rentsch@unc.UUCP (04/16/87)
In article <2577@intelca.UUCP> clif@intelca.UUCP (Clif Purkiser) writes: > While, I agree that using a global optimizing compiler is not exactly > kosher for the dhrystone benchmark it sometimes neccessary. For > instance: the GreenHills C compiler is a globally optimizing compiler > which generates good Dhrystone numbers for many architectures including > the 80386 and 68020. Unfortunately, I can not find a compiler > switch to turn off the global optimizer. This leaves me with two choices: > post the numbers with the cavet that this a global optimizing compiler > or use the results of a medicore compiler like CC. I don't really > think that global optimization is a problem as long as it is clearly > labeled. I see this drawback as saying something not about the suitability of optimizing compilers but about the suitability of the Dhrystone program as a benchmark. If Dhrystone really is prone to sweeping global optimizations (with corresponding improvements in running time), then doesn't it seem to be pretty poor as a benchmark? I heard a similar about a program used to benchmark an early FORTRAN H compiler -- the global optimizer reduced the entire program to just a print statement (after having removed loop invariants, constant folding, and removing the now empty loop). The program took a long time to compiler, but it ran VERY fast. So, here is my vote for benchmark programs which are not prone to huge improvements through global optimization.
eugene@pioneer.arpa (Eugene Miya N.) (04/16/87)
In article <104@unc.unc.UUCP> rentsch@unc.UUCP (Tim Rentsch) writes: > >So, here is my vote for benchmark programs which are not prone to >huge improvements through global optimization. Well, I won't go that ... far. That tends to imply that a user knows a lot about a machine. I've been looking at huge code restructuring lately. GO should be something you are at least aware of and be prepared to compare states. Try to run with optimization turned off for instance. This was something we discussed at a Bay Area ACM/SIGGRAPH performance meeting today. Present both numbers. My concept right now is for virtual MachoIPS and MachoFLOPs versus real MachoFLOPS and MachoIPS with the ratio as a comparison of optimization quality. Note: I might not thing this next week. From the Rock of Ages Home for Retired Hackers: --eugene miya NASA Ames Research Center eugene@ames-aurora.ARPA "You trust the `reply' command with all those different mailers out there?" "Send mail, avoid follow-ups. If enough, I'll summarize." {hplabs,hao,ihnp4,decwrl,allegra,tektronix,menlo70}!ames!aurora!eugene