<html><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">This test case was indeed enlightning, thank you. Here's what I found (you may also skip the details and directly jump to my conclusion below):<br class=""><br class="">Detailed description<br class="">====================<br class=""><br class="">In the provided test case we have the following:<br class=""><br class="">```<br class=""> %tobool.i = icmp eq i8 %1, 0<br class=""> %2 = or i32 %shr.i.i.i, 1<br class=""> %3 = icmp eq i32 %2, 5<br class=""> %4 = select i1 %tobool.i, i1 %3, i1 true<br class=""> %frombool.i = zext i1 %4 to i8<br class="">```<br class=""><br class="">At some point it would be transformed into this:<br class=""><br class="">```<br class="">%2 = icmp eq i32 %1, 5<br class="">%not.tobool.i = icmp ne i8 %has_double.0.i, 0<br class="">%3 = or i1 %2, %not.tobool.i<br class="">%frombool.i = zext i1 %3 to i8<br class="">```<br class=""><br class="">It now has the form `zext(or(icmp, icmp))`. In `visitZExt()` in InstCombineCasts.cpp this would run throgh the below code:<br class=""><br class="">```<br class="">if (SrcI && SrcI->getOpcode() == Instruction::Or) {<br class=""> // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one<br class=""> // of the (zext icmp) will be transformed.<br class=""> ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));<br class=""> ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));<br class=""> if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&<br class=""> (transformZExtICmp(LHS, CI, false) ||<br class=""> transformZExtICmp(RHS, CI, false))) {<br class=""> Value *LCast = Builder->CreateZExt(LHS, CI.getType(), LHS->getName());<br class=""> Value *RCast = Builder->CreateZExt(RHS, CI.getType(), RHS->getName());<br class=""> return BinaryOperator::Create(Instruction::Or, LCast, RCast);<br class=""> }<br class="">}<br class="">```<br class=""><br class="">Now for `transformZExtICmp(%not.tobool.i, %frombool.i, false)` it would see that this `zext icmp` would be able to be eliminated and therefore it hoists the `zext` in front of the `icmp` instructions:<br class=""><br class="">```<br class="">%2 = icmp eq i32 %1, 5<br class="">%3 = zext i1 %2 to i8<br class="">%not.tobool.i = icmp ne i8 %has_double.0.i, 0<br class="">%not.tobool.i1 = zext i1 %not.tobool.i to i8<br class="">%frombool.i = or i8 %3, %not.tobool.i1<br class="">```<br class=""><br class="">Now, in `foldCastedBitwiseLogic()` we have a reverse transformation that tries to unfold expressions like this by sinking the zext after the `or` which could lead to an endless loop switching back and forth between these transformations -- which happens for the given case. As described in r276105, I try to avoid such an endless loop by checking in `shouldOptimizeCast()` if a `zext icmp` pair might possibly be the result of the above unfolding transformation. Therefore, `shouldOptimizeCast()` also uses `transformZExtIcmp()`. Now the interesting part is that in `shouldOptimizeCast()` we get a different result from `transformZExtICmp(%not.tobool.i, %not.tobool.i1, false)` as for the `transformZExtICmp(%not.tobool.i, %frombool.i, false)` from above, despite I originally thought they basically reference the same instructions. However, it seems that the instructions have slightly changed since the unfolding, leading to a different result in `transformZExtICmp`. This means, we now conclude in `foldCastedBitwiseLogic()` that we can safely fold the casts back to `zext(or(icmp, icmp))`. Later, it seems that this would again get unfolded in `visitZExt()`, etc., leading the an endless loop.