When computing an address for a load or store, the segment base is added in before the address is sent to the CPU to actually perform the access. This has several advantages. First, because there are no alignment restrictions on segment bases, the virtual (pre-segmentation) address for an access could be aligned but produce an unaligned linear (post-segmentation) address. The opposite could also happen where an unaligned access becomes aligned. Once outside of the instruction, the majority of M5 doesn't know about segmentation and wouldn't be able to handle those sorts of situations. In CPUs like the O3 model which can do store to load forwarding, accesses to the same virtual address are expected to refer to the same piece of memory. By applying segment bases before the CPU gets the address, that remains true for x86. Because the base address for a particular segment isn't always used, both the intended base and the effective base are stored. The effective base is the value that's actually added into a virtual address.
Limit and attribute checks
Limit and attribute checks are performed by the TLB. The TLB is a convenient place to perform any type of address validation which could result in a fault, including these sorts of checks.
Architected user segments
These segments are for code or data and are specified by the ISA.
- CS - Code segment.
- DS - Default data segment.
- ES - Data segment. Implicitly used by string instructions.
- SS - Stack segment.
- FS, GS - Extra data segments.
M5 internal data segments
These segments are specific to M5 and are used internally.
- HS - A temporary segment register.
- LS - A segment register which always has a flat segment with base address 0.
Architected system segments
These segments are for system data structures and are specified by the ISA..
- TSL (LDT) - Local descriptor table.
- TSG (GDT) - Global descriptor table.
- TR (TSS) - Task state segment.
- IDTR (IDT) - Interrupt descriptor table.
The MS segment is not actually a segment. Instead, it's a flag to the TLB that a particular address should be decoded as an access to an alternative address space. All alternative address spaces are 32 bits wide, so the upper 32 bits are used to specify which one to use. The following prefixes have been defined:
- 0x100000000 - Originally planned for calls to CPUID. CPUID will likely be implemented as a special function unit/function, so this may never be used.
- 0x200000000 - Access to an MSR. The MSR number needs to be scaled by the size of a MiscReg so that M5 doesn't get confused and think adjacent MSRs overlap.
- 0x300000000 - IO port access.