middle-end/5877: negative offset to array index leads to inefficient code

[janis187 at us dot ibm dot com]

>Number:         5877
>Category:       middle-end
>Synopsis:       negative offset to array index leads to inefficient code
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          pessimizes-code
>Submitter-Id:   net
>Arrival-Date:   Thu Mar 07 16:06:01 PST 2002
>Closed-Date:
>Last-Modified:
>Originator:     Janis Johnson
>Release:        gcc version 3.1 20020307 (prerelease)
>Organization:
>Environment:
Itanium, Red Hat 7.1.94 (Roswell)
>Description:
This test case, when compiled on ia64-unknown-linux-gnu with
the current 3.1 branch or the current mainline, with any
level of optimization, splits up the RHS integer value into
individual bytes and then stores them one at a time.

The behavior started with this change, which seems to have
exposed a latent problem:

2002-02-20  Roger Sayle  <roger@eyesopen.com>
            Jakub Jelinek  <jakub@redhat.com>

        PR c/4389
        * tree.c (host_integerp): Ensure that the constant
        integer is representable in a HOST_WIDE_INT or an
        unsigned HOST_WIDE_INT when pos is zero or non-zero
        respectively.  Clarify comment.
        * c-format.c (check_format_info_recurse): Fix
        host_integerp usage; the pos argument should be zero
        when assigning to a signed HOST_WIDE_INT.

The test case has an array offset that is a negative value
but whose tree node says that it is unsigned.  The old
version of post_integerp returned 1 for that value and POS
of 0, but now it returns 0.  The creation of the array
offset and the call to post_integerp all happen within
expand_assignment.  My first attempts to fix this caused
ICEs, so I'll leave it to someone who understands this code
and get back to other things.

This behavior (eight single-byte stores instead of one
eight-byte store) is a regression from 3.0.4.
>How-To-Repeat:
Compile the following test case on ia64 and examine the
generated instructions.  There should be a single 8-byte
store to the array element within the loop rather than a
series of 1-byte stores.

long a[1004];
void
foo ()
{
  int i;
  for (i = 4; i < 1004; i++)
    a[i-4] = i;
}

>Fix:
Unknown.
>Release-Note:
>Audit-Trail:
>Unformatted: