Patchwork [5/8] Tweak bitfield alignment handling

login
register
mail settings
Submitter Richard Sandiford
Date Nov. 18, 2012, 5:36 p.m.
Message ID <87ip92x17v.fsf@talisman.default>
Download mbox | patch
Permalink /patch/199905/
State New
Headers show

Comments

Richard Sandiford - Nov. 18, 2012, 5:36 p.m.
Eric Botcazou <ebotcazou@adacore.com> writes:
>> This patch replaces:
>> 
>>       /* Stop if the mode is wider than the alignment of the containing
>> 	 object.
>> 
>> 	 It is tempting to omit the following line unless STRICT_ALIGNMENT
>> 	 is true.  But that is incorrect, since if the bitfield uses part
>> 	 of 3 bytes and we use a 4-byte mode, we could get a spurious segv
>> 	 if the extra 4th byte is past the end of memory.
>> 	 (Though at least one Unix compiler ignores this problem:
>> 	 that on the Sequent 386 machine.  */
>>       if (unit > align_)
>> 	break;
>> 
>> with two checks: one for whether the final byte of the mode is known
>> to be mapped, and one for whether the bitfield is sufficiently aligned.
>> Later patches in the series rely on this in order not to pessimise
>> memory handling.
>> 
>> However, as described in the patch, I found that extending this
>> behaviour to get_best_mode affected code generation for x86_64-linux-gnu
>> and powerpc64-linux-gnu, and led to a failure in bb-slp-5.c on both.
>> I therefore chickened out and added the original check back to
>> get_best_mode.
>> 
>> I'm certainly interested in any feedback on the comment, but at the
>> same time I'd like this series to be a no-op on targets that keep
>> the traditional .md patterns.  Any change to get_best_mode is probably
>> best done as a separate patch.
>
> I agree with your conservative approach here.
>
>> gcc/
>> 	* stor-layout.c (bit_field_mode_iterator::bit_field_mode_iterator):
>> 	Set up a default value of bitregion_end_.
>> 	(bit_field_mode_iterator::next_mode): Always apply bitregion_end_
>> 	check.  Include SLOW_UNALIGNED_ACCESS in the alignment check.
>> 	(get_best_mode): Ignore modes that are wider than the alignment.
>
> Fine with me, modulo:
>
>> @@ -2754,6 +2753,62 @@ get_best_mode (int bitsize, int bitpos,
>>    enum machine_mode widest_mode = VOIDmode;
>>    enum machine_mode mode;
>>    while (iter.next_mode (&mode)
>> +	 /* ??? For compatiblity, reject modes that are wider than the
>> +	    alignment.  This has both advantages and disadvantages.
>
> "For compatibility" is a bit vague (compatibility with what?).  I'd write "For 
> historical reasons" or something along these lines.

Yeah, that's better.  Here's what I committed after updating to the patch
for the inclusive bitregion_end_.

Richard


gcc/
	* stor-layout.c (bit_field_mode_iterator::bit_field_mode_iterator):
	Set up a default value of bitregion_end_.
	(bit_field_mode_iterator::next_mode): Always apply bitregion_end_
	check.  Include SLOW_UNALIGNED_ACCESS in the alignment check.
	(get_best_mode): Ignore modes that are wider than the alignment.
John David Anglin - Nov. 20, 2012, 2:57 a.m.
On Sun, 18 Nov 2012, Richard Sandiford wrote:

>        HOST_WIDE_INT start = bitpos_ - (bitpos_ % unit);
>        if (bitregion_start_ && start < bitregion_start_)
>  	break;
> -      if (bitregion_end_ && start + unit > bitregion_end_ + 1)
> +      if (start + unit > bitregion_end_ + 1)

This causes:

