The ranger currently handles "inferred" non-nulls with a special cache..
inferred ranges are thing like
a_2 = *b_1
the dereference of b_1 implies that any use of b_1 after that statement
is non-null (or a trap would have happened). The problem is that most
ranges are calculated as a result of the statement, not as a use. To
deal with that, I created a mini cache for non-nulls.
The first time a request is made for a pointers range, a quick visit to
each use is made to see if that use infers the non-null property, and a
bit is set for the block it occurs in. When calculating ranges, this
bitmap for the name is consulted during range propagation to see if a
range is actually non-null
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97609 is trapping because
an interesting sequence of events is happening which results in a stmt
which is in the middle of being adjusted has a query made in-flight, and
, well, boom. a PR happens.
The routine I was calling infer_value_range() was doing more work than
the cache requires, and some of this extra work was looking closely at
the stmt... and the stmt wasnt valid gimple at the time. It turns out,
the cache doesnt need this extra work, and its actually easier to simply
call the routine we really want directly.
On a side note, going into the next stage 1, I plan to replace this
approach with a more generalize solution that has been briefly mentioned.
We can add another API entry point to range-ops and the ranger to deal
with side effects. currently we can ask for
1) the range ON (before) a stmt is executed (range_of_expr)
2) the range ON an edge
3) the range OF a stmt (the result of the stmt)
I want to add a more generalized version of 3 which would be range AFTER
stmt which works for things other than the result of a stmt...
.4) range_after_stmt which would include things like the dereference:
a_2 = *b_1 // range after would provide b_1 = [1, +INF]
d_4 = f_4 / g_8 // range after would provide g_8 = ~[0,0]
Regardless... thats months away. for now this patch avoids calling
those things that dont need to be called and bypasses the erroneous
in-flight analysis of the stmt.
Bootstrapped on x86_64-pc-linux-gnu, no regressions, pushed.
Andrew
