1686,6 @@ const NavGen = struct { defer self.gpa.free(type_name); try self.spv.debugName(result_id, type_name); - if (target.os.tag == .vulkan) { - try self.spv.decorate(result_id, .Block); // Decorate all structs as block for now... - } - return result_id; }, .optional => { @@ -2073,12 +1951,13 @@ const NavGen = struct { .exploded_vector => |range| { assert(self.ty.isVector(zcu)); assert(self.ty.vectorLen(zcu) == range.len); - const consituents = try ng.gpa.alloc(IdRef, range.len); - defer ng.gpa.free(consituents); - for (consituents, 0..range.len) |*id, i| { + const constituents = try ng.gpa.alloc(IdRef, range.len); + defer ng.gpa.free(constituents); + for (constituents, 0..range.len) |*id, i| { id.* = range.at(i); } - return ng.constructVector(self.ty, consituents); + const result_ty_id = try ng.resolveType(self.ty, .direct); + return ng.constructComposite(result_ty_id, constituents); }, } } @@ -2282,7 +2161,7 @@ const NavGen = struct { .child = tmp.ty.toIntern(), }); - const vector = try ng.constructVectorSplat(vector_ty, id); + const vector = try ng.constructCompositeSplat(vector_ty, id); return .{ .ty = vector_ty, .value = .{ .spv_vectorwise = vector }, @@ -3045,7 +2924,7 @@ const NavGen = struct { const spv_err_decl_index = self.object.error_push_constant.?.push_constant_ptr; const push_constant_id = self.spv.declPtr(spv_err_decl_index).result_id; - const zero_id = try self.constInt(Type.u32, 0, .direct); + const zero_id = try self.constInt(Type.u32, 0); // We cannot use OpInBoundsAccessChain to dereference cross-storage class, so we have to use // a load. const tmp = self.spv.allocId(); @@ -3187,7 +3066,7 @@ const NavGen = struct { const storage_class = self.spvStorageClass(nav.getAddrspace()); assert(storage_class != .Generic); // These should be instance globals - const ptr_ty_id = try self.ptrType(ty, storage_class); + const ptr_ty_id = try self.ptrType(ty, storage_class, .indirect); try self.spv.sections.types_globals_constants.emit(self.spv.gpa, .OpVariable, .{ .id_result_type = ptr_ty_id, @@ -3208,7 +3087,7 @@ const NavGen = struct { try self.spv.declareDeclDeps(spv_decl_index, &.{}); - const ptr_ty_id = try self.ptrType(ty, .Function); + const ptr_ty_id = try self.ptrType(ty, .Function, .indirect); if (maybe_init_val) |init_val| { // TODO: Combine with resolveAnonDecl? @@ -3265,8 +3144,8 @@ const NavGen = struct { } fn intFromBool2(self: *NavGen, value: Temporary, result_ty: Type) !Temporary { - const zero_id = try self.constInt(result_ty, 0, .direct); - const one_id = try self.constInt(result_ty, 1, .direct); + const zero_id = try self.constInt(result_ty, 0); + const one_id = try self.constInt(result_ty, 1); return try self.buildSelect( value, @@ -3648,12 +3527,12 @@ const NavGen = struct { .strange_integer => switch (info.signedness) { .unsigned => { const mask_value = if (info.bits == 64) 0xFFFF_FFFF_FFFF_FFFF else (@as(u64, 1) << @as(u6, @intCast(info.bits))) - 1; - const mask_id = try self.constInt(ty.scalarType(zcu), mask_value, .direct); + const mask_id = try self.constInt(ty.scalarType(zcu), mask_value); return try self.buildBinary(.bit_and, value, Temporary.init(ty.scalarType(zcu), mask_id)); }, .signed => { // Shift left and right so that we can copy the sight bit that way. - const shift_amt_id = try self.constInt(ty.scalarType(zcu), info.backing_bits - info.bits, .direct); + const shift_amt_id = try self.constInt(ty.scalarType(zcu), info.backing_bits - info.bits); const shift_amt = Temporary.init(ty.scalarType(zcu), shift_amt_id); const left = try self.buildBinary(.sll, value, shift_amt); return try self.buildBinary(.sra, left, shift_amt); @@ -3687,7 +3566,7 @@ const NavGen = struct { const div = try self.buildBinary(.s_div, lhs, rhs); const rem = try self.buildBinary(.s_rem, lhs, rhs); - const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0, .direct)); + const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0)); const