<div dir="ltr">Hi Philip<div><br></div><div>Thanks for your response.<br><div><br></div><div>Yes, previously, I also thought there is one pass which could change the predicate from ">=" to "<", like some gvn passes did, update "-" to "+".</div><div>but I can't find it.</div><div><br></div><div>I tried to use -O3 to build the code, the target IR file shows all conditions are uniformed to "EQ". and then took the IRCE pass, no optimization occurs.</div><div>I didn't check the code why it didn't happen, because I know there is another thing which causes IRCE doesn't work when trip count comparison uses "EQ/NE", but it's irrelevant to this thread.<br></div></div><div><br></div><div>B.R</div><div>Jie He</div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, 5 Feb 2021 at 09:05, Philip Reames <<a href="mailto:listmail@philipreames.com">listmail@philipreames.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>
<p>)Quick response, may be off base as I'm taking a guess based on a
quick read through.)</p>
<p>IRCE - like many llvm passes - assumes IR has been canonicalized
by other passes. If you run your IR through -O3 and then IRCE, do
you get what you expect?</p>
<p>Philip<br>
</p>
<div>On 2/4/21 12:43 AM, Jie He via llvm-dev
wrote:<br>
</div>
<blockquote type="cite">
<div dir="ltr">
<pre id="gmail-m_-3890627333865380045gmail-comment_text_0" style="white-space:pre-wrap;width:50em;color:rgb(0,0,0)">Hi</pre>
<pre id="gmail-m_-3890627333865380045gmail-comment_text_0" style="white-space:pre-wrap;width:50em;color:rgb(0,0,0)">Recently, I tried to use llvm to optimize my code generated by a managed VM. this VM will instrument many range check codes before memory access operations.
I found the optimization IRCE wouldn't work when the range check with the predicate "uge". I wrote the following c/c++ code to simulate my code, where len is the max buf length:
void testIRCE(unsigned int * buf, unsigned int len, unsigned int iteration_count) {
if (iteration_count > 0) {
unsigned int i = 0;
do {
if (i >= len) { // range check
printf("overflow\n");
return;
}
buf[i] = i;
i ++;
} while (i < iteration_count);
}
}
the above code wouldn't be optimised as expected into 2 loops (iteration range splitting). I checked the llvm code, found in the function InductiveRangeCheck::parseRangeCheckICmp(), it wouldn't deal with uge cases, I'm not sure if it is intentional.
but I tried to modify the llvm IR code by replacing the uge to ult, and interchanging the operands of next branch instruction. the optimization IRCE works as expected.
my test command is below:
first, get a clean llvm IR file.
./clang++ -O3 -Xclang -disable-llvm-passes ~/testRCE.cpp -emit-llvm -S -o ~/testRCE.TBBA.ll
second, optimize it with other loop optimization.
./opt -gvn -simplifycfg -loop-simplify -loop-predication -licm -dce -mem2reg -dce -jump-threading -lcssa -simplifycfg -loop-simplify -dce -stats -debug-pass=Executions ~/testRCE.TBBA.ll -S -o ~/testRCE.mem2reg.ll
finally, take IRCE. </pre>
<span style="color:rgb(0,0,0);white-space:pre-wrap">./opt -irce-skip-profitability-checks -irce -dce -S ~/testRCE.mem2reg.ll -o ~/testRCE.irce.ll</span>--
<br>
<div dir="ltr"><br>
</div>
<div dir="ltr"><br>
</div>
<div dir="ltr">Best Regards<br>
He Jie 何杰</div>
</div>
<br>
<fieldset></fieldset>
<pre>_______________________________________________
LLVM Developers mailing list
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a>
</pre>
</blockquote>
</div>
</blockquote></div><br clear="all"><div><br></div>-- <br><div dir="ltr" class="gmail_signature">Best Regards<br>He Jie 何杰</div>