[llvm] r309758 - [SCEV/IndVars] Always compute loop exiting values if the backedge count is 0

[Hans Wennborg via llvm-commits]

On Tue, Aug 1, 2017 at 4:12 PM, Sanjoy Das
<sanjoy at playingwithpointers.com> wrote:
> Hi,
>
> On Tue, Aug 1, 2017 at 3:45 PM, Davide Italiano <davide at freebsd.org> wrote:
>> Sanjoy, Hans, can we backport this one to 5.0?
>
> SGTM, but I'd like to bake this in tree for a day or two before we do that.

Sounds good to me too.


>>
>> Thanks,
>>
>> --
>> Davide
>>
>> On Tue, Aug 1, 2017 at 3:37 PM, Sanjoy Das via llvm-commits
>> <llvm-commits at lists.llvm.org> wrote:
>>> Author: sanjoy
>>> Date: Tue Aug  1 15:37:58 2017
>>> New Revision: 309758
>>>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=309758&view=rev
>>> Log:
>>> [SCEV/IndVars] Always compute loop exiting values if the backedge count is 0
>>>
>>> If SCEV can prove that the backedge taken count for a loop is zero, it does not
>>> need to "understand" a recursive PHI to compute its exiting value.
>>>
>>> This should fix PR33885.
>>>
>>> Modified:
>>>     llvm/trunk/lib/Analysis/ScalarEvolution.cpp
>>>     llvm/trunk/test/Transforms/IndVarSimplify/exit_value_test2.ll
>>>
>>> Modified: llvm/trunk/lib/Analysis/ScalarEvolution.cpp
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolution.cpp?rev=309758&r1=309757&r2=309758&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/lib/Analysis/ScalarEvolution.cpp (original)
>>> +++ llvm/trunk/lib/Analysis/ScalarEvolution.cpp Tue Aug  1 15:37:58 2017
>>> @@ -7597,6 +7597,25 @@ const SCEV *ScalarEvolution::computeSCEV
>>>              const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
>>>              if (const SCEVConstant *BTCC =
>>>                    dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
>>> +
>>> +              // This trivial case can show up in some degenerate cases where
>>> +              // the incoming IR has not yet been fully simplified.
>>> +              if (BTCC->getValue()->isZero()) {
>>> +                Value *InitValue = nullptr;
>>> +                bool MultipleInitValues = false;
>>> +                for (unsigned i = 0; i < PN->getNumIncomingValues(); i++) {
>>> +                  if (!LI->contains(PN->getIncomingBlock(i))) {
>>> +                    if (!InitValue)
>>> +                      InitValue = PN->getIncomingValue(i);
>>> +                    else if (InitValue != PN->getIncomingValue(i)) {
>>> +                      MultipleInitValues = true;
>>> +                      break;
>>> +                    }
>>> +                  }
>>> +                  if (!MultipleInitValues && InitValue)
>>> +                    return getSCEV(InitValue);
>>> +                }
>>> +              }
>>>                // Okay, we know how many times the containing loop executes.  If
>>>                // this is a constant evolving PHI node, get the final value at
>>>                // the specified iteration number.
>>>
>>> Modified: llvm/trunk/test/Transforms/IndVarSimplify/exit_value_test2.ll
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/IndVarSimplify/exit_value_test2.ll?rev=309758&r1=309757&r2=309758&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/test/Transforms/IndVarSimplify/exit_value_test2.ll (original)
>>> +++ llvm/trunk/test/Transforms/IndVarSimplify/exit_value_test2.ll Tue Aug  1 15:37:58 2017
>>> @@ -3,15 +3,14 @@
>>>
>>>  ; Check IndVarSimplify should not replace exit value because or else
>>>  ; udiv will be introduced by expand and the cost will be high.
>>> -;
>>> -; CHECK-LABEL: @_Z3fooPKcjj(
>>> -; CHECK-NOT: udiv
>>>
>>>  declare void @_Z3mixRjj(i32* dereferenceable(4), i32)
>>>  declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture)
>>>  declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture)
>>>
>>>  define i32 @_Z3fooPKcjj(i8* nocapture readonly %s, i32 %len, i32 %c) {
>>> +; CHECK-LABEL: @_Z3fooPKcjj(
>>> +; CHECK-NOT: udiv
>>>  entry:
>>>    %a = alloca i32, align 4
>>>    %tmp = bitcast i32* %a to i8*
>>> @@ -50,3 +49,26 @@ while.end:
>>>    call void @llvm.lifetime.end.p0i8(i64 4, i8* %tmp)
>>>    ret i32 %tmp4
>>>  }
>>> +
>>> +define i32 @zero_backedge_count_test(i32 %unknown_init, i32* %unknown_mem) {
>>> +; CHECK-LABEL: @zero_backedge_count_test(
>>> +entry:
>>> +  br label %loop
>>> +
>>> +loop:
>>> +  %iv = phi i32 [ 0, %entry], [ %iv.inc, %loop ]
>>> +  %unknown_phi = phi i32 [ %unknown_init, %entry ], [ %unknown_next, %loop ]
>>> +  %iv.inc = add i32 %iv, 1
>>> +  %be_taken = icmp ne i32 %iv.inc, 1
>>> +  %unknown_next = load volatile i32, i32* %unknown_mem
>>> +  br i1 %be_taken, label %loop, label %leave
>>> +
>>> +leave:
>>> +; We can fold %unknown_phi even though the backedge value for it is completely
>>> +; unknown, since we can prove that the loop's backedge taken count is 0.
>>> +
>>> +; CHECK: leave:
>>> +; CHECK: ret i32 %unknown_init
>>> +  %exit_val = phi i32 [ %unknown_phi, %loop ]
>>> +  ret i32 %exit_val
>>> +}
>>>
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>>
>>
>>
>> --
>> Davide
>>
>> "There are no solved problems; there are only problems that are more
>> or less solved" -- Henri Poincare