authorgravatar for jacobly@ziglang.orgJacob Young <jacobly@ziglang.org> 2025-01-02 04:42:27-05:00
committergravatar for jacobly@ziglang.orgJacob Young <jacobly@ziglang.org> 2025-01-16 20:42:08-05:00
logdf7661b8d31f2a49ea2679d6edc433cc1ab53e80
tree14b55f12d579416a4557d2603b1dc9b78cbf0095
parente5d5a8bc4ea6b27dc3540ad4800a1231ff50b33d

x86_64: optimize value copying slightly


1 files changed, 68 insertions(+), 9 deletions(-)

src/arch/x86_64/CodeGen.zig+68-9
......@@ -327,6 +327,13 @@ pub const MCValue = union(enum) {
327327 };
328328 }
329329
330 fn isAddress(mcv: MCValue) bool {
331 return switch (mcv) {
332 .immediate, .register, .register_offset, .lea_frame => true,
333 else => false,
334 };
335 }
336
330337 fn address(mcv: MCValue) MCValue {
331338 return switch (mcv) {
332339 .none,
......@@ -23005,17 +23012,69 @@ fn genSetMem(
2300523012
2300623013 try self.genSetMem(base, disp, ty, .{ .register = src_reg }, opts);
2300723014 },
23008 else => try self.genInlineMemcpy(
23009 dst_ptr_mcv,
23010 src_mcv.address(),
23011 .{ .immediate = abi_size },
23012 ),
23015 else => try self.genInlineMemcpy(dst_ptr_mcv, src_mcv.address(), .{ .immediate = abi_size }, .{ .no_alias = true }),
2301323016 },
2301423017 .air_ref => |src_ref| try self.genSetMem(base, disp, ty, try self.resolveInst(src_ref), opts),
2301523018 }
2301623019}
2301723020
23018fn genInlineMemcpy(self: *CodeGen, dst_ptr: MCValue, src_ptr: MCValue, len: MCValue) InnerError!void {
23021fn genInlineMemcpy(self: *CodeGen, dst_ptr: MCValue, src_ptr: MCValue, len: MCValue, opts: struct {
23022 no_alias: bool,
23023}) InnerError!void {
23024 if (opts.no_alias and dst_ptr.isAddress() and src_ptr.isAddress()) switch (len) {
23025 else => {},
23026 .immediate => |len_imm| switch (len_imm) {
23027 else => {},
23028 1 => if (self.register_manager.tryAllocReg(null, abi.RegisterClass.gp)) |reg| {
23029 try self.asmRegisterMemory(.{ ._, .mov }, reg.to8(), try src_ptr.deref().mem(self, .{ .size = .byte }));
23030 try self.asmMemoryRegister(.{ ._, .mov }, try dst_ptr.deref().mem(self, .{ .size = .byte }), reg.to8());
23031 return;
23032 },
23033 2 => if (self.register_manager.tryAllocReg(null, abi.RegisterClass.gp)) |reg| {
23034 try self.asmRegisterMemory(.{ ._, .mov }, reg.to16(), try src_ptr.deref().mem(self, .{ .size = .word }));
23035 try self.asmMemoryRegister(.{ ._, .mov }, try dst_ptr.deref().mem(self, .{ .size = .word }), reg.to16());
23036 return;
23037 },
23038 4 => if (self.register_manager.tryAllocReg(null, abi.RegisterClass.gp)) |reg| {
23039 try self.asmRegisterMemory(.{ ._, .mov }, reg.to32(), try src_ptr.deref().mem(self, .{ .size = .dword }));
23040 try self.asmMemoryRegister(.{ ._, .mov }, try dst_ptr.deref().mem(self, .{ .size = .dword }), reg.to32());
23041 return;
23042 },
23043 8 => if (self.target.cpu.arch == .x86_64) {
23044 if (self.register_manager.tryAllocReg(null, abi.RegisterClass.gp)) |reg| {
23045 try self.asmRegisterMemory(.{ ._, .mov }, reg.to64(), try src_ptr.deref().mem(self, .{ .size = .qword }));
