[LLVMdev] Checked arithmetic

[Jonathan S. Shapiro]

On Wed, 2008-03-26 at 21:55 +0100, Duncan Sands wrote:
> Hi Chris,
> 
> > > what's the point?  We have this today with apint codegen (if you turn on
> > > LegalizeTypes).  For example, this function
> > 
> > The desired code is something like:
> > 
> > foo:
> >     addl %eax, %ecx
> >     jo overflow_happened
> >     use(%ecx)
> 
> ...is there any chance
> of codegen being capable one day of combining carry arithmetic coming
> from apint and the appropriate conditional branch into a "jo"?

This particular use case pattern (cond branch using CC value generated
as a side effect of a useful operation) appears often enough to be worth
getting right.

So if it's beyond what codegen can do, then codegen wants to be fixed.

It's clear enough that I can get working code emitted today -- that's
not the issue. The concern is that languages supporting non-hardware
arithmetic precisions rely on this particular pattern a whole lot.

I definitely need to dig in harder and understand how LLVM works before
I try to seriously advocate anything on this topic (or any other), but
from a surface understanding of the IR, and experience building several
prior compilers, I would say that the current IR is very well suited to
C and not yet an ideal match to Ada, COBOL, or some other languages.

That's not a criticism. Just a statement of first impressions.


shap