rem_is_not_zero = try self.buildCmp(.i_ne, rem, zero); @@ -3863,7 +3742,7 @@ const NavGen = struct { // = (rhs < 0) == (value < lhs) // = (rhs < 0) == (lhs > value) .signed => blk: { - const zero = Temporary.init(rhs.ty, try self.constInt(rhs.ty, 0, .direct)); + const zero = Temporary.init(rhs.ty, try self.constInt(rhs.ty, 0)); const rhs_lt_zero = try self.buildCmp(.s_lt, rhs, zero); const result_gt_lhs = try self.buildCmp(scmp, lhs, result); break :blk try self.buildCmp(.l_eq, rhs_lt_zero, result_gt_lhs); @@ -3872,11 +3751,8 @@ const NavGen = struct { const ov = try self.intFromBool(overflowed); - return try self.constructStruct( - result_ty, - &.{ result.ty, ov.ty }, - &.{ try result.materialize(self), try ov.materialize(self) }, - ); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, &.{ try result.materialize(self), try ov.materialize(self) }); } fn airMulOverflow(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -3928,11 +3804,11 @@ const NavGen = struct { const result = try self.normalize(low_bits, info); // Shift the result bits away to get the overflow bits. - const shift = Temporary.init(full_result.ty, try self.constInt(full_result.ty, info.bits, .direct)); + const shift = Temporary.init(full_result.ty, try self.constInt(full_result.ty, info.bits)); const overflow = try self.buildBinary(.srl, full_result, shift); // Directly check if its zero in the op_ty without converting first. - const zero = Temporary.init(full_result.ty, try self.constInt(full_result.ty, 0, .direct)); + const zero = Temporary.init(full_result.ty, try self.constInt(full_result.ty, 0)); const overflowed = try self.buildCmp(.i_ne, zero, overflow); break :blk .{ result, overflowed }; @@ -3946,7 +3822,7 @@ const NavGen = struct { // Overflow happened if the high-bits of the result are non-zero OR if the // high bits of the low word of the result (those outside the range of the // int) are nonzero. - const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0, .direct)); + const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0)); const high_overflowed = try self.buildCmp(.i_ne, zero, high_bits); // If no overflow bits in low_bits, no extra work needs to be done. @@ -3955,7 +3831,7 @@ const NavGen = struct { } // Shift the result bits away to get the overflow bits. - const shift = Temporary.init(lhs.ty, try self.constInt(lhs.ty, info.bits, .direct)); + const shift = Temporary.init(lhs.ty, try self.constInt(lhs.ty, info.bits)); const low_overflow = try self.buildBinary(.srl, low_bits, shift); const low_overflowed = try self.buildCmp(.i_ne, zero, low_overflow); @@ -3974,7 +3850,7 @@ const NavGen = struct { // overflow should be -1 when // (lhs > 0 && rhs < 0) || (lhs < 0 && rhs > 0) - const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0, .direct)); + const zero = Temporary.init(lhs.ty, try self.constInt(lhs.ty, 0)); const lhs_negative = try self.buildCmp(.s_lt, lhs, zero); const rhs_negative = try self.buildCmp(.s_lt, rhs, zero); const lhs_positive = try self.buildCmp(.s_gt, lhs, zero); @@ -4003,13 +3879,13 @@ const NavGen = struct { // bit for the expected overflow bits. // To do that, shift out everything bit the sign bit and // then check what remains. - const shift = Temporary.init(full_result.ty, try self.constInt(full_result.ty, info.bits - 1, .direct)); + const shift = Temporary.init(full_result.ty, try self.constInt(full_result.ty, info.bits - 1)); // Use SRA so that any sign bits are duplicated. Now we can just check if ALL bits are set // for negative cases. const overflow = try self.buildBinary(.sra, full_result, shift); - const long_all_set = Temporary.init(full_result.ty, try self.constInt(full_result.ty, -1, .direct)); - const long_zero = Temporary.init(full_result.ty, try self.constInt(full_result.ty, 0, .direct)); + const long_all_set = Temporary.init(full_result.ty, try self.constInt(full_result.ty, -1)); + const long_zero = Temporary.init(full_result.ty, try self.constInt(full_result.ty, 0)); const mask = try self.buildSelect(expected_overflow_bit, long_all_set, long_zero); const overflowed = try self.buildCmp(.i_ne, mask, overflow); @@ -4022,7 +3898,7 @@ const NavGen = struct { // Truncate result if required. const result = try self.normalize(low_bits, info); - const all_set = Temporary.init(lhs.ty, try self.constInt(lhs.ty, -1, .direct)); + const all_set = Temporary.init(lhs.ty, try self.constInt(lhs.ty, -1)); const mask = try self.buildSelect(expected_overflow_bit, all_set, zero); // Like with unsigned, overflow happened if high_bits are not the ones we expect, @@ -4038,7 +3914,7 @@ const NavGen = struct { } // Shift the result bits away to get the overflow bits. - const shift = Temporary.init(lhs.ty, try self.constInt(lhs.ty, info.bits - 1, .direct)); + const shift = Temporary.init(lhs.ty, try self.constInt(lhs.ty, info.bits - 1)); // Use SRA so that any sign bits are duplicated. Now we can just check if ALL bits are set // for negative cases. const low_overflow = try self.buildBinary(.sra, low_bits, shift); @@ -4052,11 +3928,8 @@ const NavGen = struct { const ov = try self.intFromBool(overflowed); - return try self.constructStruct( - result_ty, - &.{ result.ty, ov.ty }, - &.{ try result.materialize(self), try ov.materialize(self) }, - ); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, &.{ try result.materialize(self), try ov.materialize(self) }); } fn airShlOverflow(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -4092,11 +3965,8 @@ const NavGen = struct { const overflowed = try self.buildCmp(.i_ne, base, right); const ov = try self.intFromBool(overflowed); - return try self.constructStruct( - result_ty, - &.{ result.ty, ov.ty }, - &.{ try result.materialize(self), try ov.materialize(self) }, - ); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, &.{ try result.materialize(self), try ov.materialize(self) }); } fn airMulAdd(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -4163,7 +4033,7 @@ const NavGen = struct { const operand_id = try self.resolve(ty_op.operand); const result_ty = self.typeOfIndex(inst); - return try self.constructVectorSplat(result_ty, operand_id); + return try self.constructCompositeSplat(result_ty, operand_id); } fn airReduce(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -4297,10 +4167,10 @@ const NavGen = struct { // Fall back to manually extracting and inserting components. - const components = try self.gpa.alloc(IdRef, result_ty.vectorLen(zcu)); - defer self.gpa.free(components); + const constituents = try self.gpa.alloc(IdRef, result_ty.vectorLen(zcu)); + defer self.gpa.free(constituents); - for (components, 0..) |*id, i| { + for (constituents, 0..) |*id, i| { const elem = try mask.elemValue(pt, i); if (elem.isUndef(zcu)) { id.* = try self.spv.constUndef(scalar_ty_id); @@ -4315,14 +4185,15 @@ const NavGen = struct { } } - return try self.constructVector(result_ty, components); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, constituents); } fn indicesToIds(self: *NavGen, indices: []const u32) ![]IdRef { const ids = try self.gpa.alloc(IdRef, indices.len); errdefer self.gpa.free(ids); for (indices, ids) |index, *id| { - id.* = try self.constInt(Type.u32, index, .direct); + id.* = try self.constInt(Type.u32, index); } return ids; @@ -4676,7 +4547,7 @@ const NavGen = struct { break :blk result_id; } - const dst_ptr_ty_id = try self.ptrType(dst_ty, .Function); + const dst_ptr_ty_id = try self.ptrType(dst_ty, .Function, .indirect); const tmp_id = try self.alloc(src_ty, .{ .storage_class = .Function }); try self.store(src_ty, tmp_id, src_id, .