<br class=""><br class="">Conclusion<br class="">==========<br class=""><br class="">To summarize this: For the given test case, I do not yet fully know the exact reason why our check cannot prevent this endless loop. However, it's more ore less apparent that we cannot guarantee after the transformation from `zext(or(icmp, icmp))` to `or(zext(icmp), zext(icmp))` that these `zext(icmp)` are not changed in the meantime by another InstCombine optimization before they are passed to `foldCastedBitwiseLogic()`. That means, the results of `transformZExtICmp()` in `shouldOptimizeCast()` here might differ from those in `visitZExt()`. Now I see that the original check in `foldCastedBitwiseLogic()` which generally forbids `icmp` instructions makes a lot of sense.<br class=""><br class="">However, what I do not yet understand is this check in `visitZExt()` that guards the unfolding transformation. Why is it restricted to `or` instructions only? I also wonder if the unfolding really valuable? - which, I admit, might be a naive question :) It is only tested in the `zext-or-icmp.ll`, and when I remove this unfolding it seems that it doesn't influence the other test cases (I ran all LLVM test cases via the check target).<br class=""><br class="">Best,<br class="">Matthias<br class=""><br class=""><br class=""><blockquote type="cite" class="">Am 20.07.2016 um 14:13 schrieb Matthias Reisinger <<a href="mailto:matthias.j.reisinger@gmail.com" class="">matthias.j.reisinger@gmail.com</a>>:<br class=""><br class="">Thank for bringing this to our notice and for providing the test case, I’ll investigate this. I apologize for the inconvenience.<br class=""><br class="">Best,<br class="">Matthias<br class=""><br class=""><blockquote type="cite" class="">Am 20.07.2016 um 14:04 schrieb Tobias Grosser <<a href="mailto:tobias@grosser.es" class="">tobias@grosser.es</a>>:<br class=""><br class="">Thank you Benjamin. Sorry for the breakage.<br class=""><br class="">Best,<br class="">Tobias<br class=""><br class="">On Wed, Jul 20, 2016, at 01:47 PM, Benjamin Kramer via llvm-commits<br class="">wrote:<br class=""><blockquote type="cite" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: 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;" class="">This sends 'opt -instcombine' into an infinite loop on the attached<br class="">test case, can you take a look? I reverted this change (and r276105)<br class="">in r276106.<br class=""><br class="">On Tue, Jul 19, 2016 at 6:39 PM, Tobias Grosser via llvm-commits<br class=""><llvm-commits@lists.llvm.org> wrote:<br class=""><blockquote type="cite" class="">Author: grosser<br class="">Date: Tue Jul 19 11:39:17 2016<br class="">New Revision: 275989<br class=""><br class="">URL: http://llvm.org/viewvc/llvm-project?rev=275989&view=rev<br class="">Log:<br class="">[InstCombine] Enable cast-folding in logic(cast(icmp), cast(icmp))<br class=""><br class="">Summary:<br class="">Currently, InstCombine is already able to fold expressions of the form `logic(cast(A), cast(B))` to the simpler form `cast(logic(A, B))`, where logic designates one of `and`/`or`/`xor`. This transformation is implemented in `foldCastedBitwiseLogic()` in InstCombineAndOrXor.cpp. However, this optimization will not be performed if both `A` and `B` are `icmp` instructions. The decision to preclude casts of `icmp` instructions originates in r48715 in combination with r261707, and can be best understood by the title of the former one:<br class=""><br class=""><blockquote type="cite" class="">Transform (zext (or (icmp), (icmp))) to (or (zext (cimp), (zext icmp))) if at least one of the (zext icmp) can be transformed to eliminate an icmp.