[Vm-dev] patch: fix double alignment detection on Solaris/SPARC with gcc-4.4.x

[Andrew Gaylard]

On Mon, Jul 20, 2009 at 7:02 PM, David T. Lewis<lewis at mail.msen.com> wrote:
>
> On Sat, Jul 18, 2009 at 05:25:48PM +0200, Andrew Gaylard wrote:
>>
>> The main issue that I'm facing at the moment is that gcc-4.4.0 generates
>> code that causes unaligned accesses in fetchFloatAtinto; the same code
>> works in gcc-4.2.4 on SPARC.  Until I or someone else (hint!) fixes this,
>> I'm continuing to build with gcc-4.2.4.
>
> fetchFloatAtInto() is implemented in platforms/Cross/vm/sqMemoryAccess.h
> in the form of cpp macros. It might be tricky to debug this, but I note
> that the implementation depends on DOUBLE_WORD_ALIGNMENT, which is a macro
> that would have been set when you run configure. You might playing with
> this setting to see if it gives you the results you want. For example, as
> a complete SWAG, it might be the case that SPARC plus gcc-4.2.4 needs this
> macro defined but for some reason configure is not setting it for you.

David, thanks for the tip.  This message from configure showed
you are right (the 'yes' should be 'no'):

checking whether unaligned access to doubles is ok... yes

It turns out that newer GCCs, from 4.3.x onward, are smart
enough to optimise out the code used to detect strict alignment:

bash-3.2$ gcc -O2 double-test.c
double-test.c: In function 'main':
double-test.c:6: warning: passing argument 1 of 'f' makes integer from
pointer without a cast
double-test.c:1: note: expected 'int' but argument is of type 'char *'
bash-3.2$ ./a.out
bash-3.2$ echo $?
0

But when compiling without optimisation:

bash-3.2$ gcc double-test.c
double-test.c: In function 'main':
double-test.c:6: warning: passing argument 1 of 'f' makes integer from
pointer without a cast
double-test.c:1: note: expected 'int' but argument is of type 'char *'
bash-3.2$ ./a.out
Bus Error (core dumped)

... as expected.

This patch complicates the code sufficiently (by using the results
of the function) so that it triggers a bus error both with and without
optimisation, as would be expected.  Of course, configure needs
to be regenerated after applying this patch.

Running our internal benchmark reveals that it is worth it getting
4.4.1 to work:

gcc-4.4.1:

real    61m45.975s
user    56m16.247s
sys     4m57.253s

gcc-4.2.4:

real    67m29.260s
user    61m40.666s
sys     5m7.937s

This means the VM is sped up by 9%.

- Andrew

--- Squeak-3.10-4/platforms/unix/config/acinclude.m4.orig
2008-09-02 20:49:43.000000000 +0200
+++ Squeak-3.10-4/platforms/unix/config/acinclude.m4    2009-07-30
16:16:55.828929831 +0200
@@ -230,10 +230,33 @@


 AC_DEFUN([AC_C_DOUBLE_ALIGNMENT],
 [AC_CACHE_CHECK([whether unaligned access to doubles is ok],
ac_cv_double_align,
-  AC_TRY_RUN([f(int i){*(double *)i=*(double *)(i+4);}
-              int main(){char b[[12]];f(b);return 0;}],
+  AC_TRY_RUN([
+
+int f(void* i){
+       *(double *)i=*(double *)(i+4);
+       return *(char*)i;
+}
+int main(){
+       char b[[12]];
+       b[[0]]=1;
+       b[[1]]=2;
+       b[[2]]=3;
+       b[[3]]=4;
+       b[[4]]=0;
+       b[[5]]=0;
+       b[[6]]=0;
+       b[[7]]=0;
+       b[[8]]=0;
+       b[[9]]=0;
+       b[[10]]=0;
+       b[[11]]=0;
+
+       return f(b);
+}
+],
+
     ac_cv_double_align="yes", ac_cv_double_align="no"))
 test "$ac_cv_double_align" = "no" && AC_DEFINE(DOUBLE_WORD_ALIGNMENT)])

 AC_DEFUN([AC_C_DOUBLE_ORDER],