{}); @@ -4851,7 +4722,7 @@ const NavGen = struct { const elem_ptr_ty_id = try self.resolveType(elem_ptr_ty, .direct); const array_ptr_id = try self.resolve(ty_op.operand); - const len_id = try self.constInt(Type.usize, array_ty.arrayLen(zcu), .direct); + const len_id = try self.constInt(Type.usize, array_ty.arrayLen(zcu)); const elem_ptr_id = if (!array_ty.hasRuntimeBitsIgnoreComptime(zcu)) // Note: The pointer is something like *opaque{}, so we need to bitcast it to the element type. @@ -4860,11 +4731,8 @@ const NavGen = struct { // Convert the pointer-to-array to a pointer to the first element. try self.accessChain(elem_ptr_ty_id, array_ptr_id, &.{0}); - return try self.constructStruct( - slice_ty, - &.{ elem_ptr_ty, Type.usize }, - &.{ elem_ptr_id, len_id }, - ); + const slice_ty_id = try self.resolveType(slice_ty, .direct); + return try self.constructComposite(slice_ty_id, &.{ elem_ptr_id, len_id }); } fn airSlice(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -4872,16 +4740,9 @@ const NavGen = struct { const bin_op = self.air.extraData(Air.Bin, ty_pl.payload).data; const ptr_id = try self.resolve(bin_op.lhs); const len_id = try self.resolve(bin_op.rhs); - const ptr_ty = self.typeOf(bin_op.lhs); const slice_ty = self.typeOfIndex(inst); - - // Note: Types should not need to be converted to direct, these types - // dont need to be converted. - return try self.constructStruct( - slice_ty, - &.{ ptr_ty, Type.usize }, - &.{ ptr_id, len_id }, - ); + const slice_ty_id = try self.resolveType(slice_ty, .direct); + return try self.constructComposite(slice_ty_id, &.{ ptr_id, len_id }); } fn airAggregateInit(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -4936,11 +4797,8 @@ const NavGen = struct { else => unreachable, } - return try self.constructStruct( - result_ty, - types[0..index], - constituents[0..index], - ); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, constituents[0..index]); }, .vector => { const n_elems = result_ty.vectorLen(zcu); @@ -4951,7 +4809,8 @@ const NavGen = struct { elem_ids[i] = try self.resolve(element); } - return try self.constructVector(result_ty, elem_ids); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, elem_ids); }, .array => { const array_info = result_ty.arrayInfo(zcu); @@ -4968,7 +4827,8 @@ const NavGen = struct { elem_ids[n_elems - 1] = try self.constant(array_info.elem_type, sentinel_val, .indirect); } - return try self.constructArray(result_ty, elem_ids); + const result_ty_id = try self.resolveType(result_ty, .direct); + return try self.constructComposite(result_ty_id, elem_ids); }, else => unreachable, } @@ -4984,7 +4844,7 @@ const NavGen = struct { const elem_ty = array_ty.childType(zcu); const abi_size = elem_ty.abiSize(zcu); const size = array_ty.arrayLenIncludingSentinel(zcu) * abi_size; - return try self.constInt(Type.usize, size, .direct); + return try self.constInt(Type.usize, size); }, .many, .c => unreachable, } @@ -5060,7 +4920,7 @@ const NavGen = struct { const zcu = self.pt.zcu; // Construct new pointer type for the resulting pointer const elem_ty = ptr_ty.elemType2(zcu); // use elemType() so that we get T for *[N]T. - const elem_ptr_ty_id = try self.ptrType(elem_ty, self.spvStorageClass(ptr_ty.ptrAddressSpace(zcu))); + const elem_ptr_ty_id = try self.ptrType(elem_ty, self.spvStorageClass(ptr_ty.ptrAddressSpace(zcu)), .indirect); if (ptr_ty.isSinglePointer(zcu)) { // Pointer-to-array. In this case, the resulting pointer is not of the same type // as the ptr_ty (we want a *T, not a *[N]T), and hence we need to use accessChain. @@ -5115,8 +4975,8 @@ const NavGen = struct { const is_vector = array_ty.isVector(zcu); const elem_repr: Repr = if (is_vector) .direct else .indirect; - const ptr_array_ty_id = try self.ptrType2(array_ty, .Function, .direct); - const ptr_elem_ty_id = try self.ptrType2(elem_ty, .Function, elem_repr); + const ptr_array_ty_id = try self.ptrType(array_ty, .Function, .direct); + const ptr_elem_ty_id = try self.ptrType(elem_ty, .Function, elem_repr); const tmp_id = self.spv.allocId(); try self.func.prologue.emit(self.spv.gpa, .OpVariable, .