<div dir="ltr"><div>Hi,</div><div><br></div>I would say that GCC is wrong and should also have a version where a could be equal to b. There is no restrict keyword, so they could be equal.<div><br></div><div>Cheers,</div><div><br></div><div>Matthieu</div></div><br><div class="gmail_quote"><div dir="ltr">Le jeu. 20 sept. 2018 à 16:56, Jonas Paulsson via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>> a écrit :<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div text="#000000" bgcolor="#FFFFFF">
<p>Hi,</p>
<p>I found a difference between Clang and GCC in alias handling.
This was with a benchmark where Clang was considerably slower, and
in a hot function which does many more loads from the same address
due to stores between the uses. In other words, a value is loaded
and used, another value is stored, and then the first value is
loaded once again before its second use. This happens many times,
with three loads instead of one for each value. GCC only emits one
load.</p>
<p>The values are the arguments to this function:<br>
</p>
<tt>void su3_projector( su3_vector *a, su3_vector *b, su3_matrix *c
){</tt><br>
<tt>register int i,j;</tt><br>
<tt>register double tmp,tmp2;</tt><br>
<tt> for(i=0;i<3;i++)for(j=0;j<3;j++){</tt><br>
<tt> tmp2 = a->c[i].real * b->c[j].real;</tt><br>
<tt> tmp = a->c[i].imag * b->c[j].imag;</tt><br>
<tt> c->e[i][j].real = tmp + tmp2;</tt><br>
<tt> tmp2 = a->c[i].real * b->c[j].imag;</tt><br>
<tt> tmp = a->c[i].imag * b->c[j].real;</tt><br>
<tt> c->e[i][j].imag = tmp - tmp2;</tt><br>
<tt> }</tt><br>
<tt>}</tt><br>
<tt><br>
The types are:</tt><br>
<tt>typedef struct { complex e[3][3]; } su3_matrix;</tt><br>
<tt>typedef struct { complex c[3]; } su3_vector;<br>
</tt><br>
So the question here is if the su3_vector and su3_matrix pointers
may alias? If they may alias, then clang is right in reloading after
each store. If the standard says they cannot alias, then gcc is
right in only loading the values once each.<br>
<br>
It seems to me that either GCC is too aggressive or LLVM is too
conservative, but I don't know which one it is... As far as I
understand, there is the fact of the different struct types of the
arguments (which means they cannot alias), but also the question if
su3_vector is included in su3_matrix (which would mean they may
alias).<br>
<br>
I made a reduced test case, where the same difference seems to be
present. It has just one struct type which contains a matrix of
double:s. A store to an element of the struct via a pointer is
surrounded with two loads of a global double variable. Only Clang
emits two loads.<br>
<br>
<tt>typedef struct {</tt><tt><br>
</tt><tt> double c[3][3];</tt><tt><br>
</tt><tt>} STRUCT_TY;</tt><tt><br>
</tt><tt><br>
</tt><tt>double e = 0.0;</tt><tt><br>
</tt><tt>STRUCT_TY *f;</tt><tt><br>
</tt><tt>int g = 0;</tt><tt><br>
</tt><tt>void h() {</tt><tt><br>
</tt><tt> int i = e;</tt><tt><br>
</tt><tt> f->c[0][i] = g;</tt><tt><br>
</tt><tt> g = e;</tt><tt><br>
</tt><tt>}</tt><tt><br>
</tt><tt><br>
</tt><tt>clang -O3-march=z13 :</tt><tt><br>
</tt><tt><br>
</tt><tt>h: # @h </tt><tt><br>
</tt><tt># %bb.0: # %entry </tt><tt><br>
</tt><tt> larl %r1, e </tt><tt><br>
</tt><tt> ld %f0, 0(%r1) // LOAD E</tt><tt><br>
</tt><tt> lrl %r2, g </tt><tt><br>
</tt><tt> cfdbr %r0, 5, %f0 // CONVERT E </tt><tt><br>
</tt><tt> lgfr %r0, %r0 // EXTEND E </tt><tt><br>
</tt><tt> cdfbr %f0, %r2 </tt><tt><br>
</tt><tt> lgrl %r2, f </tt><tt><br>
</tt><tt> sllg %r3, %r0, 3 </tt><tt><br>
</tt><tt> std %f0, 0(%r3,%r2) // STORE F EL</tt><tt>EMENT</tt><tt>
</tt><tt><br>
</tt><tt> ld %f0, 0(%r1) // 2nd LOAD E </tt><tt><<<<<<<<br>
</tt><tt> cfdbr %r0, 5, %f0 // CONVERT </tt><tt><br>
</tt><tt> strl %r0, g // 2nd USE </tt><tt><br>
</tt><tt> br %r14 </tt><tt><br>
</tt><tt><br>
</tt><tt>gcc -O3-march=z13 :</tt><tt><br>
</tt><tt><br>
</tt><tt>h:</tt><tt><br>
</tt><tt>.LFB0:</tt><tt><br>
</tt><tt> .cfi_startproc</tt><tt><br>
</tt><tt> larl %r1,e</tt><tt><br>
</tt><tt> ld %f0,0(%r1)</tt><tt> // LOAD E</tt><tt><br>
</tt><tt> lrl %r2,g</tt><tt><br>
</tt><tt> lgrl %r3,f</tt><tt><br>
</tt><tt> cfdbr %r1,5,%f0</tt><tt> // CONVERT E<br>
</tt><tt> cdfbr %f0,%r2</tt><tt><br>
</tt><tt> lgfr %r2,%r1 // EXTEND E</tt><tt><br>
</tt><tt> sllg %r2,%r2,3</tt><tt><br>
</tt><tt> std %f0,0(%r2,%r3) // STORE F ELEMENT</tt><tt><br>
</tt><tt> strl %r1,g // 2nd USE</tt><tt><br>
</tt><tt> br %r14</tt><tt><br>
<br>
</tt>I hope somebody with enough experience and knowledge can guide
the way here as this seems to be quite important.<br>
<br>
/Jonas<br>
<br>
<br>
<font size="2"><br>
</font><tt><font size="2"></font><font size="2"></font></tt>
</div>
_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
</blockquote></div><br clear="all"><div><br></div>-- <br><div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div>Quantitative analyst, Ph.D.<br>Blog: <a href="http://blog.audio-tk.com/" target="_blank">http://blog.audio-tk.com/</a><br>LinkedIn: <a href="http://www.linkedin.com/in/matthieubrucher" target="_blank">http://www.linkedin.com/in/matthieubrucher</a></div></div></div></div></div>