> For optimized code, there are many optimizations that can break
> coverage info. Code motion is one of them. This patch actually tries
> to fix the broken coverage info, as illustrated by the unittest.
No, you seems to be misunderstanding what coverage means. For coverage to be
correct, it is necessary that every insn generated for a particular source
construct be associated with this source construct in the debug info. If you
associate an insn with another source construct and it is executed, then
you'll have the other source construct marked as covered, although it might
not be actually covered during the execution.
Code motion (as any other optimization passes) doesn't break coverage per se.
For example, there is nothing wrong with hoisting a instruction out of a loop
and keeping its source location if it is always executed; coverage info will
be correct after the hoisting. In more complex cases, it might be necessary
to clear the source location of the hoisted instruction.
> If we clear the debug info for instructions moved to other BB, is it
> acceptable?
Yes, clearing is acceptable in principle, but should be done with extreme care
since you drop information, so reorder_insns isn't the appropriate place as
it's too big a hammer.
FWIW, we have related patchlets in our internal tree, like:
* loop-invariant.c (move_invariant_reg): Clear the locator of the
invariant's insn after it has been moved.
* tree-ssa-loop-im.c (move_computations_stmt): Clear the location and
block of the invariant's statement after it has been moved.
As for the more general problem of jumpiness in GDB for highly optimized code,
this should be changed in GDB if your users cannot deal with it. The debug
info describes the relationship between the generated code and the source code
and, at high optimization levels, this relationship is not isomorphic at all.
It's up to the source level debugger to filter out the non-isomorphic part of
the mapping if it deems desirable to do so.
--
Eric Botcazou
