Operation frame filtering, building, and execution

• US 20070157188A1
• Filed: 12/30/2005
• Published: 07/05/2007
• Est. Priority Date: 12/30/2005
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Jourdan, Stephan, Farcy, Alexandre, Hammarlund, Per, Miller, John
• Granted Patent: US 7,533,247 B2
• Time in Patent Office: Days
• US Class Current: 717/159
• CPC Class Codes: G06F 9/3802 Instruction prefetching

Abstract

The present subject matter relates to operation frame filtering, building, and execution. Some embodiments include identifying a frame signature, counting a number of execution occurrences of the frame signature, and building a frame of operations to execute instead of operations identified by the frame signature.

Claims

1. A method comprising:

identifying a first sequence of instructions operations by an execution core and caching a frame signature, wherein the frame signature is a unique identifier of the first sequence of operations;

counting occurrences of the first sequence of operations and caching the count with the frame signature;

if the count reaches a threshold, building a frame at the next occurrence of the sequence of operations, wherein the frame includes a second sequence of operations;

detecting a subsequent need to execute the first sequence of operations; and

executing the second sequence of operations from the frame instead of the first sequence of operations.

2. The method of claim 1, wherein building a frame includes:

building the second sequence of operations as a function of the first sequence of instructions, wherein the second sequence of instructions causes the execution core to produce a same result as the first sequence of operations would have produced.

3. The method of claim 1, wherein the second sequence of operations includes the operations of the first sequence of operations reorganized to optimize execution core performance.

4. The method of claim 1, wherein the second sequence of operations includes a smaller number of operations than the first sequence of operations, wherein the second sequence of operations, when executed, produce the same result as the first sequence of operations.

5. The method of claim 1, wherein the first and second sequences of operations are sequences of uOP micro-operations.

6. The method of claim 1, further comprising:

storing the frame in a frame cache including two arrays, wherein;

the first array includes the frame signature and a second array pointer wherein the second array pointer points to a storage location of the second sequence of operations, and

the second array includes the second sequence of operations.

7. The method of claim 7, wherein the second array includes two arrays, wherein

a data linkage array providing links from line ends of second operation sequence lines in a data array to a start of a next line of the second operation sequence, and

the data array including the second sequence of operations stored in a plurality of lines.

8. The method of claim 7, wherein a length of data array lines is matched to the width of the execution core.

9. The method of claim 8, wherein the data array line length is the width of six uOP micro-operations.

10. A processor comprising:

a pipeline including;

a frame filter that receives output from pipeline retirement logic;

a frame cache that provides a cached operation sequence to the pipeline; and

a frame builder that can receive a copy of an operation sequence sent to an execution core, optimize the copy of the operation sequence, and store the optimized copy of the operation sequence to the frame cache as the cached operation sequence, wherein the cached operation sequence is a frame.

11. The processor of claim 10, wherein the frame filter:

identifies sequences of operations executed by the execution core and received from the pipeline retirement logic;

caches a frame signature for each sequence of operations, wherein the frame signature is a unique identifier of a sequence of operations;

counts occurrences of each sequence of operations and caches the count with the corresponding frame signature; and

if the count reaches a threshold for a given sequence of operations, the frame filter sends a command and a signature of the given sequence of operations to the frame cache to cause a frame to be built upon the next occurrence of the given sequence of operations.

12. The processor of claim 11, wherein the frame cache includes:

a tag array, wherein the tag array has lines including a frame signature and an address of a first line of a cached operation sequence in a data array;

the data array, wherein the data array includes a linkage array and a data array including operation lines, wherein the linkage array provides a link from a data array line end to a start of another data array line.

13. The processor of claim 12, wherein the width of a data array operation line is a width of the execution core.

14. The processor of claim 12, wherein the frame cache, upon request for an operation sequence for which it has a signature stored, determines if a frame has been built by the frame builder for the operation sequence, wherein:

if the frame has been built, the frame cache provides an operation sequence of the frame; and

if the frame has not been built, the frame cache issues an operation which causes the frame builder to copy a sequence of operations flowing through the pipeline before the sequence operations reaches the execution core.

15. The processor of claim 10, wherein the execution core is an out-of-order execution core.

16. The processor of claim 10, wherein a sequence of operations comprises a sequence of uOP micro-operations.

17. The processor of claim 10, further comprising:

a prediction unit, wherein the prediction unit receives input comprising input from an instruction cache and the frame cache.

18. The processor of claim 17, wherein the prediction unit is a branch prediction unit.

19. A system comprising:

a processor including a frame cache to provide an output in response to a request for an operation sequence, wherein the response is;

a cached operation sequence if the cached operation sequence is available,

a command to cause a frame to be built by another processor portion and stored in the frame cache, or

an indication that the frame cache does not have any information regarding the requested operation sequence; and

a graphics card operative in the system.

20. The system of claim 19, wherein the indication that the frame cache does not have any information regarding the requested operation sequence is not responding to the request.

21. The system of claim 19, wherein the frame cache includes:

a tag array, wherein the tag array has lines including a frame signature and an address of a first line of a cached operation sequence in a data array;

the data array, wherein the data array includes a linkage array and a data array including operation lines, wherein the linkage array provides a link from a data array line end to a start of another data array line.

22. The system of claim 19, wherein the command to cause a frame to be built by another processor portion causes a frame builder to copy an operation sequence sent to an execution core of the processor and store the copy of the operation sequence to the frame cache.

23. The system of claim 22, wherein the frame builder, prior to storing the copy of the operation sequence to the frame cache, modifies the copy of the operation sequence.

24. The system of claim 19, wherein a sequence of operations includes one or more processor micro-operations.

25. The system of claim 19, wherein the processor further includes:

a prediction unit that can receive input from an instruction cache and from the frame cache.

26. The system of claim 25, wherein the prediction unit is a branch prediction unit.

27. A method comprising:

generating an optimized frame as a function of a sequence of operations; and

storing the optimized frame in a frame cache, wherein subsequent occurrences of the sequence of operations are executed from the optimized frame instead of the actual sequence of operations.

28. The method of claim 27, wherein the sequence of operations is a sequence of micro-operations.

29. The method of claim 27, wherein an optimized frame is generated only when the sequence of operations is executed by a processor execution core an equal or greater number of times than a threshold number.

30. The method of claim 27, further comprising:

determining if a frame has been built for a sequence of operations prior to execution of the sequence of operations.