Putting Pointer Analysis to Work Rakesh Ghiya and Laurie J. Hendren Presented by Shey Liggett & Jason Bartkowiak

Introduction  Paper addresses the problem of how to apply pointer analysis to a wide variety of compiler applications.  Shows how to put points-to analysis and connection analysis to work  Compute read/write sets for indirections stack-directed pointers: points-to information heap directed pointers: connection analysis + anchor handles  Based on the read/write sets extend traditional optimizations.

Stack vs. Heap Directed Pointers  Stack directed pointers: pointer to stack objects. Objects on the stack have appropriate variable names int t, *pt1; pt1 = &t;  Heap directed pointers: pointers to heap objects Dynamically-allocated objects int *pt2; pt2 = malloc();

Approach  Resolve all pointer relationships on the stack using points-to analysis  Further analyze heap pointers using connection analysis  Examine how the combination of the above analyses can be used to compute applicable information

Pointer Analysis in General  Identify the set of locations read/written by a given statement or program region. x S: x = y + z; Read(S) = {y,z} Write(S) = {x} *q T: p = *q; Read(T) = {q,*q} Write(T) = {p} … *q U: *q = y; Read(U) = {y,q} Write(U) = {*q}  In order to relate read/write sets of statements: resolve indirect references into a set of static locations

Points-to Analysis (As explained by Emami)  Approximate relationships between named objects (stored-based).  Calculate pointer targets in terms of triplets of the form (x, y, D) / (x, y, P) Variable x definitely/possibly contains the address of the location corresponding to y.  Heap locations are abstracted as one symbolic stack location named heap

Points-to Analysis (cont.) C: (s, ptA, D), (t, ptB, D) mapping: U: (c, 1_c, D), (d, 1_d, D) Read(U) = {c, d, 1_c.x, 1_d.y} Write(U): {1_c.x} Read(V) = {c, d, 1_c.y, 1_d.x} Write(V): {1_c.y} Read(sum) = Read(U) + Read(V) Write(sum): Write(U) +Write(V)

Points-to Analysis (cont.) Read(sum) = {c, d, 1_c.x, 1_d.y, 1_c.y, 1_d.x} Write(sum): {1_c.x, 1_c.y} Unmapping: Read(C) = {s, t, ptA.x, ptB.y, ptA.y, ptB.x} Write(C): {ptA.x, ptA.y}

Points-to Analysis (cont.) D: (s, heap, P), (t, heap, P) mapping: flip:(a, heap, P), (b, heap, P) Read(S) = Read(T) = {b, a, heap) Write(S) = Write(T) = {heap} False dependence between S and T

Connection Analysis  Computing connection relationships between pointers (instead of explicitly computing potential targets of pointers)  Performed after point-to analysis  Focuses on heap-directed pointers  Two heap directed pointers are connected if they possibly point to heap objects belonging to the same data structure.  They are NOT connected if they definitely point to objects belonging to disjoint data structures

Connection Analysis (cont.)

 Problem: computing read/write sets based on connection analysis

Introducing Anchor Handles  Motivation: The same programmer defined name may refer to different objects at different program points  Solution: Invent enough new names: Anchor handles  Calculating read/write sets: anchor handle p is read/written each time any pointer connected to p is read/written.

HeapWrite(S) = HeapRead(T) = Detect flow dependence from S to T Introducing Anchor Handles (cont.)

Function level information: HeapWrite(t_flip) = HeapRead(t_flip) = Useful to prefetch a->x and b->y (but not a->y and b->x) No changes to the “listness” of the data structure

Introducing Anchor Handles (cont.)  Select the locations to be anchored  Generate anchor handles for each:  heap directed formal parameters  heap directed global pointer accessed in function  call site that can read/write a heap location  heap related indirect reference - *p if (p,heap,P)  Use SSA numbers to further reduce number of anchors anchor the same location (pointer a hasn’t been updated between them) same handle can be used to anchor all indirect references involving a given definition of a pointer.

Applications - extend several scalar compiler optimizations  Loop Invariant Removal (LIR) Variables that do not change in a loop (always evaluate the same value). Remove from the loop.  Location Invariant Removal (LcIR) Memory reference that accesses the same memory location in all iterations of a loop.  Common Sub-expression Elimination (CSE) Computations that are always performed at least twice on a given execution path. Eliminate second and later occurrences.

Example of LIR For(I= )temp = *a; {for(I=) Array[I] = *a;{ }Array[I] = temp; } (a)(b) Loop Invariants

Example of LcIR

Another Example of LcIR For(I= )temp = r->t; {for(I=) r->t = p->I;{ }temp=p->I; } r->t = temp; (a)(b) Location Invariants

Applications - LIR, LcIR

Example of CSE For(I= )temp = (a*b)/c; {for(I=) Array[I] = (a*b)/c;{ Array2[I] = (a*b)/c; Array[I] = temp; }Array2[I] = temp; } (a)(b) Common Sub-expression Elimination

Applications - CSE

Experimental Results Analysis Efficiency (UltraSparc) quite efficient for moderate size benchmarks average number of anchor handles per indirect reference is 0.50

Experimental Results (cont.) Optimization Opportunities expr invariants cannot always be identified without pointer analysis limited applications of LcIR. Numerous applications for CSE, LIR

Experimental Results (cont.) Benefits of using heap read/write sets LIR and CSE: number of optimizations increases moderately, for all benchmarks; stack analysis is able to detect most of them (heap read/write info doesn’t bring any added advantage in the case of address exposed variables, or if the code fragment doesn’t involve any write to heap) LcIR: increases in the two applications.

Experimental Results (cont.) Runtime Improvement Measure additional benefits of the analyses over a state-of-the-art optimizing compiler

Experimental Results (cont.) Runtime Improvement Optimized versions achieve significant reduction in the number of memory references (7% to 35.56%). There may not always be a direct correlation between the number of times optimizations are applied and the actual run time improvement. Percentage decrease is always equal or higher for Hopt compared to Sopt. Some of the applications show significant speedup over “gcc -O3”

More Applications  Improving array dependence tests. In C, arrays are mostly implemented using pointers to dynamically-allocated storage. Pointer based array references pose problems for array dependence tester. Pointer analyses can make it more effective.  Program Understanding/Debugging. Based on the summary of read/write sets can make observations about the effect of a function on data structures passed to it (which fields are/not updated by the function).  Guide data prefetching for recursive heap data structures

Contributions  Provided a new method for computing read/write sets for connection analysis, introducing the notion of anchor handles.  Demonstrated a variety of applications: extending standard scalar compiler optimizations, array dependence testers and program understanding.  Provided extensive measurements. Demonstrate up to 10% improvement over gcc -O3.

Conclusions  Pointer analysis is an important part of an optimizing C compiler, and one can achieve significant benefits from such an analysis.  Future work will be in three major directions:  Effect of stack and heap read/write sets on fine-grain parallelism and instruction scheduling  Benefits of context-sensitive, flow-sensitive analyses vs. flow-insensitive analyses  Continue to develop new transformations for pointer- intensive programs