23046 try self.asmMemoryRegister(.{ ._, .mov }, try dst_ptr.deref().mem(self, .{ .size = .qword }), reg.to64());
23047 return;
23048 }
23049 },
23050 16 => if (self.hasFeature(.avx)) {
23051 if (self.register_manager.tryAllocReg(null, abi.RegisterClass.sse)) |reg| {
23052 try self.asmRegisterMemory(.{ .v_dqu, .mov }, reg.to128(), try src_ptr.deref().mem(self, .{ .size = .xword }));
23053 try self.asmMemoryRegister(.{ .v_dqu, .mov }, try dst_ptr.deref().mem(self, .{ .size = .xword }), reg.to128());
23054 return;
23055 }
23056 } else if (self.hasFeature(.sse2)) {
23057 if (self.register_manager.tryAllocReg(null, abi.RegisterClass.sse)) |reg| {
23058 try self.asmRegisterMemory(.{ ._dqu, .mov }, reg.to128(), try src_ptr.deref().mem(self, .{ .size = .xword }));
23059 try self.asmMemoryRegister(.{ ._dqu, .mov }, try dst_ptr.deref().mem(self, .{ .size = .xword }), reg.to128());
23060 return;
23061 }
23062 } else if (self.hasFeature(.sse)) {
23063 if (self.register_manager.tryAllocReg(null, abi.RegisterClass.sse)) |reg| {
23064 try self.asmRegisterMemory(.{ ._ps, .movu }, reg.to128(), try src_ptr.deref().mem(self, .{ .size = .xword }));
23065 try self.asmMemoryRegister(.{ ._ps, .movu }, try dst_ptr.deref().mem(self, .{ .size = .xword }), reg.to128());
23066 return;
23067 }
23068 },
23069 32 => if (self.hasFeature(.avx)) {
23070 if (self.register_manager.tryAllocReg(null, abi.RegisterClass.sse)) |reg| {
23071 try self.asmRegisterMemory(.{ .v_dqu, .mov }, reg.to256(), try src_ptr.deref().mem(self, .{ .size = .yword }));
23072 try self.asmMemoryRegister(.{ .v_dqu, .mov }, try dst_ptr.deref().mem(self, .{ .size = .yword }), reg.to256());
23073 return;
23074 }
23075 },
23076 },
23077 };
2301923078 try self.spillRegisters(&.{ .rsi, .rdi, .rcx });
2302023079 try self.genSetReg(.rsi, .usize, src_ptr, .{});
2302123080 try self.genSetReg(.rdi, .usize, dst_ptr, .{});
......@@ -24057,7 +24116,7 @@ fn airMemset(self: *CodeGen, inst: Air.Inst.Index, safety: bool) !void {
2405724116 len_reg,
2405824117 .s(elem_abi_size),
2405924118 );
24060 try self.genInlineMemcpy(second_elem_ptr_mcv, dst_ptr, len_mcv);
24119 try self.genInlineMemcpy(second_elem_ptr_mcv, dst_ptr, len_mcv, .{ .no_alias = false });
2406124120
2406224121 self.performReloc(skip_reloc);
2406324122 },
......@@ -24082,7 +24141,7 @@ fn airMemset(self: *CodeGen, inst: Air.Inst.Index, safety: bool) !void {
2408224141 } }, .{});
2408324142
2408424143 const bytes_to_copy: MCValue = .{ .immediate = elem_abi_size * (len - 1) };
24085 try self.genInlineMemcpy(second_elem_ptr_mcv, dst, bytes_to_copy);
24144 try self.genInlineMemcpy(second_elem_ptr_mcv, dst, bytes_to_copy, .{ .no_alias = false });
2408624145 },
2408724146 .c, .many => unreachable,
2408824147 }
......@@ -24165,7 +24224,7 @@ fn airMemcpy(self: *CodeGen, inst: Air.Inst.Index) !void {
2416524224 else => src,
2416624225 };
2416724226
24168 try self.genInlineMemcpy(dst_ptr, src_ptr, len);
24227 try self.genInlineMemcpy(dst_ptr, src_ptr, len, .{ .no_alias = true });
2416924228
2417024229 return self.finishAir(inst, .unreach, .{ bin_op.lhs, bin_op.rhs, .none });
2417124230}