Potential for Dynamic code elimination

1
Potential for Dynamic code elimination

• presented by
• Martin Labrecque
• Peter Yiannacouras

2
Target Instructions (1)

• Dead instructions
• their output is overwritten without ever being
  used
• ADD r1 r2 r3
• .
• .
• .
• .
• MOV r1 0x0001

always unused or conditional to control path
3
Target Instructions (2)

• Silent instructions
• they don't change the value at their output
  location
• vect contains numbers lt 100
• for (i .. N)
• r1 vecti
• ltsilentgt r3 (r1 lt 100)? 0
  1
• ltsilentgt r5 r5 r3
• ltsilentgt r6 r5 4

dependence path
4
Target Instructions (3)

• Value predicted instructions
• their output is known in advance
• they may modify the state of the machine
• for (i .. N)
• MOV r1 r1 8
• ADD r2 cos(r1)

dependence path
5
Plan of talk

• Background about the target instructions
• How to deal with them
• A safe removal framework
• Result from removals
• Conclusions

6
How many are they?

• Silent 29
• Dead 5
• Predicting stores
• 33 less writeback
• Predicting loads
• 12 speedup
• how long to confirm a dead instruction?
• are the 3 types of instructions chained?
• if so, can we take advantage of that?

7
How removal works dead instructions
removed
D
feeds only
C
Dead ? removed
feeds no instruction
C
overwrites unused value
B

• Speculation ends
• ? when output is killed ? when
  assumptions fail

8
How removal works silent and value predictable
D
stays for verification
C
silent / predicted
silent / predicted
B
silent / predicted
A
9
Infrastructure

• SimpleScalar simulator
• Monitor all instructions
• Dependence graph around target instructions
• Set some history and speculation limits
• Don't remove
• branches
• instructions with no output (ex NOP)
• squashed instructions
• if only outputs to ZERO register(s)
• trap instructions (ignored)

10
Some results
11
ResultsPercent Dead/Silent/Predicted

• FP benches do well on silent/vapred
• mgrid has 45 on both

12
Number of Candidates

• Only a few instructions are candidates
• On average, 64 are live at a time (408)

13
Removal Reoccurrence
Distribution of Percent of Time a Candidate is
Removed

• Means at 62, 64, and 60 respectively

14
Predictor Size for Candidates

• Given that an instruction was removed on last
  iteration

15
Percent Memory Operations

• Usually not much
• Harder to track memory operations

16
Removals Per Chain

• Dead very rarely has more than one!
• Silent up to 12

17
Removals Per Chain Combined

• 5x more removals in total
• Up to 50 in a chain

18
Speculation Failure

• Unbounded speculation for dead
• Unbounded Rollback cost

19
Parameters for Dead Removal

• Exponentially decreasing chance of verifying dead
• Rollback cost increases
• Choose parameters to optimize

Distance Till Dead Speculation is confirmed
20
Parameters for Silent/Value Predicted Removal

• Choose so as to not hinder chaining
• too big gt wasted overhead

21
Conclusions

• Lots of Reuse gt Predictors neednt be large
• Can ignore memory operations with little loss
• Dead
• Safe recursive dead elimination is worthless
• Balance speculation rollback for single
  removals
• Silent/Value Predicted
• Chains of 12 can be removed
• Much more effective when combined (chains of 50)

22
Applications

• Helper threads
• Dont need guaranteed correctness
• Can use removals to speed ahead of real thread
• Program Phase Analysis
• Dynamic Specialization at instruction level

23
Extra slides
24
3 Parameters for Dead Removal
backward distance
forward distance
predecessor distance
25
1 Parameter for Silent and Predicted
backward distance
0 forward distance