<br class=""></blockquote><br class="">Apparently, it introduced a transformation that is a reverse of the transformation that is done in `foldCastedBitwiseLogic()`. Its purpose is to expose pairs of `zext icmp` that would subsequently be optimized by `transformZExtICmp()` in InstCombineCasts.cpp. Therefore, in order to avoid an endless loop of switching back and forth between these two transformations, the one in `foldCastedBitwiseLogic()` has been restricted to exclude `icmp` instructions which is mirrored in the responsible check:<br class=""><br class="">`if ((!isa<ICmpInst>(Cast0Src) || !isa<ICmpInst>(Cast1Src)) && ...`<br class=""><br class="">This check seems to sort out more cases than necessary because:<br class="">- the reverse transformation is obviously done for `or` instructions only<br class="">- and also not every `zext icmp` pair is necessarily the result of this reverse transformation<br class=""><br class="">Therefore we now remove this check and replace it by a more finegrained one in `shouldOptimizeCast()` that now rejects only those `logic(zext(icmp), zext(icmp))` that would be able to be optimized by `transformZExtICmp()`, which also avoids the mentioned endless loop. That means we are now able to also simplify expressions of the form `logic(cast(icmp), cast(icmp))` to `cast(logic(icmp, icmp))` (`cast` being an arbitrary `CastInst`).<br class=""><br class="">As an example, consider the following IR snippet<br class=""><br class="">```<br class="">%1 = icmp sgt i64 %a, %b<br class="">%2 = zext i1 %1 to i8<br class="">%3 = icmp slt i64 %a, %c<br class="">%4 = zext i1 %3 to i8<br class="">%5 = and i8 %2, %4<br class="">```<br class=""><br class="">which would now be transformed to<br class=""><br class="">```<br class="">%1 = icmp sgt i64 %a, %b<br class="">%2 = icmp slt i64 %a, %c<br class="">%3 = and i1 %1, %2<br class="">%4 = zext i1 %3 to i8<br class="">```<br class=""><br class="">This issue became apparent when experimenting with the programming language Julia, which makes use of LLVM. Currently, Julia lowers its `Bool` datatype to LLVM's `i8` (also see https://github.com/JuliaLang/julia/pull/17225). In fact, the above IR example is the lowered form of the Julia snippet `(a > b) & (a < c)`. Like shown above, this may introduce `zext` operations, casting between `i1` and `i8`, which could for example hinder ScalarEvolution and Polly on certain code.<br class=""><br class="">Reviewers: grosser, vtjnash, majnemer<br class=""><br class="">Subscribers: majnemer, llvm-commits<br class=""><br class="">Differential Revision: https://reviews.llvm.org/D22511<br class=""><br class="">Contributed-by: Matthias Reisinger<br class=""><br class="">Modified:<br class=""> llvm/trunk/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp<br class=""> llvm/trunk/test/Transforms/InstCombine/zext.ll<br class=""><br class="">Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp<br class="">URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp?rev=275989&r1=275988&r2=275989&view=diff<br class="">==============================================================================<br class="">--- llvm/trunk/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (original)<br class="">+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp Tue Jul 19 11:39:17 2016<br class="">@@ -1212,6 +1212,13 @@ bool InstCombiner::shouldOptimizeCast(Ca<br class=""> isa<CmpInst>(CastSrc) && CI->getDestTy()->isVectorTy())<br class=""> return false;<br class=""><br class="">+ // Don't optimize the cast if it is a (zext icmp) that can already be<br class="">+ // eliminated.