{ @@ -5171,7 +5031,7 @@ const NavGen = struct { const scalar_ty = vector_ty.scalarType(zcu); const storage_class = self.spvStorageClass(vector_ptr_ty.ptrAddressSpace(zcu)); - const scalar_ptr_ty_id = try self.ptrType(scalar_ty, storage_class); + const scalar_ptr_ty_id = try self.ptrType(scalar_ty, storage_class, .indirect); const vector_ptr = try self.resolve(data.vector_ptr); const index = try self.resolve(extra.lhs); @@ -5193,7 +5053,7 @@ const NavGen = struct { if (layout.tag_size == 0) return; const tag_ty = un_ty.unionTagTypeSafety(zcu).?; - const tag_ptr_ty_id = try self.ptrType(tag_ty, self.spvStorageClass(un_ptr_ty.ptrAddressSpace(zcu))); + const tag_ptr_ty_id = try self.ptrType(tag_ty, self.spvStorageClass(un_ptr_ty.ptrAddressSpace(zcu)), .indirect); const union_ptr_id = try self.resolve(bin_op.lhs); const new_tag_id = try self.resolve(bin_op.rhs); @@ -5252,23 +5112,23 @@ const NavGen = struct { } else 0; if (!layout.has_payload) { - return try self.constInt(tag_ty, tag_int, .direct); + return try self.constInt(tag_ty, tag_int); } const tmp_id = try self.alloc(ty, .{ .storage_class = .Function }); if (layout.tag_size != 0) { - const tag_ptr_ty_id = try self.ptrType(tag_ty, .Function); + const tag_ptr_ty_id = try self.ptrType(tag_ty, .Function, .indirect); const ptr_id = try self.accessChain(tag_ptr_ty_id, tmp_id, &.{@as(u32, @intCast(layout.tag_index))}); - const tag_id = try self.constInt(tag_ty, tag_int, .direct); + const tag_id = try self.constInt(tag_ty, tag_int); try self.store(tag_ty, ptr_id, tag_id, .{}); } const payload_ty = Type.fromInterned(union_ty.field_types.get(ip)[active_field]); if (payload_ty.hasRuntimeBitsIgnoreComptime(zcu)) { - const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, .Function); + const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, .Function, .indirect); const pl_ptr_id = try self.accessChain(pl_ptr_ty_id, tmp_id, &.{layout.payload_index}); - const active_pl_ptr_ty_id = try self.ptrType(payload_ty, .Function); + const active_pl_ptr_ty_id = try self.ptrType(payload_ty, .Function, .indirect); const active_pl_ptr_id = self.spv.allocId(); try self.func.body.emit(self.spv.gpa, .OpBitcast, .{ .id_result_type = active_pl_ptr_ty_id, @@ -5332,10 +5192,10 @@ const NavGen = struct { const tmp_id = try self.alloc(object_ty, .{ .storage_class = .Function }); try self.store(object_ty, tmp_id, object_id, .{}); - const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, .Function); + const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, .Function, .indirect); const pl_ptr_id = try self.accessChain(pl_ptr_ty_id, tmp_id, &.{layout.payload_index}); - const active_pl_ptr_ty_id = try self.ptrType(field_ty, .Function); + const active_pl_ptr_ty_id = try self.ptrType(field_ty, .Function, .indirect); const active_pl_ptr_id = self.spv.allocId(); try self.func.body.emit(self.spv.gpa, .OpBitcast, .{ .id_result_type = active_pl_ptr_ty_id, @@ -5365,7 +5225,7 @@ const NavGen = struct { const base_ptr_int = base_ptr_int: { if (field_offset == 0) break :base_ptr_int field_ptr_int; - const field_offset_id = try self.constInt(Type.usize, field_offset, .direct); + const field_offset_id = try self.constInt(Type.usize, field_offset); const field_ptr_tmp = Temporary.init(Type.usize, field_ptr_int); const field_offset_tmp = Temporary.init(Type.usize, field_offset_id); const result = try self.buildBinary(.i_sub, field_ptr_tmp, field_offset_tmp); @@ -5415,7 +5275,7 @@ const NavGen = struct { } const storage_class = self.spvStorageClass(object_ptr_ty.ptrAddressSpace(zcu)); - const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, storage_class); + const pl_ptr_ty_id = try self.ptrType(layout.payload_ty, storage_class, .indirect); const pl_ptr_id = try self.accessChain(pl_ptr_ty_id, object_ptr, &.