Path: meolyon!mwhh!news.Hanse.DE!news.rrz.uni-hamburg.de!news.dkrz.de!news.tu-harburg.de!news.dfn.de!gina.zfn.uni-bremen.de!marvin.pc-labor.uni-bremen.de!news.uni-stuttgart.de!news.belwue.de!fu-berlin.de!news.mathworks.com!newsfeed.internetmci.com!news.sprintlink.net!cs.utexas.edu!math.ohio-state.edu!cis.ohio-state.edu!ida.liu.se!mpe
From: mpe@ida.liu.se (Mikael Pettersson)
Newsgroups: gnu.gcc.bug
Subject: gcc 2.7.0 bug PowerPC
Date: 10 Aug 1995 18:27:43 -0400
Organization: GNUs Not Usenet
Lines: 132
Sender: daemon@cis.ohio-state.edu
Approved: bug-gcc@prep.ai.mit.edu
Distribution: gnu
Message-ID: <9508102226.AA11314@sen14.ida.liu.se>
GCC 2.7.0, configured for "powerpc-parsytec-parix", generates
incorrect code for the BUG function below. This happens both
with and without -O.
The problem is that the first word of array "tmp118" has
been given the same address in the activation record as
the last word of array "tmp117".
/* bug.c */
#include <stdio.h>
extern void gc(void*), print(void*);
char *sp_low, *sp_high;
void BUG(void)
{
{
char mySP;
if( &mySP < sp_low ) {
void *tmp119[5];
gc(tmp119);
return;
}
}
{
void *tmp115[2];
tmp115[0] = (void*)256;
tmp115[1] = (void*)64;
{
void *tmp116[2];
tmp116[0] = (void*)257;
tmp116[1] = tmp115;
{
void *tmp117[3];
tmp117[0] = (void*)515;
tmp117[1] = (void*)2;
tmp117[2] = tmp116;
{
void *tmp118[2];
tmp118[0] = (void*)260; /* overwrites tmp117[2] */
tmp118[1] = tmp117;
print(tmp118);
}
}
}
}
}
void gc(void *ignore) {}
void print(void *x)
{
if( (char*)x < sp_low || (char*)x > sp_high ) {
printf("%u", x);
} else {
unsigned hdr = *(unsigned*)x;
unsigned slots;
void **xx;
printf("%u(", hdr & 63);
slots = hdr >> 8;
xx = (void**)x + 1;
if( slots > 0 )
while( print(*xx++), --slots > 0 )
printf(", ");
printf(")");
}
}
int main(void)
{
char mySP;
sp_high = &mySP;
sp_low = &mySP - 4096;
BUG();
printf("\n");
return 0;
}
The program outputs:
4(3(2, 260))
while it _should_ output:
4(3(2, 1(0(64))))
This is what gcc -O -S generates for function BUG:
.align 2
.globl BUG
.globl .BUG
.csect BUG[DS]
BUG:
.long .BUG, TOC[tc0], 0
.csect .text[PR]
.BUG:
mflr 0
stw 0,8(1)
stwu 1,-96(1)
lwz 9,LC..4(2)
lwz 9,0(9)
addi 0,1,56
cmplw 1,0,9
bc 4,4,L..14
addi 3,1,64
bl .gc
cror 31,31,31
b L..13
L..14:
li 0,256
stw 0,64(1) /* tmp115[0] = 256 */
li 0,64
stw 0,68(1) /* tmp115[1] = 64 */
li 0,257
stw 0,72(1) /* tmp116[0] = 257 */
addi 0,1,64
stw 0,76(1) /* tmp116[1] = &tmp115 */
li 0,515
stw 0,80(1) /* tmp117[0] = 515 */
li 0,2
stw 0,84(1) /* tmp117[1] = 2 */
li 0,260 /* no asignment to tmp117[2] ?? */
stw 0,88(1) /* tmp118[0] = 260 OVERWRITES tmp117[2] */
addi 0,1,80
stw 0,92(1) /* tmp118[1] = &tmp117 */
addi 3,1,88
bl .print /* print(&tmp118) */
cror 31,31,31
L..13:
addi 1,1,96
lwz 0,8(1)
mtlr 0
blr
