[PATCH] D22630: Loop rotation

[Sebastian Pop via llvm-commits]

sebpop added inline comments.

================
Comment at: llvm/lib/Transforms/Scalar/LoopRotation.cpp:431
@@ +430,3 @@
+    DEBUG(dbgs() << "\nSplitting the edge of Loop:"; L->dumpVerbose(););
+    LoopLatch = SplitEdge(LoopLatch, L->getHeader(), DT, LI);
+  }
----------------
eli.friedman wrote:
> sebpop wrote:
> > hiraditya wrote:
> > > eli.friedman wrote:
> > > > The loop latch could be an indirectbr.
> > > I'll fix this, thanks.
> > I have the impression that the loop analysis will not find a natural loop if the edge from the latch to the header is an indirect branch.
> > Are you sure this is possible?  Do you have a test-case where we recognize a loop with an indirect branch for the latch edge?
> indirectbr loop in C:
> 
>     int g(), h(), z();
>     int f() {
>         static void* x[] = {&&F, &&G, &&H};
>         F: 
>         goto *x[z()];
>         
>         G: return g();
>         H: return h();
>     }
> 
> `clang -x c - -o - -S -O2 -emit-llvm | opt -analyze -loops` gives:
> 
>     Printing analysis 'Natural Loop Information' for function 'f':
>     Loop at depth 1 containing: %F<header><latch><exiting>
Thanks Eli for the testcase: we will add it to the testsuite and we will add a check "if (isa<IndirectBrInst>(LoopLatch->getTerminator()))" before entering the loop rotation for that loop.


https://reviews.llvm.org/D22630