This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

increase alignment of global structs in increase_alignment pass


Hi Richard,
As discussed in private mail, this version of patch attempts to
increase alignment
of global struct decl if it contains an an array field(s) and array's
offset is a multiple of the alignment of vector type corresponding to
it's scalar type and recursively checks for nested structs.
eg:
static struct
{
  int a, b, c, d;
  int k[4];
  float f[10];
};
k is a candidate array since it's offset is 16 and alignment of
"vector (4) int" is 8.
Similarly for f.

I haven't been able to create a test-case where there are
multiple candidate arrays and vector alignment of arrays are different.
I suppose in this case we will have to increase alignment
of the struct by the max alignment ?
eg:
static struct
{
  <fields>
  T1 k[S1]
  <fields>
  T2 f[S2]
  <fields>
};

if V1 is vector type corresponding to T1, and V2 corresponding vector
type to T2,
offset (k) % align(V1) == 0 and offset (f) % align(V2) == 0
and align (V1) > align(V2) then we will increase alignment of struct
by align(V1).

Testing showed FAIL for g++.dg/torture/pr31863.C due to program timeout.
Initially it appeared to me, it went in infinite loop. However
on second thoughts I think it's probably not an infinite loop, rather
taking (extraordinarily) large amount of time
to compile the test-case with the patch.
The test-case  builds quickly for only 2 instantiations of ClassSpec
(ClassSpec<Key, A001, 1>,
 ClassSpec<Key, A002, 2>)
Building with 22 instantiations (upto ClassSpec<Key, A0023, 22>) takes up
to ~1m to compile.
with:
23  instantiations: ~2m
24 instantiations: ~5m
For 30 instantiations I terminated cc1plus after 13m (by SIGKILL).

I guess it shouldn't go in an infinite loop because:
a) structs cannot have circular references.
b) works for lower number of instantiations
However I have no sound evidence that it cannot be in infinite loop.
I don't understand why a decl node is getting visited more than once
for that test-case.

Using a hash_map to store alignments of decl's so that decl node gets visited
only once prevents the issue.
The patch is cross-tested on aarch64*-*-* and arm*-*-* and passes with
using hash_map
workaround.

Thanks,
Prathamesh

Attachment: patch.diff
Description: Text document

Attachment: ChangeLog
Description: Binary data


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]