Path: meolyon!mwhh!news.Hanse.DE!news.rrz.uni-hamburg.de!news.dkrz.de!news.tu-harburg.de!news.dfn.de!Germany.EU.net!howland.reston.ans.net!math.ohio-state.edu!magnus.acs.ohio-state.edu!cis.ohio-state.edu!meolyon.hanse.DE!amylaar
From: amylaar@meolyon.hanse.DE (Joern Rennecke)
Newsgroups: gnu.gcc.bug
Subject: Re: gcc 2.7.0 bug HP-PA (bug fix)
Date: 9 Sep 1995 15:37:05 -0400
Organization: GNUs Not Usenet
Lines: 343
Sender: daemon@cis.ohio-state.edu
Approved: bug-gcc@prep.ai.mit.edu
Distribution: gnu
Message-ID: <m0srTmG-000DkHC@meolyon>
~Newsgroups: gnu.gcc.bug
Path: amylaar
~From: amylaar@meolyon.hanse.de (Joern Rennecke)
~Subject: Re: gcc 2.7.0 bug HP-PA (bug fix)
Distribution: gnu
Organization: Private site running Linux
Message-ID: <1995Sep9.173310.4622@meolyon.hanse.de>
~References: <9508101742.AA11105@sen14.ida.liu.se>
~Date: Sat, 9 Sep 1995 17:33:10 GMT
mpe@ida.liu.se (Mikael Pettersson) writes:
>GCC 2.7.0, configured with --build=hppa1.1-hp-hpux9.05
>and --with-gnu-as (from gnu binutils 2.5.2), miscompiles
>the following program (when not optimizing, -O0):
>/* bug.c */
>#include <stdio.h>
>void rml_gc(void) {}
>char *rmlSPLIM;
>static void sclam826(unsigned *tmp1509)
>{
> char mySP;
> if( &mySP > rmlSPLIM ) {
> void *tmp1527[3];
> rml_gc();
> }
> {
> unsigned tmp796 = *tmp1509;
> unsigned tmp1510[5];
> printf("0x%x\n", tmp796);
> tmp1510[4] = 0x55555555;
> printf("0x%x\n", tmp796); /* tmp796 is wrong here */
> }
>}
>int main(void)
>{
> unsigned tmp1509;
> rmlSPLIM = (char*)&tmp1509 + 4096;
> tmp1509 = 0xaaaaaaaa;
> sclam826(&tmp1509);
> return 0;
>}
>The (incorrect) output of the program is:
>0xaaaaaaaa
>0x55555555
>The assembly code for sclam826() is:
> .align 4
> .PARAM sclam826,ARGW0=GR
>sclam826
> .PROC
> .CALLINFO FRAME=192,CALLS,SAVE_RP,SAVE_SP,ENTRY_GR=3
> .ENTRY
> stw %r2,-20(0,%r30)
> copy %r3,%r1
> copy %r30,%r3
> stwm %r1,192(0,%r30)
> stw %r26,-36(0,%r3)
> addil LR'rmlSPLIM-$global$,%r27
> ldo RR'rmlSPLIM-$global$(%r1),%r19
> ldw 0(0,%r19),%r20
> ldo 8(%r3),%r19
> comb,>>=,n %r20,%r19,L$0003
> .CALL
> bl rml_gc,%r2
> nop
>L$0003
> ldw -36(0,%r3),%r19
> ldw 0(0,%r19),%r20
> stw %r20,32(0,%r3) /* ok */
> ldil LR'L$C0000,%r19
> ldo RR'L$C0000(%r19),%r26
> ldw 32(0,%r3),%r25 /* ok */
> .CALL ARGW0=GR,ARGW1=GR
> bl printf,%r2
> nop
> ldil L'1431655765,%r20
> ldo R'1431655765(%r20),%r19
> stw %r19,32(0,%r3) /* BUG HERE */
> ldil LR'L$C0000,%r19
> ldo RR'L$C0000(%r19),%r26
> ldw 32(0,%r3),%r25 /* wrong value now */
> .CALL ARGW0=GR,ARGW1=GR
> bl printf,%r2
> nop
>L$0002
> ldw -20(0,%r3),%r2
> ldo 64(%r3),%r30
> ldwm -64(0,%r30),%r3
> bv,n 0(%r2)
> .EXIT
> .PROCEND
>As one can see from the lines marked /* ok */ and /* BUG HERE */,
>GCC has placed "tmp796" and "tmp1510[4]" in the same location
>in the activation record.
The underlying problem seems to be present for all targets with
FRAME_GROWS_DOWNWARD un#defined. Fix appended.
