Added sections about debugging mis-compilations and incorrect code generation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@8584 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Misha Brukman 2003-09-17 18:51:47 +00:00
parent 0bb806bd9a
commit f1d01fb6a4

View File

@ -20,8 +20,10 @@
<li><a href="#passes">Bugs in LLVM passes</a> <li><a href="#passes">Bugs in LLVM passes</a>
</ul> </ul>
<li><a href="#miscompilations">Miscompilations</a> <li><a href="#miscompilations">Miscompilations</a>
<li><a href="#codegen">Incorrect code generation (JIT and LLC)</a>
<p><b>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></b><p> <p><b>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a> and
<a href="http://misha.brukman.net">Misha Brukman</a></b><p>
</ol><p></font> </ol><p></font>
</td><td valign=top align=right> </td><td valign=top align=right>
<img src="Debugging.gif" width=444 height=314> <img src="Debugging.gif" width=444 height=314>
@ -138,7 +140,7 @@ being linked together (the "<tt><b>llvm-gcc</b> -v</tt>" output should include
the full list of objects linked). Then run:<p> the full list of objects linked). Then run:<p>
<pre> <pre>
<b>as</b> &lt; /dev/null &gt; null.bc <b>llvm-as</b> &lt; /dev/null &gt; null.bc
<b>gccld</b> -debug-pass=Arguments null.bc <b>gccld</b> -debug-pass=Arguments null.bc
</pre><p> </pre><p>
@ -180,13 +182,80 @@ to reproduce the problem to the llvmbugs mailing list.<p>
<!-- *********************************************************************** --> <!-- *********************************************************************** -->
</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0> </ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b> <tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
<a name="miscompilations">Miscompilations <a name="miscompilations">Miscompilations</a>
</b></font></td></tr></table><ul> </b></font></td></tr></table><ul>
<!-- *********************************************************************** --> <!-- *********************************************************************** -->
Fortunately we haven't had too many miscompilations. Because of this, this A miscompilation occurs when a pass does not correctly transform a program, thus
section is a TODO. Basically, use bugpoint to track down the problem.<p> producing errors that are only noticed during execution. This is different from
producing invalid LLVM code (i.e., code not in SSA form, using values before
defining them, etc.) which the verifier will check for after a pass finishes its
run.<p>
To debug a miscompilation, you should choose which program you wish to run the
output through, e.g. C backend, the JIT, or LLC, and a selection of passes, one
of which may be causing the error, and run, for example:
<pre>
<b>bugpoint</b> -run-cbe -mode=compile [... optimization passes ...] file-to-test.bc
</pre>
<tt>bugpoint</tt> will try to narrow down your list of passes to the one pass
that causes an error, and simplify the bytecode file as much as it can to assist
you. It will print a message letting you know how to reproduce the resulting
error.
<!-- *********************************************************************** -->
</ul><table width="100%" bgcolor="#330077" border=0 cellpadding=4 cellspacing=0>
<tr><td align=center><font color="#EEEEFF" size=+2 face="Georgia,Palatino"><b>
<a name="codegen">Incorrect code generation</a>
</b></font></td></tr></table><ul>
<!-- *********************************************************************** -->
Similarly to debugging incorrect compilation by mis-behaving passes, you can
debug incorrect code generation by either LLC or the JIT, using
<tt>bugpoint</tt>. The process <tt>bugpoint</tt> follows in this case is to try
to narrow the code down to a function that is miscompiled by one or the other
method, but since for correctness, the entire program must be run,
<tt>bugpoint</tt> will compile the code it deems to not be affected with the C
Backend, and then link in the shared object it generates.<p>
To debug the JIT:
<pre>
<b>bugpoint</b> -run-jit -mode=codegen -output=[correct output file] [bytecodefile]
</pre>
Similarly, to debug the LLC, one would run:
<pre>
<b>bugpoint</b> -run-llc -mode=codegen -output=[correct output file] [bytecodefile]
</pre>
At the end of a successful <tt>bugpoint</tt> run, you will be presented
with two bytecode files: a <em>safe</em> file which can be compiled with the C
backend and the <em>test</em> file which either LLC or the JIT
mis-codegenerates, and thus causes the error.<p>
To reproduce the error that <tt>bugpoint</tt> found, it is sufficient to do the
following:
<ol>
<li>Regenerate the shared object from the safe bytecode file:<br>
<pre>
<b>llvm-dis</b> -c safe.bc -o safe.c<br>
<b>gcc</b> -shared safe.c -o safe.so
</pre></li>
<li>If debugging LLC, compile test bytecode native and link with the shared object:<br>
<pre>
<b>llc</b> test.bc -o test.s -f<br>
gcc test.s safe.so -o test.llc<br>
./test.llc [program options]
</pre></li>
<p>
If debugging the JIT, load the shared object and supply the test bytecode:<br>
<pre>
<b>lli</b> -load=safe.so test.bc [program options]
</pre></li>
</ol>
<!-- *********************************************************************** --> <!-- *********************************************************************** -->
</ul> </ul>
@ -196,6 +265,6 @@ section is a TODO. Basically, use bugpoint to track down the problem.<p>
<address><a href="mailto:sabre@nondot.org">Chris Lattner</a></address> <address><a href="mailto:sabre@nondot.org">Chris Lattner</a></address>
<!-- Created: Tue Aug 6 15:00:33 CDT 2002 --> <!-- Created: Tue Aug 6 15:00:33 CDT 2002 -->
<!-- hhmts start --> <!-- hhmts start -->
Last modified: Fri May 23 09:48:53 CDT 2003 Last modified: Wed Sep 17 13:49:15 CDT 2003
<!-- hhmts end --> <!-- hhmts end -->
</font></body></html> </font></body></html>