EECS 583 – Homework 2
Winter 2004
Assigned: Monday, February 16, 2004
Due: Sunday, March 7, 2004
This homework focuses on extending the Elcor optimization system to do 2 optimizations
1. Local common subexpression elimination (CSE)
2. Register renaming
#1: Local CSE
Your first mission is to implement the optimization local common subexpression
elimination. The top-level function should take an entire procedure and walk over each
basic block or hyperblock invoking local CSE on the block. The local CSE algorithm
should identify and eliminate all redundant expressions within a block. Elimination is
accomplished by creating 2 move operations. So, it should perform the following
transformation:
r1 = r2 + r3 if T r1 = r2 + r3 if T
… r100 = r1 if T
…
r4 = r2 + r3 if T r4 = r100 if T
Subsequent copy propagation and dead code elimination will remove the copies (you
don’t need to implement these). Note that you are implementing a local algorithm only,
so any inter-block redundant expressions are not eliminated, only intra-block
redundancies. Your implementation should only handle arithmetic instructions (see
opcode_properties.h for useful functions to identify the appropriate opcodes). Branches,
pbrs, loads, stores, cmpps, and pseudo operations should be excluded from optimization.
You should use reaching definitions (DU/UD chains) to accomplish the optimization. A
sample piece of code will be provided to invoke reaching definition analysis and access
the information. Your optimization should work on either a basic block or hyperblock.
Thus, your optimization should be predicate cognizant. This is actually easier than it
sounds, as most Elcor analyses (ie liveness, reaching defs) are already predicate
cognizant, so this is not that difficult, but does require a bit of extra work.
Elcor already has a version of global CSE, see Opti/el_opti_common_subexpr_elim.cpp.
You are welcome to look at this and steal any code from it that you may find useful.
However, this implementation is rather difficult to understand and it uses available
expression analysis. Your implementation (which will use reaching defs) will be quite
different, so in reality the amount of useful code you can get from here will be limited
and it may be easier to ignore this code to avoid unnecessary confusion.
