Apparatus and method for reducing execution latency of floating point operations having special case operands
View Patent ↗An apparatus and method for floating-point special case handling. In one embodiment, a processor may include a first execution unit configured to execute a longer-latency floating-point instruction, and a second execution unit configured to execute a shorter-latency floating-point instruction. In response to the longer-latency floating-point instruction being issued to the first execution unit, the second execution unit may be further configured to detect whether a result of the longer-latency floating-point instruction is determinable from one or more operands of the longer-latency floating-point instruction independently of the first execution unit executing the longer-latency floating-point instruction. In response to detecting that the result is determinable, the second execution unit may be further configured to flush the longer-latency floating-point instruction from the first execution unit and to determine the result.
1. A processor, comprising: a first execution unit configured to execute a longer-latency floating-point instruction, wherein said longer-latency floating-point instruction defines a floating-point computation to be performed on one or more operands; and
a second execution unit configured to execute a shorter-latency floating-point instruction;
wherein said longer-latency floating-point instruction is configured to execute in a greater number of execution cycles than said shorter-latency floating-point instruction;
wherein in response to said longer-latency floating-point instruction being issued to said first execution unit, said second execution unit is further configured to detect from one or more operands of said longer-latency floating-point instruction whether a result of said longer-latency floating-point instruction is determinable without actually performing said floating-point computation to completion; and
wherein in response to detecting that said result is determinable, said second execution unit is further configured to flush said longer-latency floating-point instruction from said first execution unit and to determine said result.
2. The processor as recited in claim 1 , wherein in response to said second execution unit determining said result, execution of said longer-latency floating-point instruction completes in fewer execution cycles than if said first execution unit had determined said result.
3. The processor as recited in claim 1 , wherein said second execution unit detecting whether said result is determinable occurs concurrently with said first execution unit executing said longer-latency floating-point instruction.
4. The processor as recited in claim 1 , wherein said first execution unit is configured to execute floating-point division or square root instructions.
5. The processor as recited in claim 1 , wherein said second execution unit is configured to execute floating-point addition or multiplication instructions.
6. The processor as recited in claim 1 , wherein said second execution unit is further configured to determine an exception status of said longer-latency floating-point instruction in response to detecting that said result of said longer-latency instruction is determinable without actually performing said floating-point computation to completion.
7. The processor as recited in claim 1 , wherein said first execution unit is further configured to perform a mantissa arithmetic portion of said longer-latency floating-point instruction, and wherein said second execution unit is further configured to perform an exponent arithmetic portion of said longer-latency floating-point instruction.
8. The processor as recited in claim 1 , wherein execution behavior of said longer-latency and said shorter-latency floating point instructions is defined according to Institute of Electrical and Electronics Engineers (IEEE)754-1985 Standard for Binary Floating-Point Arithmetic.
9. The processor as recited in claim 1 , further comprising instruction fetch logic configured to issue a first instruction from one of a plurality of threads during one execution cycle and to issue a second instruction from another one of said plurality of threads during a successive execution cycle.
10. A method, comprising: issuing a longer-latency floating-point instruction to a first execution unit, wherein said longer-latency floating-Point instruction defines a floating-point computation to be performed on one or more operands;
in response to said issuing, detecting from one or more operands of said longer-latency floating-point instruction whether a result of said longer-latency floating-point instruction is determinable without actually performing said floating-point computation to completion; and
in response to detecting that said result is determinable, flushing said longer-latency floating-point instruction from said first execution unit and determining said result independently of said first execution unit.
11. The method as recited in claim 10 , wherein in response to determining said result independent of said first execution unit, execution of said longer-latency floating-point instruction completes in fewer execution cycles than if said first execution unit had determined said result.
12. The method as recited in claim 10 , wherein detecting whether said result is determinable occurs concurrently with said first execution unit executing said longer-latency floating-point instruction.
13. The method as recited in claim 10 , wherein said first execution unit is configured to execute floating-point division or square root instructions.
14. The method as recited in claim 10 , further comprising executing a shorter-latency floating-point instruction in a second execution unit, wherein said shorter-latency floating-point instruction includes a floating-point addition or multiplication instruction, and wherein said longer-latency floating-point instruction is configured to execute in a greater number of execution cycles than said shorter-latency floating-point instruction.
15. The method as recited in claim 14 , further comprising performing a mantissa arithmetic portion of said longer-latency floating-point instruction in said first execution unit, and performing an exponent arithmetic portion of said longer-latency floating-point instruction in said second execution unit.
16. The method as recited in claim 14 , wherein execution behavior of said longer-latency and said shorter-latency floating point instructions is defined according to Institute of Electrical and Electronics Engineers (IEEE)754-1985 Standard for Binary Floating-Point Arithmetic.
17. The method as recited in claim 10 , further comprising determining an exception status of said longer-latency floating-point instruction in response to detecting that said result of said longer-latency instruction is determinable without actually performing said floating-point computation to completion.
18. The method as recited in claim 10 , further comprising:
issuing a first instruction from one of a plurality of threads during one execution cycle; and
issuing a second instruction from another one of said plurality of threads during a successive execution cycle.
19. A system, comprising: a system memory; and a processor coupled to said system memory, wherein said processor comprises: a first execution unit configured to execute a longer-latency floating-point instruction, wherein said longer-latency floating-point instruction defines a floating-point computation to be performed on one or more operands; and
a second execution unit configured to execute a shorter-latency floating-point instruction;
wherein said longer-latency floating-point instruction is configured to execute in a greater number of execution cycles than said shorter-latency floating-point instruction;
wherein in response to said longer-latency floating-point instruction being issued to said first execution unit, said second execution unit is further configured to detect from one or more operands of said longer-latency floating-point instruction whether a result of said longer-latency floating-point instruction is determinable without actually performing said floating-point computation to completion; and
wherein in response to detecting that said result is determinable, said second execution unit is further configured to flush said longer-latency floating-point instruction from said first execution unit and to determine said result.
20. The system as recited in claim 19 , wherein in response to said second execution unit determining said result, execution of said longer-latency floating-point instruction completes in fewer execution cycles than if said first execution unit had determined said result.
21. The system as recited in claim 19 , wherein said second execution unit detecting whether said result is determinable occurs concurrently with said first execution unit executing said longer-latency floating-point instruction.
22. The system as recited in claim 19 , wherein said first execution unit is configured to execute floating-point division or square root instructions.
23. The system as recited in claim 19 , wherein said second execution unit is configured to execute floating-point addition or multiplication instructions.
24. The system as recited in claim 19 , wherein said second execution unit is further configured to determine an exception status of said longer-latency floating-point instruction in response to detecting that said result of said longer-latency instruction is determinable without actually performing said floating-point computation to completion.
25. The system as recited in claim 19 , wherein said first execution unit is further configured to perform a mantissa arithmetic portion of said longer-latency floating-point instruction, and wherein said second execution unit is further configured to perform an exponent arithmetic portion of said longer-latency floating-point instruction.
26. The system as recited in claim 19 , wherein execution behavior of said longer-latency and said shorter-latency floating point instructions is defined according to Institute of Electrical and Electronics Engineers (IEEE)754-1985 Standard for Binary Floating-Point Arithmetic.
27. The system as recited in claim 19 , wherein said processor further comprises instruction fetch logic configured to issue a first instruction from one of a plurality of threads during one execution cycle and to issue a second instruction from another one of said plurality of threads during a successive execution cycle.