{layout.payload_index}); const active_pl_ptr_id = self.spv.allocId(); @@ -5456,7 +5316,7 @@ const NavGen = struct { ty: Type, options: AllocOptions, ) !IdRef { - const ptr_fn_ty_id = try self.ptrType(ty, .Function); + const ptr_fn_ty_id = try self.ptrType(ty, .Function, .indirect); // SPIR-V requires that OpVariable declarations for locals go into the first block, so we are just going to // directly generate them into func.prologue instead of the body. @@ -5475,7 +5335,7 @@ const NavGen = struct { switch (options.storage_class) { .Generic => { - const ptr_gn_ty_id = try self.ptrType(ty, .Generic); + const ptr_gn_ty_id = try self.ptrType(ty, .Generic, .indirect); // Convert to a generic pointer return self.castToGeneric(ptr_gn_ty_id, var_id); }, @@ -5724,7 +5584,7 @@ const NavGen = struct { assert(cf.block_stack.items.len > 0); // Check if the target of the branch was this current block. - const this_block = try self.constInt(Type.u32, @intFromEnum(inst), .direct); + const this_block = try self.constInt(Type.u32, @intFromEnum(inst)); const jump_to_this_block_id = self.spv.allocId(); const bool_ty_id = try self.resolveType(Type.bool, .direct); try self.func.body.emit(self.spv.gpa, .OpIEqual, .{ @@ -5804,7 +5664,7 @@ const NavGen = struct { try self.store(operand_ty, block_result_var_id, operand_id, .{}); } - const next_block = try self.constInt(Type.u32, @intFromEnum(br.block_inst), .direct); + const next_block = try self.constInt(Type.u32, @intFromEnum(br.block_inst)); try self.structuredBreak(next_block); }, .unstructured => |cf| { @@ -5968,7 +5828,7 @@ const NavGen = struct { // Functions with an empty error set are emitted with an error code // return type and return zero so they can be function pointers coerced // to functions that return anyerror. - const no_err_id = try self.constInt(Type.anyerror, 0, .direct); + const no_err_id = try self.constInt(Type.anyerror, 0); return try self.func.body.emit(self.spv.gpa, .OpReturnValue, .{ .value = no_err_id }); } else { return try self.func.body.emit(self.spv.gpa, .OpReturn, {}); @@ -5992,7 +5852,7 @@ const NavGen = struct { // Functions with an empty error set are emitted with an error code // return type and return zero so they can be function pointers coerced // to functions that return anyerror. - const no_err_id = try self.constInt(Type.anyerror, 0, .direct); + const no_err_id = try self.constInt(Type.anyerror, 0); return try self.func.body.emit(self.spv.gpa, .OpReturnValue, .{ .value = no_err_id }); } else { return try self.func.body.emit(self.spv.gpa, .OpReturn, {}); @@ -6026,7 +5886,7 @@ const NavGen = struct { else err_union_id; - const zero_id = try self.constInt(Type.anyerror, 0, .direct); + const zero_id = try self.constInt(Type.anyerror, 0); const is_err_id = self.spv.allocId(); try self.func.body.emit(self.spv.gpa, .OpINotEqual, .{ .id_result_type = bool_ty_id, @@ -6134,7 +5994,8 @@ const NavGen = struct { types[eu_layout.errorFieldIndex()] = Type.anyerror; types[eu_layout.payloadFieldIndex()] = payload_ty; - return try self.constructStruct(err_union_ty, &types, &members); + const err_union_ty_id = try self.resolveType(err_union_ty, .direct); + return try self.constructComposite(err_union_ty_id, &members); } fn airWrapErrUnionPayload(self: *NavGen, inst: Air.Inst.Index) !?IdRef { @@ -6145,18 +6006,19 @@ const NavGen = struct { const eu_layout = self.errorUnionLayout(payload_ty); if (!eu_layout.payload_has_bits) { - return try self.constInt(Type.anyerror, 0, .direct); + return try self.constInt(Type.anyerror, 0); } var members: [2]IdRef = undefined; - members[eu_layout.errorFieldIndex()] = try self.constInt(Type.anyerror, 0, .direct); + members[eu_layout.errorFieldIndex()] = try self.constInt(Type.anyerror, 0); members[eu_layout.payloadFieldIndex()] = try self.convertToIndirect(payload_ty, operand_id); var types: [2]Type = undefined; types[eu_layout.errorFieldIndex()] = Type.anyerror; types[eu_layout.payloadFieldIndex()] = payload_ty; - return try self.constructStruct(err_union_ty, &types, &members); + const err_union_ty_id = try self.resolveType(err_union_ty, .direct); + return try self.constructComposite(err_union_ty_id, &members); } fn airIsNull(self: *NavGen, inst: Air.Inst.Index, is_pointer: bool, pred: enum { is_null, is_non_null }) !?IdRef { @@ -6204,7 +6066,7 @@ const NavGen = struct { if (is_pointer) { if (payload_ty.hasRuntimeBitsIgnoreComptime(zcu)) { const storage_class = self.spvStorageClass(operand_ty.ptrAddressSpace(zcu)); - const bool_ptr_ty_id = try self.ptrType(Type.bool, storage_class); + const bool_ptr_ty_id = try self.ptrType(Type.bool, storage_class, .indirect); const tag_ptr_id = try self.accessChain(bool_ptr_ty_id, operand_id, &.{1}); break :blk try self.load(Type.bool, tag_ptr_id, .{}); } @@ -6267,7 +6129,7 @@ const NavGen = struct { .id_result_type = bool_ty_id, .id_result = result_id, .operand_1 = error_id, - .operand_2 = try self.constInt(Type.anyerror, 0, .direct), + .operand_2 = try self.constInt(Type.anyerror, 0), }, ), } @@ -6335,8 +6197,8 @@ const NavGen = struct { const payload_id = try self.convertToIndirect(payload_ty, operand_id); const members = [_]IdRef{ payload_id, try self.constBool(true, .indirect) }; - const types = [_]Type{ payload_ty, Type.bool }; - return try self.constructStruct(optional_ty, &types, &members); + const optional_ty_id = try self.resolveType(optional_ty, .direct); + return try self.constructComposite(optional_ty_id, &members); } fn airSwitchBr(self: *NavGen, inst: Air.Inst.Index) !void { @@ -6752,13 +6614,13 @@ const NavGen = struct { fn builtin3D(self: *NavGen, result_ty: Type, builtin: spec.BuiltIn, dimension: u32, out_of_range_value: anytype) !IdRef { if (dimension >= 3) { - return try self.constInt(result_ty, out_of_range_value, .direct); + return try self.constInt(result_ty, out_of_range_value); } const vec_ty = try self.pt.vectorType(.{ .len = 3, .child = result_ty.toIntern(), }); - const ptr_ty_id = try self.ptrType(vec_ty, .Input); + const ptr_ty_id = try self.ptrType(vec_ty, .Input, .indirect); const spv_decl_index = try self.spv.builtin(ptr_ty_id, builtin); try self.func.decl_deps.put(self.spv.gpa, spv_decl_index, {}); const ptr = self.spv.declPtr(spv_decl_index).result_id; diff --git a/src/codegen/spirv/Module.zig b/src/codegen/spirv/Module.zig index f4af74b3aa..8e7f5ad4bc 100644 --- a/src/codegen/spirv/Module.zig +++ b/src/codegen/spirv/Module.zig @@ -10,6 +10,8 @@ const Module = @This(); const std = @import("std"); const Allocator = std.mem.Allocator; const assert = std.debug.assert; +const autoHashStrat = std.hash.autoHashStrat; +const Wyhash = std.hash.Wyhash; const spec = @import("spec.zig"); const Word = spec.Word; @@ -19,6 +21,19 @@ const IdResultType = spec.IdResultType; const Section = @import("Section.zig"); +/// Helper HashMap type to hash deeply +fn DeepHashMap(K: type, V: type) type { + return std.HashMapUnmanaged(K, V, struct { + pub fn hash(ctx: @This(), key: K) u64 { + _ = ctx; + var hasher = Wyhash.init(0); + autoHashStrat(&hasher, key, .Deep); + return hasher.final(); + } + pub const eql = std.hash_map.getAutoEqlFn(K, @This()); + }, std.hash_map.default_max_load_percentage); +} + /// This structure represents a function that isc in-progress of being emitted. /// Commonly, the contents of this structure will be merged with the appropriate /// sections of the module and re-used. Note that the SPIR-V module system makes @@ -159,8 +174,13 @@ cache: struct { // This cache is required so that @Vector(X, u1) in direct representation has the // same ID as @Vector(X, bool) in indirect representation. vector_types: std.AutoHashMapUnmanaged(struct { IdRef, u32 }, IdRef) = .empty, + array_types: std.AutoHashMapUnmanaged(struct { IdRef, IdRef }, IdRef) = .empty, + function_types: DeepHashMap(struct { IdRef, []const IdRef }, IdRef) = .empty, builtins: std.AutoHashMapUnmanaged(struct { IdRef, spec.BuiltIn }, Decl.Index) = .empty, + decorations: std.AutoHashMapUnmanaged(struct { IdRef, spec.Decoration }, void) = .empty, + + bool_const: [2]?IdRef = .