TOR AAMODT

1
The Predictability of Computations that Produce
Unpredictable Outcomes

• TOR AAMODT
• (aamodt_at_eecg.utoronto.ca
• Andreas Moshovos Paul Chow
• Electrical and Computer Engineering
• University of Toronto
• Canada

2
Outcome-Based Prediction
History of Outcomes leading up to Branch
X TNTTNTT ...NTN... TNTTNTT
History
Next time we encounter X after TNTTNT we can
predict T
Outcome of Branch X
Why this works Locality in the outcome stream
3
Problem

• Unpredictable Branches THE Problem.
• No Outcome-Locality

4
Operation-Based Prediction
add
ld
slt
bne

• Find locality in the computations that produce
  the outcome

5
This Work

• First work that looks at the fundamental program
  behaviour that would facilitate operation-based
  prediction.
• Related work
• Characterization of slices
• Prefetching loads / pre-execution of branches

6
Ideally...

• Slice (i.e., slice trace) will always be the
  same.
• Slice will contain very few operations spanning
  large portion of original program.
• Easy (fast) to pre-compute.

7
Terminology

• Lead earliest instruction in slice
• Target branch we want to precompute

add
ld
slt
bne
8
What Should a Slice be?

• Commited Instructions
• 32, 64, 128, or 256 window
• Ignore Control Flow
• retain side-effect of JAL on r31
• Memory Dependence
• follow resolved load-store dependence M
• Restrict Instructions
• R max 1/4, U no restriction

FETCH ... COMMIT
older
9
Methodology

• 12 programs from SPEC2000
• Baseline Outcome Prediction Hardware
• 64K Gshare 64K bimodal w/ 64K selector
• 64 entry RAS
• sim-outorder (SimpleScalar 3.0)
• 8-way, 128 entry RUU, 64 entry-fetch buffer
• 64K dual LI, 256K unified L2
• 64 entry LSQ
• Perfect Memory Disambiguation

10
Measuring Slice Locality

• locality(1) Probability same slice was seen
  last time. High value of locality(1) indicates
  that last-operation based slice prediction would
  work well.
• locality(N) Probability same slice seen in last
  N unique slices.

11
Measuring Slice Locality

• Save the FOUR unique, most recent slice traces
  per static branch (only on misprediction).
• Each time a mispredicted branch is encountered
  check whether the slice trace was the most
  recent, 2nd most recent, etc...

12
Measuring Slice Locality

• All results are weighted averages.
• Result for each static branch weighted
  proportionally to the number of times the
  operation-based predictor mispredicted it.
• Characteristics of branches that cause most
  mispredictions emphasized.

13
Unrestricted Slices 32UM
Better
Locality
gcc equake ammp
bzip
Saving ONE slice captures most of locality.
14
Restricted vs. Unrestricted
Better
Locality
32UM
32RM
gcc equake ammp
bzip
Most slices have few instructions.
15
Effect of Memory Dependence
Better
Locality
64R
64RM
gcc equake ammp
bzip
Tracking Dependence Does Not Affect Locality Much.
16
Window Size
Better
Locality
32RM
64RM
128RM
256RM
gcc equake ammp
bzip
Locality good even for large windows.
17
Effect of Selection Context 128RM
Better
Locality
On Mispredict
Always
gcc equake ammp
bzip
Focusing on Mispredictions Improves Locality.
18
Idealized Predictor
Lead PC

• Spawn and execute instantaneously when lead
  operation is encountered.
• Store up to 4 slice traces per lead operation

19
Idealized Predictor

• Match operations register dependencies as
  instructions are fetched.
• After matching there is usually only one
  prediction per target, if any (gt80 of time)...
• Tie-breaker 1 longest lead-target distance.
• Tie-breaker 2 most recently detected slice.

20
Correcting Mispredictions
32RM
64RM
128RM
gcc equake ammp
bzip
High Coverage of Mispredicted Branches
21
Interaction with Outcome-Based Predictor
32RM
64RM
128RM
gcc equake ammp
bzip
Very Little Destructive Interference
22
Summary

• Slice-locality for mispredicted branches
• average of 70 for restricted slices on a 64
  entry window following load-store dependencies
  (12 SPEC2000 benchmarks).
• Accuracy of idealized predictor
• 74 of mispredicted branches eliminated

23
Conclusion

• First work that looks at the fundamental program
  behaviour, slice-locality, that would facilitate
  predicting slice traces to pre-execute outcomes.
• SPEC2000 benchmarks show very high slice-locality
  for mispredicted branches.