Joern Rennecke
*** gcc-2.7.0/function.c Sat Jul 8 01:56:26 1995
--- gcc-2.7.0-x/function.c Sat Sep 9 19:26:59 1995
*************** static tree blocks_nreverse PROTO((tree)
*** 439,444 ****
--- 439,445 ----
static int all_blocks PROTO((tree, tree *));
static int *record_insns PROTO((rtx));
static int contains PROTO((rtx, int *));
+ static rtx assign_stack_local2 PROTO((enum machine_mode, int, int, int *));
/* Pointer to chain of `struct function' for containing functions. */
struct function *outer_function_chain;
*************** get_frame_size ()
*** 638,643 ****
--- 639,646 ----
#endif
}
+ int dummy_slotsize;
+
/* Allocate a stack slot of SIZE bytes and return a MEM rtx for it
with machine mode MODE.
*************** assign_stack_local (mode, size, align)
*** 654,662 ****
--- 657,676 ----
int size;
int align;
{
+ return assign_stack_local2(mode, size, align, &dummy_slotsize);
+ }
+
+ rtx
+ assign_stack_local2 (mode, size, align, sizep)
+ enum machine_mode mode;
+ int size;
+ int align;
+ int *sizep;
+ {
register rtx x, addr;
int bigend_correction = 0;
int alignment;
+ int frame_offset_old;
if (align == 0)
{
*************** assign_stack_local (mode, size, align)
*** 672,677 ****
--- 686,697 ----
else
alignment = align / BITS_PER_UNIT;
+ /* All space allocated on the frame has to be accounted for in the
+ allocated slots sizes, combine_temp_slots won't think that adjacent
+ slots really are adjacent. OTOH, having stack slots start before their
+ address calls for trouble, thus the alignment padding is added to the
+ size of the slot the padding is aligned with. */
+ frame_offset_old = frame_offset;
/* Round frame offset to that alignment.
We must be careful here, since FRAME_OFFSET might be negative and
division with a negative dividend isn't as well defined as we might
*************** assign_stack_local (mode, size, align)
*** 681,686 ****
--- 701,714 ----
frame_offset = FLOOR_ROUND (frame_offset, alignment);
#else
frame_offset = CEIL_ROUND (frame_offset, alignment);
+ {
+ static int *last_sizep = &dummy_slotsize;
+ /* when reload rounds the stacksize, last_sizep gets reset to
+ &dummy_slotsize */
+ *last_sizep += frame_offset - frame_offset_old;
+ last_sizep = sizep;
+ *sizep = size;
+ }
#endif
/* On a big-endian machine, if we are allocating more space than we will use,
*************** assign_stack_local (mode, size, align)
*** 690,695 ****
--- 718,724 ----
#ifdef FRAME_GROWS_DOWNWARD
frame_offset -= size;
+ *sizep = frame_offset_old - frame_offset;
#endif
/* If we have already instantiated virtual registers, return the actual
*************** assign_stack_temp (mode, size, keep)
*** 832,837 ****
--- 861,883 ----
if (best_p->size - rounded_size >= alignment)
{
p = (struct temp_slot *) oballoc (sizeof (struct temp_slot));
+ /* make sure that a size increase from assign_stack_local2
+ to p->size will end up in the appropriate block. Also
+ keeps temp_slots ordered so that combine_temp_slots might
+ be simplified. */
+ #ifdef FRAME_GROWS_DOWNWARD
+ p->next = best_p->next;
+ best_p->next = p;
+ #else
+ {
+ struct temp_slot *q;
+ *p = *best_p;
+ best_p->next = p;
+ q = p;
+ p = best_p;
+ best_p = q;
+ }
+ #endif
p->in_use = p->addr_taken = 0;
p->size = best_p->size - rounded_size;
p->slot = gen_rtx (MEM, BLKmode,
*************** assign_stack_temp (mode, size, keep)
*** 839,846 ****
rounded_size));
p->address = 0;
p->rtl_expr = 0;
- p->next = temp_slots;
- temp_slots = p;
stack_slot_list = gen_rtx (EXPR_LIST, VOIDmode, p->slot,
stack_slot_list);
--- 885,890 ----
*************** assign_stack_temp (mode, size, keep)
*** 855,875 ****
/* If we still didn't find one, make a new temporary. */
if (p == 0)
{
- int frame_offset_old = frame_offset;
p = (struct temp_slot *) oballoc (sizeof (struct temp_slot));
/* If the temp slot mode doesn't indicate the alignment,
