This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

flow problem avoided, left with global-alloc


After thinking about Richard Henderson's comments about flow, and talking
with Feeley, I set a flag to have only one computed goto per
procedure, instead of distributing the computed goto's throughout
the code.  (If you want a computed goto, you jump to the "computed
goto basic block".) That cut the compile time and memory requirements
appreciably.  Here are the new times for "cc1 -O1 -fPIC -mcpu=supersparc"
with gcc-2.96 19991005:

time in flow: 1.000000 (1%)
time in global-alloc: 48.390000 (71%)
time in flow2: 0.770000 (1%)

By the way, I'm wondering if the flow analysis considers as possible
targets of computed goto's only those labels whose addresses are
taken, rather than all labels.

Brad Lucier


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]