On 07/29/2016 09:35 AM, Bin Cheng wrote:
Hi,
This is prerequisite patch for fixing PR34114 which reveals a weakness of GCC
in analyzing niter for loop with NE_EXPR exit condition. For such loops, we
quite often need to check if delta (final - start) of loop range is multiple of
IV's step. This patch proves multiple_of_p (top, bottom) for more cases:
1) Handle (X + 4294967293) as (X - 3) if the expression is of unsigned int
type. At the moment it's not recognized because (4294967293 % 3 != 0)
2) Handle top var if it's defined as:
top = (X & ~(bottom - 1) ; bottom is power of 2
3) Handle top var if it's defined as:
Y = X % bottom
top = X - Y
Bootstrap and test on x86_64 and AArch64 along with next patch. Is it OK?
Thanks,
bin
2016-07-27 Bin Cheng <bin.ch...@arm.com>
* fold-const.c (multiple_of_p): Improve MULT_EXPR, PLUS_EXPR,
PLUS_EXPR case. Handle SSA_NAME case.
OK.
Note that it would seem that this is an ideal kind of behaviorial change
to test in the new unit testing framework. See the code inside the #if
CHECKING_P at the bottom of fold-const.c.
Essentially you can build up the appropriate tree nodes and pass them to
multiple_of_p, then test the result -- without having to find/create C
code that will trigger your code.
I'm not requiring it for this change, but I might start for this kind of
change in the future :-) In the mean time, consider adding the tests,
you may find that they're a great way to verify behavior of these low
level folders and similar routines.
Jeff
