Walter Bright wrote:
grauzone wrote:
bearophile wrote:
Don:
Actually Walter loves goto, so DMD copes really well with it.
When possible it's better to use structured programming and avoid
gotos. That construct can avoid gotos in some common situations.
And regarding the compiler back-end, I think it's also better to
start thinking what's good for LDC :-)
I don't know how relevant this is, but: LLVM uses SSA for registers,
and it seems to be simpler to convert code to SSA if there are no gotos:
"We show that it is possible to generate SSA form in a single pass
(even during parsing) if the program contains only structured control
flow (i.e., no gotos). For such programs the dominator tree can be
built on the fly, too."
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.45.4503
I don't know what they're talking about. The D optimizer has no trouble
whatsoever building dominator graphs, no matter what rat's nest of gotos
are used. This technology was well understood in 1980.
In fact, the front end deconstructs *everything* into gotos, and that's
all the optimizer works with.
The challenge the paper addresses is constructing the SSA form in one
shot. There are indeed algorithms that build the SSA in several passes.
My vague recollection from a compiler construction class is that there
are a number of static analyses that need to do a fixed-point
computation on loops. Such analyses can deal with any jumps except a
particular one: jump forward from the outside of a loop in the middle of
it. Does anyone know what analyses were those?
Andrei
