fix ICE out of SRA try_instantiate_multiple_fields

[Olivier Hainque <hainque at adacore dot com>]

Hello

Compiling the Ada testcase below on mips-irix/n32 ICEs out of an invalid
operation produced by the SRA try_instantiate_multiple_fields circuitry. The
case is extracted from the a-except runtime library unit, so the ICE presented
here actually breaks bootstrap for Ada on the target.

    package P is
       function Letter return String;
    end;

    package body P is
       function X return String is
       begin
          return "X";
       end;

       function Letter return String is
       begin
          return X;
       end;
    end;

The compiler eventually complains about an invalid BIT_FIELD_REF conversion
operand, like so:

    $ ../prev-gcc/gnat1 -gnatp -O1 [...] p.adb
    ...
    p.adb:14: error: invalid operand to unary operator
    BIT_FIELD_REF <SR.2_4, 32, 32>;
    p.adb:14: note: in statement
    SR.5_6 = (unsigned int) BIT_FIELD_REF <SR.2_4, 32, 32>;

Indeed, post-gimplification NOP_EXPRs are expected to apply to "is_gimple_val"
operands only, and SRA turns

    P.Letter ()
    {
    ...
    <bb 2>:
      D.1314 ={v} p__x ();
      D.1313 ={v} D.1314;
      R5b ={v} D.1313;
      D.1315 ={v} R5b;
      return D.1315;

    }

into

    P.Letter ()
    {
    ...
    <bb 2>:
      SR.1_1 = 0;
      SR.2_2 = 0;
      D.1314 ={v} p__x ();
      SR.3_3 = VIEW_CONVERT_EXPR<UNSIGNED_64>(D.1314);
      SR.2_4 = SR.3_3;
      SR.4_5 = SR.1_1 & 0x0ffffffff00000000;
      SR.5_6 = (unsigned int) BIT_FIELD_REF <SR.2_4, 32, 32>;
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
      SR.6_7 = (UNSIGNED_64) SR.5_6;
      SR.1_8 = SR.4_5 | SR.6_7;
      SR.7_9 = SR.1_8 & 4294967295;
      SR.8_10 = (unsigned int) BIT_FIELD_REF <SR.2_4, 32, 0>;
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
      SR.9_11 = (UNSIGNED_64) SR.8_10;
      SR.10_12 = SR.9_11 << 32;
      SR.1_13 = SR.7_9 | SR.10_12;
      SR.1_14 = SR.2_4;
      SR.11_15 = SR.1_14;
      D.1315 ={v} VIEW_CONVERT_EXPR<struct >(SR.11_15);
      return D.1315;

The transformed assignments initially involved Ada "fat" pointers to
unconstrained objects (strings), represented as a record with two regular
pointer fields, one to the bounds and one to the data.

IIUC, try_instantiate_multiple_fields is meant to simplify the code generated
for accesses to groups of bitfields. Pointer fields are usually not good
candidates for such coalescing, and the unexpected code expansion above I
think illustrates this pretty well.

Pointer fields are most of the time not coalesced by virtue of their size
typically being equal to that of a word, which is not the case on mips-irix
with the n32 ABI where the crash showed up.

The attached simple patch is a suggestion to address that by early returning
from try_instantiate_multiple_fields when the candidate field has
POINTER_TYPE.

Tested first on mips-irix/n32 by checking that the testcase doesn't ICE
anymore and that bootstrap proceeds past the previous failure point, then with
a regular bootstrap and regression test cycle on x86_64-suse-linux.

Thanks in advance,

Olivier

2008-05-06  Olivier Hainque  <hainque@adacore.com>

	* tree-sra.c (try_instantiate_multiple_fields): Early return
	if field has POINTER_TYPE.

	testsuite/
	* gnat.dg/fatp_sra.adb: New test.

Attachment: fatp-sra.dif
Description: Text document