Run Time Optimization 15-745: Optimizing Compilers Pedro Artigas Motivation A good reason Compiling a language that contains runtime constructs Java dynamic class loading Perl or Matlab eval(“statement”) Faster than interpreting A better reason May use program information only available at run time 2 Example of run-time information The processor that will be used to run the program inc ax is faster on a Pentium III add ax,1 is faster on a Pentium 4 No need to recompile if generating code at run time The actual program input/run-time behavior Is my profile information accurate for the current program input? YES! 3 The life cycle of a program Compile Link Load/Run One Object File One Binary One Process Global Analysis Whole Program Analysis Analysis? No observation! Larger scope, better information about program behavior 4 New strategies are possible Pessimistic x Optimistic approaches Ex: Does int *a points to the same location as int *b ? Compile time/Pessimistic: Prove that in ANY execution those pointers point to different addresses Run Time/Optimistic: Up to now in the current execution a and b point to different locations Assume this holds If the assumption breaks, invalidate generated code and generate new code 5 A sanity check Using run time information does not require run time code generation Example: Versioning ISA may allow cheaper tests IA-64 Transmeta if (a!=b) { <generate code assuming a!=b> } else { <generate code assuming a==b> } 6 Drawbacks Code generation has to be FAST Rule of thumb: almost linear on program size Code quality: Compromise on quality to achieve fast code generation shoot for good, not great Also this usually means: No time for classical Iterative Data Flow Analysis at run time 7 No classical IDFA: Solutions Quasi-Static and/or Staged Compilation Perform IDFA at compile time Specialize the dynamic code generator for the obtained information That is, encode the obtained data flow information in the “binary” Do not rely on classical IDFA Use algorithms that do not require it Ex: Dominator based value numbering (coming up!) Generate code in a style that does not require it Ex: One entry multiple exits traces as in deco and dynamo 8 Code generation Strategies Compiling a language that requires run-time code generation: Compile adaptively: Use a very simple and fast code generation scheme Re-compile frequently used regions using more advanced techniques 9 Adaptive Compilation: Motivation re-compilation 2 level recompilation Optimizing Only Fast Only Optimal(Oracle) total cost Very simple code generation Elaborate code generation Fast compiler Optimizing compiler Cost threshold Higher compilation cost Problem: execution count Higher execution cost We may not know in advance how frequently a region will execute Measure frequencies and re-compile dynamically 10 Code generation Strategies Compiling selected regions that benefit from run-time code generation: Pick only the regions that should benefit the most Which regions? Select them statically Use profile information Re-compile (that is select then dynamically) Usually all of the above 11 Code Optimization Unit 1 2 What is the run-time unit of optimization? 3 4 Option: Procedures/static code regions Option: Traces 1 2 3 4 4 Similar to static compilers Start at the target of a backward branch Include all the instructions in a path May include procedure calls and returns Branches Fall through = remain in the trace Target = exit the trace 12 Current strategies Static region Trace JIT compilers Java JITs Matlab JITs ? Run-time performance engines Dyc Fabius Dynamo Deco 13 Run-Time code generation: Case studies Two examples of algorithms that are suitable for run-time code generation Run time CSE/PRE replacement: Dominator based value numbering Run time Register Allocation: Linear scan register allocation 14 Sidebar With traces CSE/PRE become almost trivial No need for register allocation if optimizing a binary (ex: dynamo) A+B PRE A+B CSE A+B A+B 15 Review: Local value numbering Store expressions already computed (in a hash table) Store variable nameVN mapping in the VN array Store VNvariable name mapping in the Name array Same value numbersame value Expression was computed in the past, check if result is available New expression, add to the table for each basic block Table.empty() for each computed expression (“x=y op z”) if V=Table.lookup(“y op z”) VN[“x”]=V if VN[Name[V]]==V //expression is still there replace “x = y op z” with “x = Name[V]” else Name[V]=“x” else VN[“x”]=new_value_number() Table.insert(“y op z”,VN[“x”]) Name[VN[“x”]]=“x” 16 Local value numbering Works in linear time on program size Assuming accesses to the array and the hash table occur in constant time Can we make it work in a scope larger than a basic block? (Hint: Yes) What are the potential problems? 17 Problems How to propagate the hash table contents across basic blocks? How to make sure that is safe to access the location containing the expression in other basic blocks? How do we make sure if the location containing the expression is fresh? Remember: no IDFA 18 Control flow issues On split points things are simple Just keep the content of the hash table from the predecessor What about merge points? We do not know if the same expression was computed in all incoming paths We do not want to check the fact anyway (why?) Reset the state of the hash table to a safe state it had in the past Which program point in the past? The immediate dominator of the merge block 19 Data flow issues Making sure the def of an expression is fresh and reaches the blocks of interest How? By construction! SSA All names are fresh (Single Assignment) All defs dominate its’ uses (regular uses not functions) As, by construction, we introduce new defs using functions at every point this would not