[Bug middle-end/40952] New: version 150336 broke bootstrap on ia64-hp-hpux11.23

[Bug middle-end/40952] New: version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

Starting with version 150336, GCC does not bootstrap on ia64-hp-hpux11.23. In
stage2 gengtype dies with a Memory fault:

Program received signal SIGSEGV, Segmentation fault
  si_code: 1 - SEGV_MAPERR - Address not mapped to object.
0x40107a0:0 in nreverse_pairs (list=0x40014350)

I have updated to 150344 but that did not fix this problem.

Since it happens in stage2, I am afraid the patch may be causing us to generate
bad code on IA64 HP-UX.


-- 
           Summary: version 150336 broke bootstrap on ia64-hp-hpux11.23
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: sje at cup dot hp dot com
 GCC build triplet: ia64-hp-hpux11.23
  GCC host triplet: ia64-hp-hpux11.23
GCC target triplet: ia64-hp-hpux11.23


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #1 from bonzini at gnu dot org  2009-08-03 20:54 -------
Well, stage2 is better than stage3 :-)

Can you post a preprocessed testcase?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #2 from bonzini at gnu dot org  2009-08-04 06:37 -------
I found this pasto, can you check?

Index: gcc/calls.c
===================================================================
--- gcc/calls.c (revision 150359)
+++ gcc/calls.c (working copy)
@@ -2353,9 +2353,9 @@ expand_call (tree exp, rtx target, int i
                                 &caller_unsignedp,
                                 TREE_TYPE (current_function_decl), 1);
       callee_promoted_mode
-       = promote_function_mode (TREE_TYPE (caller_res), callee_mode,
+       = promote_function_mode (TREE_TYPE (funtype), callee_mode,
                                 &callee_unsignedp,
-                                TREE_TYPE (funtype), 1);
+                                funtype, 1);
       if (caller_mode != VOIDmode
          && (caller_promoted_mode != callee_promoted_mode
              || ((caller_mode != caller_promoted_mode


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2009-08-04 06:37:23
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #3 from sje at cup dot hp dot com  2009-08-04 14:59 -------
No, the patch in comment #2 did not fix the problem.  I did a non-bootstrap
build followed by a testsuite run hoping I would get a failure there with a
nice small test case but I didn't get any unexpected failures.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #4 from sje at cup dot hp dot com  2009-08-04 17:09 -------
Created an attachment (id=18300)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18300&action=view)
Testcase

This test case shows the difference in generated code between r150335 and
r150336.  It does not actually generate wrong code, just different (technically
better) code.

In this case on IA64 HP-UX we no longer generate some addp4 instructions to
extend pointer arguments from 32 bits to 64 bits.  In this code the addp4's are
not needed but I am guessing somewhere in gengtype they are not getting
generated and are needed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #5 from sje at cup dot hp dot com  2009-08-04 17:46 -------
Created an attachment (id=18301)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18301&action=view)
Better test case

Here is a better test case that actually shows bad code being generated.  In
r150335, we extend r32 with 'addp4 r32 = 0, r32' and that is fine.  In the new
code we combine the addp4 with a move and generate 'addp4 r8 = r32, r0'  This
is wrong, if we generated 'addp4 - r0, r32' then we would be OK.  addp4 is not
a symetrical operation and the first argument should be an offset (often zero)
and the second argument should be a 32 bit pointer that we want to extend to 64
bits before dereferencing.  addp4 uses the REG_POINTER attribute to distinguish
between offsets and pointers.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #6 from sje at cup dot hp dot com  2009-08-04 21:04 -------
I am currently looking at assign_parm_setup_reg, in the old code with test2.c,
we set promoted_nominal_mode being set to DImode and then we get a conversion
because promoted_nominal_mode != data->promoted_mode.  In the new code
promoted_nominal_mode is set to SImode (the same as data->promoted_node) and so
we don't enter the if statement where the ptr_extend conversion is done in the
old code.  promoted_nominal_mode was set by a call to promote_function_node and
is now being set by a call to promote_mode.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #7 from sje at cup dot hp dot com  2009-08-04 21:23 -------
I think I got things backwards in my last comment.  In the old code
promoted_nominal_mode was set by a call to promote_mode which had the ifdef
POINTERS_EXTEND_UNSIGNED in it to set unsignedp for REFERENCE and POINTER
types.

In the new code we call promote_function_mode and this code doesn't have that
ifdef.

So I think, for HP-UX, I need to define TARGET_PROMOTE_FUNCTION_MODE, assuming
we don't want the ifdef POINTERS_EXTEND_UNSIGNED in promote_function_mode.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #8 from bonzini at gnu dot org  2009-08-04 22:30 -------
Created an attachment (id=18304)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18304&action=view)
plausible patch


The patch adds back the mode of operation corresponding to
promote_nominal_mode's assignment.  It turns out that it's needed in another
place and that it can be defined relatively easily in the documentation.

I'll try bootstrapping it tomorrow.