{ null, null }, } = .{}, /// Set of Decls, referred to by Decl.Index. @@ -201,7 +221,10 @@ pub fn deinit(self: *Module) void { self.cache.int_types.deinit(self.gpa); self.cache.float_types.deinit(self.gpa); self.cache.vector_types.deinit(self.gpa); + self.cache.array_types.deinit(self.gpa); + self.cache.function_types.deinit(self.gpa); self.cache.builtins.deinit(self.gpa); + self.cache.decorations.deinit(self.gpa); self.decls.deinit(self.gpa); self.decl_deps.deinit(self.gpa); @@ -477,20 +500,69 @@ pub fn floatType(self: *Module, bits: u16) !IdRef { return entry.value_ptr.*; } -pub fn vectorType(self: *Module, len: u32, child_id: IdRef) !IdRef { - const entry = try self.cache.vector_types.getOrPut(self.gpa, .{ child_id, len }); +pub fn vectorType(self: *Module, len: u32, child_ty_id: IdRef) !IdRef { + const entry = try self.cache.vector_types.getOrPut(self.gpa, .{ child_ty_id, len }); if (!entry.found_existing) { const result_id = self.allocId(); entry.value_ptr.* = result_id; try self.sections.types_globals_constants.emit(self.gpa, .OpTypeVector, .{ .id_result = result_id, - .component_type = child_id, + .component_type = child_ty_id, .component_count = len, }); } return entry.value_ptr.*; } +pub fn arrayType(self: *Module, len_id: IdRef, child_ty_id: IdRef) !IdRef { + const entry = try self.cache.array_types.getOrPut(self.gpa, .{ child_ty_id, len_id }); + if (!entry.found_existing) { + const result_id = self.allocId(); + entry.value_ptr.* = result_id; + try self.sections.types_globals_constants.emit(self.gpa, .OpTypeArray, .{ + .id_result = result_id, + .element_type = child_ty_id, + .length = len_id, + }); + } + return entry.value_ptr.*; +} + +pub fn functionType(self: *Module, return_ty_id: IdRef, param_type_ids: []const IdRef) !IdRef { + const entry = try self.cache.function_types.getOrPut(self.gpa, .{ return_ty_id, param_type_ids }); + if (!entry.found_existing) { + const result_id = self.allocId(); + entry.value_ptr.* = result_id; + try self.sections.types_globals_constants.emit(self.gpa, .OpTypeFunction, .{ + .id_result = result_id, + .return_type = return_ty_id, + .id_ref_2 = param_type_ids, + }); + } + return entry.value_ptr.*; +} + +pub fn constBool(self: *Module, value: bool) !IdRef { + if (self.cache.bool_const[@intFromBool(value)]) |b| return b; + + const result_ty_id = try self.boolType(); + const result_id = self.allocId(); + self.cache.bool_const[@intFromBool(value)] = result_id; + + switch (value) { + inline else => |value_ct| try self.sections.types_globals_constants.emit( + self.gpa, + if (value_ct) .OpConstantTrue else .OpConstantFalse, + .{ + .id_result_type = result_ty_id, + .id_result = result_id, + }, + ), + } + + return result_id; +} + /// Return a pointer to a builtin variable. `result_ty_id` must be a **pointer** /// with storage class `.Input`. pub fn builtin(self: *Module, result_ty_id: IdRef, spirv_builtin: spec.BuiltIn) !Decl.Index { @@ -534,13 +606,17 @@ pub fn decorate( target: IdRef, decoration: spec.Decoration.Extended, ) !void { - try self.sections.annotations.emit(self.gpa, .OpDecorate, .{ - .target = target, - .decoration = decoration, - }); + const entry = try self.cache.decorations.getOrPut(self.gpa, .{ target, decoration }); + if (!entry.found_existing) { + try self.sections.annotations.emit(self.gpa, .OpDecorate, .{ + .target = target, + .decoration = decoration, + }); + } } /// Decorate a result-id which is a member of some struct. +/// We really don't have to and shouldn't need to cache this. pub fn decorateMember( self: *Module, structure_type: IdRef, |