/home/dave/gnu/gcc/objdir/./prev-gcc/g++ -B/home/dave/gnu/gcc/objdir/./prev-gcc/ -B/home/dave/opt/gnu/gcc/gcc-4.8.0/hppa-linux-gnu/bin/ -nostdinc++ -B/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/src/.libs -B/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/libsupc++/.libs -I/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/include/hppa-linux-gnu -I/home/dave/gnu/gcc
/objdir/prev-hppa-linux-gnu/libstdc++-v3/include -I/home/dave/gnu/gcc/gcc/libstdc++-v3/libsupc++ -L/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/sr
c/.libs -L/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/libsupc++/.libs -c   -g -O2 -DIN_GCC   -fno-exceptions -fno-rtti -fasynchronous-unwind-tabl
es -W -Wall -Wno-narrowing -Wwrite-strings -Wcast-qual -Wmissing-format-attribut
e -pedantic -Wno-long-long -Wno-variadic-macros -Wno-overlength-strings -Werror 
-fno-common  -DHAVE_CONFIG_H -I. -I. -I../../gcc/gcc -I../../gcc/gcc/. -I../../g
cc/gcc/../include -I../../gcc/gcc/../libcpp/include  -I../../gcc/gcc/../libdecnu
mber -I../../gcc/gcc/../libdecnumber/dpd -I../libdecnumber -I../../gcc/gcc/../li
bbacktrace    ../../gcc/gcc/stor-layout.c -o stor-layout.o../../gcc/gcc/stor-layout.c: In member function 〘bool bit_field_mode_iterator::n
ext_mode(machine_mode*)〙:
../../gcc/gcc/stor-layout.c:2690:43: error: comparison between signed and unsign
ed integer expressions [-Werror=sign-compare]
       if (start + unit > bitregion_end_ + 1)
					   ^
cc1plus: all warnings being treated as errors

Dave
Mikael Pettersson - Nov. 20, 2012, 8:18 a.m.
John David Anglin writes:
 > On Sun, 18 Nov 2012, Richard Sandiford wrote:
 > 
 > >        HOST_WIDE_INT start = bitpos_ - (bitpos_ % unit);
 > >        if (bitregion_start_ && start < bitregion_start_)
 > >  	break;
 > > -      if (bitregion_end_ && start + unit > bitregion_end_ + 1)
 > > +      if (start + unit > bitregion_end_ + 1)
 > 
 > This causes:
 > 
 > /home/dave/gnu/gcc/objdir/./prev-gcc/g++ -B/home/dave/gnu/gcc/objdir/./prev-gcc/ -B/home/dave/opt/gnu/gcc/gcc-4.8.0/hppa-linux-gnu/bin/ -nostdinc++ -B/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/src/.libs -B/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/libsupc++/.libs -I/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/include/hppa-linux-gnu -I/home/dave/gnu/gcc
 > /objdir/prev-hppa-linux-gnu/libstdc++-v3/include -I/home/dave/gnu/gcc/gcc/libstdc++-v3/libsupc++ -L/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/sr
 > c/.libs -L/home/dave/gnu/gcc/objdir/prev-hppa-linux-gnu/libstdc++-v3/libsupc++/.libs -c   -g -O2 -DIN_GCC   -fno-exceptions -fno-rtti -fasynchronous-unwind-tabl
 > es -W -Wall -Wno-narrowing -Wwrite-strings -Wcast-qual -Wmissing-format-attribut
 > e -pedantic -Wno-long-long -Wno-variadic-macros -Wno-overlength-strings -Werror 
 > -fno-common  -DHAVE_CONFIG_H -I. -I. -I../../gcc/gcc -I../../gcc/gcc/. -I../../g
 > cc/gcc/../include -I../../gcc/gcc/../libcpp/include  -I../../gcc/gcc/../libdecnu
 > mber -I../../gcc/gcc/../libdecnumber/dpd -I../libdecnumber -I../../gcc/gcc/../li
 > bbacktrace    ../../gcc/gcc/stor-layout.c -o stor-layout.o../../gcc/gcc/stor-layout.c: In member function 〘bool bit_field_mode_iterator::n
 > ext_mode(machine_mode*)〙:
 > ../../gcc/gcc/stor-layout.c:2690:43: error: comparison between signed and unsign
 > ed integer expressions [-Werror=sign-compare]
 >        if (start + unit > bitregion_end_ + 1)
 > 					   ^
 > cc1plus: all warnings being treated as errors

This error also breaks m68k-linux bootstrap.

HWI32 issue?