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |bonzini at gnu dot org
                   |dot org                     |
             Status|NEW                         |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #9 from sje at cup dot hp dot com  2009-08-04 23:48 -------
Well, it got a lot further using the patch but it didn't completely bootstrap. 
The compiler built but while building libstdc++-v3 I got:

/proj/opensrc/nightly/src/trunk/libstdc++-v3/libsupc++/eh_ptr.cc: In function
'std::__exception_ptr::exception_ptr std::current_exception()':
/proj/opensrc/nightly/src/trunk/libstdc++-v3/libsupc++/eh_ptr.cc:176:31:
internal compiler error: in expand_expr_real_1, at expr.c:7427

/proj/opensrc/nightly/src/trunk/libstdc++-v3/src/ios_locale.cc: In member
function 'std::locale std::ios_base::imbue(const std::locale&)':
/proj/opensrc/nightly/src/trunk/libstdc++-v3/src/ios_locale.cc:50:20: internal
compiler error: in expand_expr_real_1, at expr.c:7427

/proj/opensrc/nightly/src/trunk/libstdc++-v3/src/locale.cc: In member function
'std::string std::locale::name() const':
/proj/opensrc/nightly/src/trunk/libstdc++-v3/src/locale.cc:124:12: internal
compiler error: in expand_expr_real_1, at expr.c:7427

and a few more, all dieing at line 7427 of expr.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #10 from bonzini at gnu dot org  2009-08-05 06:09 -------
preprocessed testcase?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #11 from sje at cup dot hp dot com  2009-08-05 21:36 -------
Created an attachment (id=18308)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18308&action=view)
 C++ test case that fails with plausible patch

This is cut down from libstdc++-v3 sources.  I am not sure it is still legal
C++ but it generates the same error that I was getting during the libstdc++-v3
build.

No special options are needed, it will fail with:

x.cc: In member function 'virtual std::string
std::generic_error_category::message(int) const':
x.cc:14:61: internal compiler error: in expand_expr_real_1, at expr.c:7427
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #12 from bonzini at gnu dot org  2009-08-05 23:09 -------
Created an attachment (id=18309)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18309&action=view)
another patch

This failure is related to NRV, which is why there is no C failure.

If I take care of it the patch is actually nicer because there's no distinction
between RESULT_DECL and PARM_DECL.  for_return == 2 now means simply "the
register in/from which a PARM_DECL/RESULT_DECL is copied".

This means that the only code that needs to special case promote_function_mode
is code copying from parameter registers (calls.c) and into the result
(expand_value_return in stmt.c).  I checked the pre-patch code and everybody
else used to guard the call to promote_mode(..., 1) with a call to the
promote_args/promote_return hook; this means that now they can call
promote_function_mode with a final argument that is 0 or 1.

BTW here is a minimal testcase:

struct A
{
  A (const A & __str);
  A (const char * __s);
};
struct B
{
  virtual A a (void) const { return A ("hi"); }
};
const B b;


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #18304|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #13 from sje at cup dot hp dot com  2009-08-06 16:57 -------
I have good news and I have bad news (as the saying goes).  My IA64 bootstrap
completed successfully and the testing also went fine.  But, I also did an
hppa2.0w-hp-hpux11.11 bootstrap and that bootstrap failed with differences in
the results of stage2 and stage3.  When I remove your patch, the bootstrap on
PA is fine.

Comparing stages 2 and 3
warning: gcc/cc1plus-checksum.o differs
warning: gcc/cc1-checksum.o differs
warning: hppa2.0w-hp-hpux11.11/libgcc/lib2funcs.o differs
warning: hppa2.0w-hp-hpux11.11/libgcc/lib2funcs_s.o differs
Bootstrap comparison failure!
gcc/fortran/io.o differs
gcc/fortran/scanner.o differs
gcc/fortran/intrinsic.o differs
make[2]: *** [compare] Error 1

I will see if I can get more info on what the object differences are.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #14 from sje at cup dot hp dot com  2009-08-06 17:44 -------
Created an attachment (id=18312)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18312&action=view)
test case where stage1 and stage2 generate different code

After applying the most recent patch the ia64 bootstrap started working but the
pa bootstrap (hppa2.0w-hp-hpux11.11) started failing.  The stage2/stage3
comparision failed.  I cut down fortran/io.c into test4.c, and when I compile
this test case with the stage1-gcc/cc1 I get unsigned extraction instructions
in the code stream but with stage2-gcc/cc1 I get signed extraction
instructions.  I think the signed extractions are wrong since format_string is
a pointer to an unsigned int, but then again we are putting the value into a
signed char.  If I make the function an 'unsigned char' then the problem goes
away.  Originally the function was just 'char' and it also had the problem


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #15 from bonzini at gnu dot org  2009-08-07 07:24 -------
Subject: Re:  version 150336 broke bootstrap on ia64-hp-hpux11.23


> After applying the most recent patch the ia64 bootstrap started working but the
> pa bootstrap (hppa2.0w-hp-hpux11.11) started failing.  The stage2/stage3
> comparision failed.  I cut down fortran/io.c into test4.c, and when I compile
> this test case with the stage1-gcc/cc1 I get unsigned extraction instructions
> in the code stream but with stage2-gcc/cc1 I get signed extraction
> instructions.  I think the signed extractions are wrong since format_string is
> a pointer to an unsigned int, but then again we are putting the value into a
> signed char.

