ieee 754r 11/12/2003 1 the purpose of a standard is to - -bless the past -arbitrate the present...

11
EEE 754R 11/12/2003 1 The purpose of a standard is to - - Bless the past - Arbitrate the present - Avoid the future

Upload: daniella-wilkins

Post on 04-Jan-2016

213 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

The purpose of a standard is to -

- Bless the past

- Arbitrate the present

- Avoid the future

Page 2: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Givens -

- Performance matters- Fast is limited by Moore's Law- Many is limited by algorithm- The speed of light is constant

Page 3: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Present High Performance -

- 10's to 1000's of FPUs per chip- 10's to 100's of cycles across chip- Dynamic logic; no clock- Pure dataflow internal model- No precise exceptions- Some have no program at all- No notion of global time

Page 4: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

754 assumes global time -

- Modes, rounding and otherwise

- Flags

- Traps

- Control flow determinism

Page 5: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Dynamic modes (rounding etc.) -

- VERY expensive to come to ameaningful consistent global state

- VERY expensive to be able to cometo a consistent global state, whetheryou actually ever do or not

- No problem with static (known tocompiler) modes

Page 6: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Traps -

- Imprecise non-resuming traps areeasy and potentially useful

- Precise, resuming traps are hopelessor VERY expensive if taken

- Precise, resuming traps are hopelessor VERY expensive even if never taken

Page 7: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Control Flow Determinism -

- No notion of global time means thathardware computes some things inthe future – this is called speculation

- Sometimes the future doesn't happen

- Waiting for the future to get here isVERY expensive

- Discarding futures that didn't happenmust be cheap

Page 8: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Proposal – Modes:

- Retain modes

- Explicitly assert that the mode ofan operation must be determinableby an inspection of the programaccording to its language rules (i.e. static binding) and must notrequire its execution

Page 9: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Proposal – Flags:

- Retain flags- State the flag propagation rules- State that inspection points shall be

individually specified in the program text- Explicitly assert that the operations

which determine the value of the flagsat any inspection point must be determinable by an inspection of the program according to its language rules (i.e. static binding) and must notrequire its execution

Page 10: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Proposal – Speculation:

- Explicitly state that arbitrary speculationof operations is permitted, but not acrossinspection points present in the text of theprogram

- Explicitly state that the speculation stateof asynchronous parts of the program(including threads, and signal andexception handlers of all sorts) is undefined

Page 11: IEEE 754R 11/12/2003 1 The purpose of a standard is to - -Bless the past -Arbitrate the present -Avoid the future

IEEE 754R 11/12/2003 1

Ignore these issues, and 754R will -

- Bless the past

- Be flouted in the present

- Be a footnote in the future