[llvm-dev] Can't get an induction variable by getInductionVariable API

[Florian Hahn via llvm-dev]

Hi,

> On Oct 23, 2020, at 09:23, ZIN MOON via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Dear All
> 
> I have a question about the getInductionVariable API.
> I was trying to get an induction variable (indvar) in the code below using the getInductionVariable API (LoopInfo class).
> 
> C file
> void test(int val) {
>   int sum = 0, i = 0;
>   for (i = 0; i < val; i++) {
>       sum += i;
>   }
> }
> 
> IR (with mem2reg pass)
> ; Function Attrs: nounwind
> define dso_local void @test(i32 %0) #0 {
>   br label %2
> 2:                                                ; preds = %6, %1
>   %.01 = phi i32 [ 0, %1 ], [ %5, %6 ]
>   %.0 = phi i32 [ 0, %1 ], [ %7, %6 ]
>   %3 = icmp slt i32 %.0, %0
>   br i1 %3, label %4, label %8 <- Why don't check this BranchInst ?
> 4:                                                ; preds = %2
>   %5 = add nsw i32 %.01, %.0
>   br label %6
> 6:                                                ; preds = %4
>   %7 = add nsw i32 %.0, 1
>   br label %2  <- checked it whether isCondition by getLatchCmpInst API(called by getInductionVariable API)
> 8:                                                ; preds = %2
>   ret void
> }
> 
> Then I got null (no indvar) because BranchInst's number of operands was less than 2.(red block).
> 
> A CmpInst(compare with induction variable) is contained in the header of the loop.
> But the getInductionVariable API checks only in a latch block(contains a branch back to the header of the loop)
> is there a missing case that contains CmpInst in the header?
> 
> And if I use the getCanonicalInductionVariable API, I can get the variable 'i'(%. 01) as an indvar.
> 
> I checked the code implementing getInductionVariable and getCanonicalInductionVariable.
> It seems that indvar and canonical indvar differ only in the starting number (0 or none).
> Why not use the same way to detect the indvar between getInductionVariable and getCanonicalInductionVariable.


Most loop transformations and utilities expect to work on ‘rotated’ loops (https://www.llvm.org/docs/LoopTerminology.html#rotated-loops). It is probably best to also run `-loop-rotate` before transforming analyzing/transforming the loop.

Note that the documentation for getInductionVariable explicitly requires the induction variable to determine the condition of the branch in the loop latch (http://llvm.org/doxygen/classllvm_1_1Loop.html#ab05e97728516fbeeaa9426496257c800) and uses InductionDescriptor::isInductionPHI to detect induction PHIs (plus some other checks). There’s also isAuxiliaryInductionVariable, which does not have the latch branch requirement or you can use InductionDescriptor directly.

Cheers,
Florian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201023/84cada57/attachment.html>