hold 20 Dominator/SSA based value numbering DVN(Block B) Table.PushScope() for each exp “n=(…)” if (exp is redundant or meaningless) //meaningless: (x0,x0) VN[“n”]= Table.lookup(“(…)” or “x0”) First process the remove(“n=(…)”) expressions else VN[“n”]=“n” Table.insert(“(…)”,VN[n]) for each exp “x=y op z” if (“v”=Table.lookup(“y op z”)) VN[“x”]=“v” Them the remove(“x=y op z”) regular ones else VN[“x”]=“x” Table.insert(“x=y op z”,VN[“x”]) for each successor s of B Propagate info Adjust the inputs about inputs for each dominator tree child c in CFG reverse post-order and call DVN DVN(c) recursively 21 Table.PopScope() VN Name VN Example u0 1 v0 u0=a0+b0 w0 v0=c0+d0 x0 w0=e0+f0 2 x0=c0+d0 y0=c0+d0 u2= (u0,u1) x2=(x0,x1) y2=(y0,y1) u3=a0+b0 3 y0 u1=a0+b0 u1 x1=e0+f0 x1 y1=e0+f0 y1 4 u2 x2 y2 u3 22 Problems Does not catch x0=a0+b0 x0=a0+b0 x1=a0+b0 x1=(x0,x2) x2=a0+b0 x2=(x0,x1) But it performs almost as well as CSE And runs much faster linear time ? (YES? NO?) 23 Homework #4 The DVN algorithm scans the CFG in a similar way as the second phase of SSA translation SSA translation phase #1 SSA translation phase #2 Placing functions assigning unique numbers to variables Combine both and save one pass Gives us a smaller constant But, at run time, it pays of! 24 Run time register allocation Graph Coloring? Not an option Even the simple stack based heuristic shown in class is O(n2) Not even counting: Building the graph Move coalescing optimization But register allocation is VERY important in terms of performance Remember, memory is REALLY slow We need a simple but effective (almost) linear time algorithm 25 Let’s start simple Start with a local (basic block) linear time algorithm Assuming only one def and one use per variable (More constrained than SSA) Assuming that if a variable is spilled it must remain spilled (Why?) Can we find an optimum linear time algorithm? (Hint: Yes) Ideas? Think about liveness first … 26 Simple Algorithm: Computing Liveness One def and one use per variable, only one block A live range is merely the interval between the def and the use Live Interval: Interval between the first def and the last use OBS: Live Range = Live Interval if there is no control flow, only one def and use We could compute live intervals using a linear scan if we store the def instructions (beginning of the interval) in a hash table 27 Example S1: S2: S3: S4: S5: S6: S7: S8: A=1 B=2 C=3 D=A E=B use(E) use(D) use(C) 28 Now Register Allocation Another linear scan Keep the active intervals in an list (active) Assumption: an interval, when spilled, will remain spilled Two scenarios #1: No problem #2: Must spill Which interval? | active | R | active | R 29 Spilling heuristic Since there is no second chance: That is a spilled variable will always remain spilled Spill the interval that ends last Intuition: As one spill must occur … Pick the one that makes the remaining allocation least constrained That is, the interval that ends last This is the provably optimum solution (given all the constraints) 30 Linear Scan Register Allocation active = {} freeregs = {all_registers} for each interval I (in order of increasing start point) for each interval J in active if J.end>I.start Expire old continue active.remove(J) intervals freeregs.insert(J.register) end for each interval J if active.length()==R spill_candidade=active.last(); if (spill_candidate.end>I.end) Must spill, pick I.register = spill_candidate.register either the last spill(spill_candidate) interval in active active.remove(spill_candidate) active.insert_sorted(I) //sorted by end point or the new else interval spill(I) else No I.register = freeregs.pop() //get any register from the free list active.insert_sorted(I) //sorted by end point constraints end for each interval I 31 Example (R=2) S1: S2: S3: S4: S5: S6: S7: S8: A A B C D E A=1 S1 B B=2 S2 C C=3 S3 D D=A S4 E E=B S5 use(E) S6 use(D) S7 use(C) S8 32 Is the second pass really linear? Invariant: active.length()<=R Complexity O(R*n) R is usually a small constant (128 at most) Therefore: O(n) 33 And we are done! Right? YES and NO Use the same algorithm as before for register assignment Program representation: Linear list of instructions Live intervals are not precise anymore given control flow and multiple def/uses Not optimum, but still FAST Code quality: within 10% of graph coloring for spec95 benchmarks (One problem with this claim) 34 The Worst problem: Obtaining precise live intervals How to obtain precise live interval information FAST? Claim of 10% relies on live interval information obtained using liveness analysis (IDFA) Most recent solutions: IDFA is SLOW, O(n3) Use the local interval algorithm for variables that only live inside one basic block Use liveness analysis for more global variables Alleviates the problem, does not fully solve it 35 More problems: Live intervals may not be precise OBS: The idea of lifetime holes leads to allocators that also try to use this holes to assign the same register to other live ranges (bin-packing) Such an allocator is used in the Alpha family of compilers (GEM compilers) 36 Other problems: Linearization order Register allocation quality depends on chosen block linearization order Choose a good order in practice layout order depth first traversal of the CFG Both only 10% slower than graph coloring 37 Graph coloring versus Linear scan Compilation cost scaling 38 Conclusion Run time code generation provides new optimization opportunities Challenges Identify new optimization opportunities Design new compilation strategies Design algorithms and implementations that: example: optimistic versus conservative minimize run time overhead Do not compromise much on code quality Recent examples indicate: extending fast local methods is a promising way to obtain fast run-time code generation 39