<br class="">+ if (auto *ZExt = dyn_cast<ZExtInst>(CI))<br class="">+ if (auto *ICmp = dyn_cast<ICmpInst>(CastSrc))<br class="">+ if (transformZExtICmp(ICmp, *ZExt, false))<br class="">+ return false;<br class="">+<br class=""> return true;<br class="">}<br class=""><br class="">@@ -1260,8 +1267,7 @@ Instruction *InstCombiner::foldCastedBit<br class=""> Value *Cast1Src = Cast1->getOperand(0);<br class=""><br class=""> // fold logic(cast(A), cast(B)) -> cast(logic(A, B))<br class="">- if ((!isa<ICmpInst>(Cast0Src) || !isa<ICmpInst>(Cast1Src)) &&<br class="">- shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {<br class="">+ if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {<br class=""> Value *NewOp = Builder->CreateBinOp(LogicOpc, Cast0Src, Cast1Src,<br class=""> I.getName());<br class=""> return CastInst::Create(CastOpcode, NewOp, DestTy);<br class=""><br class="">Modified: llvm/trunk/test/Transforms/InstCombine/zext.ll<br class="">URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/zext.ll?rev=275989&r1=275988&r2=275989&view=diff<br class="">==============================================================================<br class="">--- llvm/trunk/test/Transforms/InstCombine/zext.ll (original)<br class="">+++ llvm/trunk/test/Transforms/InstCombine/zext.ll Tue Jul 19 11:39:17 2016<br class="">@@ -73,3 +73,41 @@ define <2 x i64> @fold_xor_zext_sandwich<br class=""> ret <2 x i64> %zext2<br class="">}<br class=""><br class="">+; Assert that zexts in logic(zext(icmp), zext(icmp)) can be folded<br class="">+; CHECK-LABEL: @fold_logic_zext_icmp(<br class="">+; CHECK-NEXT: [[ICMP1:%.*]] = icmp sgt i64 %a, %b<br class="">+; CHECK-NEXT: [[ICMP2:%.*]] = icmp slt i64 %a, %c<br class="">+; CHECK-NEXT: [[AND:%.*]] = and i1 [[ICMP1]], [[ICMP2]]<br class="">+; CHECK-NEXT: [[ZEXT:%.*]] = zext i1 [[AND]] to i8<br class="">+; CHECK-NEXT: ret i8 [[ZEXT]]<br class="">+define i8 @fold_logic_zext_icmp(i64 %a, i64 %b, i64 %c) {<br class="">+ %1 = icmp sgt i64 %a, %b<br class="">+ %2 = zext i1 %1 to i8<br class="">+ %3 = icmp slt i64 %a, %c<br class="">+ %4 = zext i1 %3 to i8<br class="">+ %5 = and i8 %2, %4<br class="">+ ret i8 %5<br class="">+}<br class="">+<br class="">+; Assert that zexts in logic(zext(icmp), zext(icmp)) are also folded accross<br class="">+; nested logical operators.<br class="">+; CHECK-LABEL: @fold_nested_logic_zext_icmp(<br class="">+; CHECK-NEXT: [[ICMP1:%.*]] = icmp sgt i64 %a, %b<br class="">+; CHECK-NEXT: [[ICMP2:%.*]] = icmp slt i64 %a, %c<br class="">+; CHECK-NEXT: [[AND:%.*]] = and i1 [[ICMP1]], [[ICMP2]]<br class="">+; CHECK-NEXT: [[ICMP3:%.*]] = icmp eq i64 %a, %d<br class="">+; CHECK-NEXT: [[OR:%.*]] = or i1 [[AND]], [[ICMP3]]<br class="">+; CHECK-NEXT: [[ZEXT:%.*]] = zext i1 [[OR]] to i8<br class="">+; CHECK-NEXT: ret i8 [[ZEXT]]<br class="">+define i8 @fold_nested_logic_zext_icmp(i64 %a, i64 %b, i64 %c, i64 %d) {<br class="">+ %1 = icmp sgt i64 %a, %b<br class="">+ %2 = zext i1 %1 to i8<br class="">+ %3 = icmp slt i64 %a, %c<br class="">+ %4 = zext i1 %3 to i8<br class="">+ %5 = and i8 %2, %4<br class="">+ %6 = icmp eq i64 %a, %d<br class="">+ %7 = zext i1 %6 to i8<br class="">+ %8 = or i8 %5, %7<br class="">+ ret i8 %8<br class="">+}<br class="">+<br class=""><br class=""><br class="">_______________________________________________<br class="">llvm-commits mailing list<br class="">llvm-commits@lists.llvm.org<br class="">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits<br class=""></blockquote>_______________________________________________<br class="">llvm-commits mailing list<br class="">llvm-commits@lists.llvm.org<br class="">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits<br class="">Email had 1 attachment:<br class="">+ bugpoint-reduced-simplified.ll<br class=""> 3k (application/octet-stream)<br class=""></blockquote></blockquote><br class=""></blockquote><br class=""></body></html>