Patch

Index: gcc/stor-layout.c
===================================================================
--- gcc/stor-layout.c	2012-11-18 17:26:31.159273478 +0000
+++ gcc/stor-layout.c	2012-11-18 17:28:17.481177253 +0000
@@ -2646,6 +2646,13 @@  fixup_unsigned_type (tree type)
   bitregion_end_ (bitregion_end), align_ (MIN (align, BIGGEST_ALIGNMENT)),
   volatilep_ (volatilep), count_ (0)
 {
+  if (!bitregion_end_)
+    {
+      /* We can assume that any aligned chunk of ALIGN_ bits that overlaps
+	 the bitfield is mapped and won't trap.  */
+      bitregion_end_ = bitpos + bitsize + align_ - 1;
+      bitregion_end_ -= bitregion_end_ % align_ + 1;
+    }
 }
 
 /* Calls to this function return successively larger modes that can be used
@@ -2676,23 +2683,15 @@  bit_field_mode_iterator::next_mode (enum
       if (count_ > 0 && unit > BITS_PER_WORD)
 	break;
 
-      /* Stop if the mode is wider than the alignment of the containing
-	 object.
-
-	 It is tempting to omit the following line unless STRICT_ALIGNMENT
-	 is true.  But that is incorrect, since if the bitfield uses part
-	 of 3 bytes and we use a 4-byte mode, we could get a spurious segv
-	 if the extra 4th byte is past the end of memory.
-	 (Though at least one Unix compiler ignores this problem:
-	 that on the Sequent 386 machine.  */
-      if (unit > align_)
-	break;
-
       /* Stop if the mode goes outside the bitregion.  */
       HOST_WIDE_INT start = bitpos_ - (bitpos_ % unit);
       if (bitregion_start_ && start < bitregion_start_)
 	break;
-      if (bitregion_end_ && start + unit > bitregion_end_ + 1)
+      if (start + unit > bitregion_end_ + 1)
+	break;
+
+      /* Stop if the mode requires too much alignment.  */
+      if (unit > align_ && SLOW_UNALIGNED_ACCESS (mode_, align_))
 	break;
 
       *out_mode = mode_;
@@ -2751,6 +2750,62 @@  get_best_mode (int bitsize, int bitpos,
   enum machine_mode widest_mode = VOIDmode;
   enum machine_mode mode;
   while (iter.next_mode (&mode)
+	 /* ??? For historical reasons, reject modes that are wider than
+	    the alignment.  This has both advantages and disadvantages.
+	    Removing this check means that something like:
+
+	       struct s { unsigned int x; unsigned int y; };
+	       int f (struct s *s) { return s->x == 0 && s->y == 0; }
+
+	    can be implemented using a single load and compare on
+	    64-bit machines that have no alignment restrictions.
+	    For example, on powerpc64-linux-gnu, we would generate:
+
+		    ld 3,0(3)
+		    cntlzd 3,3
+		    srdi 3,3,6
+		    blr
+
+	    rather than:
+
+		    lwz 9,0(3)
+		    cmpwi 7,9,0
+		    bne 7,.L3
+		    lwz 3,4(3)
+		    cntlzw 3,3
+		    srwi 3,3,5
+		    extsw 3,3
+		    blr
+		    .p2align 4,,15
+	    .L3:
+		    li 3,0
+		    blr
+
+	    However, accessing more than one field can make life harder
+	    for the gimple optimizers.  For example, gcc.dg/vect/bb-slp-5.c
+	    has a series of unsigned short copies followed by a series of
+	    unsigned short comparisons.  With this check, both the copies
+	    and comparisons remain 16-bit accesses and FRE is able
+	    to eliminate the latter.  Without the check, the comparisons
+	    can be done using 2 64-bit operations, which FRE isn't able
+	    to handle in the same way.
+
+	    Either way, it would probably be worth disabling this check
+	    during expand.  One particular example where removing the
+	    check would help is the get_best_mode call in store_bit_field.
+	    If we are given a memory bitregion of 128 bits that is aligned
+	    to a 64-bit boundary, and the bitfield we want to modify is
+	    in the second half of the bitregion, this check causes
+	    store_bitfield to turn the memory into a 64-bit reference
+	    to the _first_ half of the region.  We later use
+	    adjust_bitfield_address to get a reference to the correct half,
+	    but doing so looks to adjust_bitfield_address as though we are
+	    moving past the end of the original object, so it drops the
+	    associated MEM_EXPR and MEM_OFFSET.  Removing the check
+	    causes store_bit_field to keep a 128-bit memory reference,
+	    so that the final bitfield reference still has a MEM_EXPR
+	    and MEM_OFFSET.  */
+	 && GET_MODE_BITSIZE (mode) <= align
 	 && (largest_mode == VOIDmode
 	     || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (largest_mode)))
     {