Ouch, this means stage2 is being miscompiled.  The testcase is very 
useful, but now one needs to understand _where_ the miscompiled code is 
and minimize that testcase.

If you can give me access to the HPUX machine, I can debug it.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #16 from bonzini at gnu dot org  2009-08-08 07:57 -------
Can you try running the testsuite on a non-bootstrapped hppa-hpux compiler? 
However, I have an idea and will provide a patch soon.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #17 from bonzini at gnu dot org  2009-08-08 08:18 -------
Try squashing this in:

Index: pa/pa.c
===================================================================
--- pa/pa.c     (revision 150359)
+++ pa/pa.c     (working copy)
@@ -9199,7 +9199,7 @@ pa_promote_function_mode (const_tree typ
                           const_tree fntype ATTRIBUTE_UNUSED,
                           int for_return)
 {
-  if (for_return != 1)
+  if (for_return == 0)
     return mode;
   return promote_mode (type, mode, punsignedp);
 }


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #18 from sje at cup dot hp dot com  2009-08-10 15:33 -------
I tried the pa.c change, it didn't help.  I will do a non-bootstrap build with
tests.  All my HP-UX machines are behind the HP firewall unfortunately.  I see
we have a PA box in the compile farm but I think it is running linux and not
hpux.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #19 from sje at cup dot hp dot com  2009-08-10 19:32 -------
Here are the unexpected failures I see when running the testsuite using a
non-bootstrap PA compiler.  The compiler I built includes the patch from
comment #17.

FAIL: 22_locale/ctype/narrow/char/2.cc execution test
FAIL: gcc.c-torture/execute/20020108-1.c execution,  -O0
FAIL: gcc.c-torture/execute/20020108-1.c execution,  -O1
FAIL: gcc.c-torture/execute/pr39240.c execution,  -O0
FAIL: gcc.c-torture/execute/pr39240.c execution,  -O1
FAIL: gcc.c-torture/execute/pr39240.c execution,  -O2
FAIL: gcc.c-torture/execute/pr39240.c execution,  -O3 -fomit-frame-pointer
FAIL: gcc.c-torture/execute/pr39240.c execution,  -O3 -g
FAIL: gcc.c-torture/execute/pr39240.c execution,  -Os
FAIL: gfortran.dg/compiler-directive_1.f90  -O   (test for errors, line 42)
FAIL: gfortran.dg/compiler-directive_1.f90  -O  (test for excess errors)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #20 from bonzini at gnu dot org  2009-08-10 20:09 -------
Created an attachment (id=18337)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18337&action=view)
patch fixing hpux differences

This patch squashes in the following too.  It is a stupid used-uninitialized
bug that was not signaled because unsignedp is passed by reference to
promote_function_mode.

Index: gcc/stmt.c
===================================================================
--- gcc/stmt.c  (branch promote-function-mode2)
+++ gcc/stmt.c  (working copy)
@@ -1517,8 +1517,8 @@ expand_value_return (rtx val)
   rtx return_reg = DECL_RTL (decl);
   if (return_reg != val)
     {
-      int unsignedp;
       tree funtype = TREE_TYPE (current_function_decl);
       tree type = TREE_TYPE (decl);
+      int unsignedp = TYPE_UNSIGNED (type);
       enum machine_mode old_mode = DECL_MODE (decl);
       enum machine_mode mode = promote_function_mode (type, old_mode,

I'll also bootstrap the patch overnight on x86_64-pc-linux-gnu, and plan to
rename the last argument of the hook now.


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #18309|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #21 from sje at cup dot hp dot com  2009-08-10 20:18 -------
That seems to have fixed pr39240.c in the testsuite.  I will try a complete
bootstrap again.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #22 from sje at cup dot hp dot com  2009-08-10 21:30 -------
A hppa2.0w-hp-hpux11.11 bootstrap was successful with the v3 patch and running
the broken tests also looks OK now.  I will let my normal nightly process do a
full test run tonight on PA but it looks good.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[sje at cup dot hp dot com]

------- Comment #23 from sje at cup dot hp dot com  2009-08-11 17:18 -------
I don't have complete results because my builds failed due to a bad libstdc++
checkin (that has been fixed) but the C and Fortran parts all looked good.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952

[Bug middle-end/40952] version 150336 broke bootstrap on ia64-hp-hpux11.23

[bonzini at gnu dot org]

------- Comment #24 from bonzini at gnu dot org  2009-08-12 16:29 -------
patch committed as r150700


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40952