Cache rollback acceleration via a bank based versioning cache ciruit
An embodiment of the present invention includes a circuit for tracking memory operations with trace-based execution. Each trace includes a sequence of operations that includes zero or more of the memory operations. At least some of the active memory operations access the memory in an execution order that is different from the program order. The circuit includes a first memory that caches data accessed by the memory operations. This memory is partitioned into N banks. Checkpoint entries, which are stored in a second memory also partitioned into N banks, are associated with each trace. Each entry refers to a checkpoint location in the first memory. A sub-circuit receives rollback requests and responds by overwriting checkpoint locations. Each of the N memory units consisting of a bank in the first memory and the corresponding bank in the second memory may be rolled back independently and concurrently with other memory units.
1. A circuit for tracking memory operations in an out-of-order and speculative processor, where the circuit is adapted for use with a trace unit and with an execution unit of the processor, where the trace unit sequences traces for execution thereby forming a trace sequence order, where each of the traces includes a sequence of operations that includes zero or more of the memory operations, where the execution unit executes the operations and has an architectural state, and where the circuit comprises:
a first memory configured to cache at least some of the data accessed by the memory operations, where the memory operations being executed form a set of active memory operations, where the active memory operations have a predefined program order among themselves, where the predefined program order imposes a set of ordering constraints, where at least some of the active memory operations access the memory in an execution order that is different from the predefined program order, and where the first memory is partitioned into a number N of banks;
a second memory configured to receive and hold a set of checkpoint entries, where each checkpoint entry is associated with one of the traces, where each checkpoint entry is of a checkpoint location within the first memory, where each checkpoint entry includes checkpoint data indicating a state of the checkpoint location within the first memory prior to an update of the checkpoint location, and where the update results from executing one of the active memory operations, where the second memory is partitioned into the number N of sub-units, and where each of the sub-units forms a memory unit with a corresponding one of the banks;
a first sub-circuit configured to receive a rollback request, and, in response thereto, to overwrite in the first memory a set of the checkpoint locations, where each one of the memory units is rolled back independently of the other memory units, and each one is rolled back concurrently with the other memory units until such time as each one completes rollback.
2. The circuit of claim 1 , where:
none of the operations of a particular one of the traces has any effect on the architectural state prior to a commitment of the first particular trace, and where the first particular trace becomes eligible for the commitment after the execution completes for all operations of the first particular trace; and
none of the operations of a particular one of the traces has any effect on the architectural state, provided that the particular trace does abort.
3. The circuit of claim 1 , where the execution of the operations of a particular one of the traces can complete in such a way that a first subset of the operations of the particular trace affect the architectural state and that a second subset of the operations of the particular trace do not affect the architectural state.
4. The circuit of claim 1 , where the operation sequences of a set of at least two of the traces are executed concurrently by the execution unit, where the active memory operations include a memory operation from a first particular one of the set, and where the active memory operations include a memory operation from a second particular one of the set.