use the largest possible, so no one will be disappointed. */
! p->slot = assign_stack_local (mode, size, mode == BLKmode ? -1 : 0);
! /* The following slot size computation is necessary because we don't
! know the actual size of the temporary slot until assign_stack_local
! has performed all the frame alignment and size rounding for the
! requested temporary. Otherwise combine_temp_slots won't think that
! adjacent slots really are adjacent. */
! #ifdef FRAME_GROWS_DOWNWARD
! p->size = frame_offset_old - frame_offset;
! #else
! p->size = frame_offset - frame_offset_old;
! #endif
p->address = 0;
p->next = temp_slots;
temp_slots = p;
--- 899,909 ----
/* If we still didn't find one, make a new temporary. */
if (p == 0)
{
p = (struct temp_slot *) oballoc (sizeof (struct temp_slot));
/* If the temp slot mode doesn't indicate the alignment,
use the largest possible, so no one will be disappointed. */
! p->slot = assign_stack_local2 (mode, size, mode == BLKmode ? -1 : 0,
! &p->size);
p->address = 0;
p->next = temp_slots;
temp_slots = p;
*************** combine_temp_slots ()
*** 927,933 ****
--- 961,983 ----
{
/* P comes after Q; combine P into Q. */
q->size += p->size;
+ #ifdef STACK_GROWS_DOWNWARD
delete_p = 1;
+ #else
+ /* If we'd miss a size increase from assign_stack_local2
+ because it goes into the discarded slot instead of the
+ combined one, subsequent calls would be unable to
+ combine it with a following slot. */
+ {
+ struct temp_slot *tmp = p->next;
+ *p = *q;
+ q->next = tmp;
+ tmp = p;
+ p = q;
+ q = tmp;
+ delete_q = 1;
+ }
+ #endif
break;
}
}
*** gcc-2.7.0/ChangeLog Sat Jul 8 01:54:05 1995
--- gcc-2.7.0-x/ChangeLog Sat Sep 9 19:15:42 1995
***************
*** 1,3 ****
--- 1,42 ----
+ Sat Sep 9 17:54:02 MET DST 1995 J"orn Rennecke (amylaar@meolyon.hanse.de)
+ * function.c (assign_stack_local2): new function. Like
+ assign_local_stack, but takes an additional parameter to store
+ the size of an allocated slot.
+ When FRAME_GROWS_DOWNWARD is not #defined, add the stack alignment
+ to the size of the last allocated slot.
+ (assign_stack_local): redefined as a call to assign_stack_local2.
+ assign_stack_temp(): use assign_stack_local2. Take special care
+ when splitting a block where last_sizep will point to.
+ (combine_temp_slots): When combining for !STACK_GROWS_DOWNWARD,
+ make sure the size increase gets into the combined slot.
+
+ Tue Aug 29 19:10:53 1995 Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
+ * c-common.c (decl_attribute, case A_PACKED): Check is_type first.
+ (decl_attribute, case A_T_UNION): Likewise.
+ Don't access TYPE_FIELDS if DECL is zero.
+ * c-decl.c (finish_struct): If transparent_union attribute
+ specified, validate it once we finish laying the union out.
+
+ Thu Aug 17 21:02:09 1995 J"orn Rennecke (amylaar@meolyon.hanse.de)
+ * i386.c (arithmetic_comparison_operator): New function.
+ (print_operand) : take into account that overflow flag is not
+ set like after a compare instruction.
+ * i386.md (decrement_and_branch_until_zero): use
+ arithmetic_comparison_operator to decide if there is an comparison
+ that is suitable to be expressed by the condition code from an
+ arithmetic operation.
+
+ Fri Jul 14 07:34:21 1995 J"orn Rennecke (amylaar@meolyon.hanse.de)
+ * c-decl.c(finish_struct): wait with decl_attributes() till
+ TYPE_FIELDS (t) is valid.
+
+ Thu Jul 13 22:53:04 1995 J"orn Rennecke (amylaar@meolyon.hanse.de)
+ * c-common.c(decl_attributes): bug fix for A_ALIGNED
+
+ Sat Jul 8 10:33:59 1995 J"orn Rennecke (amylaar@meolyon.hanse.de)
+ * New optimization option -fomit-default-branch . Most of it
+ is in expand_end_case(). Doc in gcc.1
+
Fri Jun 16 06:54:03 1995 Richard Kenner (kenner@vlsi1.ultra.nyu.edu)
* alpha.c (alpha_builtin_saveregs): Use ptr_mode and conversions
