authorgravatar for jacobly@ziglang.orgJacob Young <jacobly@ziglang.org> 2023-04-25 16:47:44-04:00
committergravatar for jacobly@ziglang.orgJacob Young <jacobly@ziglang.org> 2023-04-26 19:05:17-04:00
logeff0e6a726e46901bf8dededab9ef8db78398811
treef44c0baf4094900de7356c93b0add66d503b13a1
parenta8842b6cbf4242c2cfec203b21149872a6d2e246

x86_64: fix rem/mod behavior and hazards


1 files changed, 49 insertions(+), 10 deletions(-)

src/arch/x86_64/CodeGen.zig+49-10
......@@ -2601,7 +2601,13 @@ fn genIntMulDivOpMir(
26012601 else => unreachable,
26022602 .mul, .imul => {},
26032603 .div => try self.asmRegisterRegister(.xor, .edx, .edx),
2604 .idiv => try self.asmOpOnly(.cqo),
2604 .idiv => switch (self.regBitSize(ty)) {
2605 8 => try self.asmOpOnly(.cbw),
2606 16 => try self.asmOpOnly(.cwd),
2607 32 => try self.asmOpOnly(.cdq),
2608 64 => try self.asmOpOnly(.cqo),
2609 else => unreachable,
2610 },
26052611 }
26062612
26072613 const mat_rhs: MCValue = switch (rhs) {
......@@ -2631,7 +2637,8 @@ fn genIntMulDivOpMir(
26312637/// Always returns a register.
26322638/// Clobbers .rax and .rdx registers.
26332639fn genInlineIntDivFloor(self: *Self, ty: Type, lhs: MCValue, rhs: MCValue) !MCValue {
2634 const signedness = ty.intInfo(self.target.*).signedness;
2640 const abi_size = @intCast(u32, ty.abiSize(self.target.*));
2641 const int_info = ty.intInfo(self.target.*);
26352642 const dividend: Register = switch (lhs) {
26362643 .register => |reg| reg,
26372644 else => try self.copyToTmpRegister(ty, lhs),
......@@ -2646,16 +2653,32 @@ fn genInlineIntDivFloor(self: *Self, ty: Type, lhs: MCValue, rhs: MCValue) !MCVa
26462653 const divisor_lock = self.register_manager.lockReg(divisor);
26472654 defer if (divisor_lock) |lock| self.register_manager.unlockReg(lock);
26482655
2649 try self.genIntMulDivOpMir(switch (signedness) {
2656 try self.genIntMulDivOpMir(switch (int_info.signedness) {
26502657 .signed => .idiv,
26512658 .unsigned => .div,
2652 }, Type.isize, .{ .register = dividend }, .{ .register = divisor });
2659 }, ty, .{ .register = dividend }, .{ .register = divisor });
26532660
2654 try self.asmRegisterRegister(.xor, divisor.to64(), dividend.to64());
2655 try self.asmRegisterImmediate(.sar, divisor.to64(), Immediate.u(63));
2656 try self.asmRegisterRegister(.@"test", .rdx, .rdx);
2657 try self.asmCmovccRegisterRegister(divisor.to64(), .rdx, .e);
2658 try self.genBinOpMir(.add, Type.isize, .{ .register = divisor }, .{ .register = .rax });
2661 try self.asmRegisterRegister(
2662 .xor,
2663 registerAlias(divisor, abi_size),
2664 registerAlias(dividend, abi_size),
2665 );
2666 try self.asmRegisterImmediate(
2667 .sar,
2668 registerAlias(divisor, abi_size),
2669 Immediate.u(int_info.bits - 1),
2670 );
2671 try self.asmRegisterRegister(
2672 .@"test",
2673 registerAlias(.rdx, abi_size),
2674 registerAlias(.rdx, abi_size),
2675 );
2676 try self.asmCmovccRegisterRegister(
2677 registerAlias(divisor, abi_size),
2678 registerAlias(.rdx, abi_size),
2679 .z,
2680 );
2681 try self.genBinOpMir(.add, ty, .{ .register = divisor }, .{ .register = .rax });
26592682 return MCValue{ .register = divisor };
26602683}
26612684
......@@ -4928,8 +4951,24 @@ fn genMulDivBinOp(
49284951
49294952 switch (signedness) {
49304953 .signed => {
4954 const lhs_lock = switch (lhs) {
4955 .register => |reg| self.register_manager.lockReg(reg),
4956 else => null,
4957 };
4958 defer if (lhs_lock) |lock| self.register_manager.unlockReg(lock);
4959 const rhs_lock = switch (rhs) {
4960 .register => |reg| self.register_manager.lockReg(reg),
4961 else => null,
4962 };
4963 defer if (rhs_lock) |lock| self.register_manager.unlockReg(lock);
4964
4965 // hack around hazard between rhs and div_floor by copying rhs to another register
4966 const rhs_copy = try self.copyToTmpRegister(ty, rhs);
4967 const rhs_copy_lock = self.register_manager.lockRegAssumeUnused(rhs_copy);
4968 defer self.register_manager.unlockReg(rhs_copy_lock);
4969
49314970 const div_floor = try self.genInlineIntDivFloor(ty, lhs, rhs);
4932 try self.genIntMulComplexOpMir(ty, div_floor, rhs);
4971 try self.genIntMulComplexOpMir(ty, div_floor, .{ .register = rhs_copy });
49334972 const div_floor_lock = self.register_manager.lockReg(div_floor.register);
49344973 defer if (div_floor_lock) |lock| self.register_manager.unlockReg(lock);
49354974