<html><head><meta http-equiv="Content-Type" content="text/html charset=windows-1252"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><br><div><div>On Oct 20, 2013, at 12:47 AM, Chris Lattner <<a href="mailto:clattner@apple.com">clattner@apple.com</a>> wrote:</div><br class="Apple-interchange-newline"><blockquote type="cite"><meta http-equiv="Content-Type" content="text/html charset=us-ascii"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><br><div><div>On Oct 20, 2013, at 12:04 AM, Michael Gottesman <<a href="mailto:mgottesman@apple.com">mgottesman@apple.com</a>> wrote:</div><br class="Apple-interchange-newline"><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;">Author: mgottesman<br>Date: Sun Oct 20 02:04:37 2013<br>New Revision: 193045<br><br>URL:<span class="Apple-converted-space"> </span><a href="http://llvm.org/viewvc/llvm-project?rev=193045&view=rev">http://llvm.org/viewvc/llvm-project?rev=193045&view=rev</a><br>Log:<br>Teach simplify-cfg how to correctly create covered lookup tables for switches on iN with N >= 3.<br></div></blockquote><div dir="auto"><br></div>Thanks Michael. More specifically, this is handling "fully-covered" switches. A couple of comments:</div></div></blockquote><blockquote type="cite"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><div dir="auto"><br></div><div dir="auto"><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;">+ // Compute the maximum table size representable by the integer type we are<br>+ // switching upon.<br>+ const unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();</div></blockquote><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;">+ const uint64_t MaxTableSize = CaseSize > 63? UINT64_MAX : 1ULL << CaseSize;<br></div></blockquote><div dir="auto"><br></div><div dir="auto">We don't generally mark local variables const like this. Please don’t.</div></div></div></blockquote><div><br></div><div>This was a last minute thing that crept in. Its gone.</div><br><blockquote type="cite"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><div dir="auto"><br><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;">+ assert(MaxTableSize >= TableSize &&<br>+ "It is impossible for a switch to have more entries than the max "<br>+ "representable value of its input integer type's size.");<br>+<br>+ // If we have a covered lookup table, unconditionally branch to the lookup table<br>+ // BB. Otherwise, check if the condition value is within the case range. If it<br>+ // is so, branch to the new BB. Otherwise branch to SI's default destination.<br>+ const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;<br></div></blockquote><div dir="auto"><br></div><div dir="auto">This is a "fully covered" table.</div></div></div></blockquote><div><br></div><div>Ok.</div><br><blockquote type="cite"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><div dir="auto"><br><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;">+ if (GeneratingCoveredLookupTable) {<br>+ Builder.CreateBr(LookupBB);<br>+ } else {<br>+ Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(<br>+ MinCaseVal->getType(), TableSize));<br>+ Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());<br>+ }<br><br> // Populate the BB that does the lookups.<br> Builder.SetInsertPoint(LookupBB);<br>@@ -3772,7 +3788,13 @@ static bool SwitchToLookupTable(SwitchIn<br> // Remove the switch.<br> for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {<br></div></blockquote><div dir="auto"><br></div><div dir="auto">This is not a problem in your patch, but there is no reason to evaluate getNumSuccessors() each time through the loop.</div></div></div></blockquote><div><br></div><div>Sounds good.</div><br><blockquote type="cite"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><div dir="auto"><br><blockquote type="cite"><div style="font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;"> BasicBlock *Succ = SI->getSuccessor(i);<br>- if (Succ == SI->getDefaultDest()) continue;<br>+<br>+ // If we are not generating a covered lookup table, we will have a<br>+ // conditional branch from SI's parent BB to SI's default destination if our<br>+ // input value lies outside of our case range. Thus in that case leave the<br>+ // default destination BB as a predecessor of SI's parent BB.<br>+ if (Succ == SI->getDefaultDest() && !GeneratingCoveredLookupTable)<br>+ continue;<br></div></blockquote><div dir="auto"><br></div><div dir="auto">This doesn't seem like the right check. If there is a switch whose default destination is "BB" and there are explicit edges to the same block, this will incorrectly remove multiple predecessors. The right patch for this is simply:</div><div dir="auto"><br></div><div dir="auto"><div> if (GeneratingCoveredLookupTable) {<br> Builder.CreateBr(LookupBB);</div><div> SI->getDefaultDest()->removePredecessor(SI->getParent());</div><div> } else {<br> Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(<br> MinCaseVal->getType(), TableSize));<br> Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());<br> }<br></div><div> </div></div></div></div></blockquote><blockquote type="cite"><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;"><div dir="auto"><div dir="auto"><div>Without this logic in the loop.</div></div></div></div></blockquote><div><br></div><div>Ok you are right. Removing a successor which is no longer a successor is not a good thing to do = /. (*DOH*)</div><div><br></div><div>How does the attached patch look?</div><div><br></div><div></div></div></body></html>