1. D A C U C P Speculative Alias Analysis for Executable Code Manel Fernández and Roger Espasa Computer Architecture Department Universitat Politècnica de Catalunya Barcelona, Spain {mfernand,roger}@ac.upc.es

2. D A C U C P Motivation Alias analysis Provides information for memory disambiguation  Key issue in today’s optimizing compilers Formulated as a dataflow analysis  In terms of source language constructs  Trade-off between cost and precision Executable code optimizers New optimization opportunities appear  Whole program view, etc. Limited usefulness of “traditional” analyses  High level information is lost  Object code is larger than corresponding source code Memory disambiguation is one of the weak points of object code modification

3. D A C U C P Alias analysis on executable code Existing techniques Instruction inspection  Register use-def chains Residue-based alias analysis  [Debray et al. POPL ‘98]  Arithmetic computations mod-k –Small displacements from a base register  “Fine grain” disambiguation Problems of existing alias analysis Low precision  Conservative assumptions Strong resource-usage constraints  Widening: less precision joining different definitions  Context-insensitive formulation

4. D A C U C P Talk outline Motivation Speculative alias analysis Evaluation Summary

5. D A C U C P Speculative alias analysis Existing analysis are conservative Trade off between cost and precision A new variable: safeness Analysis becomes speculative  Increases precision at low cost  Analysis is not always correct Our proposals Two may-alias approaches  Region-based alias analysis  Profile-guided alias analysis Well-suited for speculative optimizations  Speculative reordering based on check-and-recovery schemes  E.g.: reordering memory operations Safeness Precision Resource usage (cost) Resource usage (cost) Alias Analysis Alias Analysis

6. D A C U C P Propagating memory regions Observations Conservative propagation of descriptors  E.g.: operating different definitions Loads produce conservative descriptors  Nothing is known about the loaded value Key ideas Propagation without losing precision  Propagating “very basic” information “Guessing” possible memory regions  Memory regions “are disjoint” –Global –Stack –Heap... I 2 load (sp),r1 I 2 add gp,r1,r1 I 1 store r0,(r1) I 2 load (sp),r0...

7. D A C U C P Region-based alias analysis Interprocedural low-level scheme Well-suited for executable code Computation of memory regions For each register r defines S  S : set of regions { Global, Stack, Heap }  T = ,  = { G,S,H } Dataflow propagation Speculative approach  Assumptions are not always correct “Coarse-grain” disambiguation  Can be applied coupled to a residue-based scheme Aggressive region-based analysis Loaded values are hardly ever aliased with other pointers  E.g.: linked lists Set load destination descriptors to T instead of ... I 2 load (sp),r1 I 2 add gp,r1,r1 I 1 store r0,(r1) I 2 load (sp),r0...

8. D A C U C P Propagating likely paths Observations Widening operation is conservative  E.g.: several definitions reaching a use Context-insensitive analysis  Context-sensitive is not feasible Key ideas Propagation without loosing precision  Reducing the number of paths “Cold” references are not important  Conflicts are not significant at run time  Ignore “cold” paths I 1 store r0,(r1) I 2 load (sp),r0... add sp,0,r1... add gp,0,r1 Hot path

9. D A C U C P Profile-guided alias analysis Interprocedural general scheme Needs profile information Only likely executed paths are considered Applied on top of any dataflow analysis  Redefine join operation Speculative approach  “Cold” paths are ignored “Likely-path” disambiguation

10. D A C U C P A combined algorithm Alias analysis scheme Phase 1  Use-def chains Phase 2  Residue-based  Region-based Phase 3  Profile-guided Phase 2 Disambiguation scheme Input: I 1,I 2 Output: {dependent, independent, likely independent, unknown} Method: if ud-chains (I 1,I 2 ) ≠ unknown return ud-chains (I 1,I 2 ); if aliasing (I 1,I 2,safe) ≠ unknown return aliasing (I 1,I 2,safe); if aliasing (I 1,I 2,unsafe) ≠ unknown return likely independent; else return unknown; End Method

11. D A C U C P Talk outline Motivation Speculative alias analysis Evaluation Summary

12. D A C U C P Methodology Benchmark suite SPECint95  Compiled on an AlphaServer with full optimizations  Intrumented using Pixie to get profiling information Experimental framework Alto executable optimizer SimpleScalar safe simulator Evaluation Static precision Misspeculation rate

13. D A C U C P Evaluating effectiveness Disambiguation query “Question” made to the memory disambiguator Relationship between two memory references  Dependent, independent, likely independent  Unknown Set of disambiguation queries Pair of references that belong to the same function Building the set  Consider every load/store in a hot path (2nd item)  Consider every load/store in previous paths (1st item) Typical behavior of compiler optimizations

14. D A C U C P Static precision

15. D A C U C P Misspeculation rate

16. D A C U C P Talk outline Motivation Speculative alias analysis Evaluation Summary

17. D A C U C P Summary Memory disambiguation is one of the weak points of object code modification Speculative alias analysis Compromise among cost, precision and safeness  Region-based alias analysis  Profile-guided alias analysis Well-suited for executable code Conclusions Precision increases from 16% up to 83% in average Misspeculation rate around 1%  In front of 2% using an “always speculate” scheme To be used for reordering memory operations with high recovery cost