[LICM] Bail if checking a global/constant for invariant.start

[aeubanks]

Summary

When we check if a load is loop invariant by finding a dominating
invariant.start call, we strip bitcasts until we get to an i8* Value,
and look for an invariant.start use of the i8* Value.

We may accidentally end up at an i8 global and look at a global's uses,
which we shouldn't do in a loop pass. Although we could make this
logic work with globals, that's not currently intended.

[aeubanks]

remove FIXME

[nikic]

LGTM, this matches what we do elsewhere.

llvm/test/Transforms/LICM/hoisting.ll

605

This makes it sound like there's a correctness (or at least profitability) reason not to hoist these, while ultimately this is only a technical limitation (right?) I'd make this "Hoisting invariant loads of globals is currently not supported" or invert the fixme to "FIXME: Support hoisting invariant loads of globals".

[aeubanks]

update test comment

llvm/lib/Transforms/Scalar/LICM.cpp

static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
   // operand type).
   while (Addr->getType() != PtrInt8Ty) {
     auto *BC = dyn_cast<BitCastInst>(Addr);
     // Avoid traversing high number of bitcast uses.
     if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
       return false;
     Addr = BC->getOperand(0);
   }
+  // If we've ended up at a global/constant, bail. We shouldn't be looking at
+  // uselists for non-local Values in a loop pass.
+  if (isa<Constant>(Addr))
+    return false;
 
   unsigned UsesVisited = 0;
   // Traverse all uses of the load operand value, to see if invariant.start is
   // one of the uses, and whether it dominates the load instruction.
   for (auto *U : Addr->users()) {
     // Avoid traversing for Load operand with high number of users.
     if (++UsesVisited > MaxNumUsesTraversed)
       return false;

llvm/test/Transforms/LICM/hoisting.ll

 loopexit:
   ret i32 %sum
 }
 
 declare void @g(i1)
 
 @a = external global i8
 
 ; FIXME: Don't hoist invariant loads of globals.
 define void @test_fence6() {
 ; CHECK-LABEL: @test_fence6(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[I:%.*]] = call {}* @llvm.invariant.start.p0i8(i64 1, i8* @a)
+; CHECK-NEXT:    br label [[F:%.*]]
+; CHECK:       f:
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i8, i8* @a, align 1
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[TMP0]], 0
 ; CHECK-NEXT:    [[T:%.*]] = icmp eq i8 [[TMP1]], 0
-; CHECK-NEXT:    br label [[F:%.*]]
-; CHECK:       f:
 ; CHECK-NEXT:    tail call void @g(i1 [[T]])
 ; CHECK-NEXT:    br label [[F]]
 ;
 entry:
   %i = call {}* @llvm.invariant.start.p0i8(i64 1, i8* @a)
   br label %f
 
 f:
   %0 = load i8, i8* @a
   %1 = and i8 %0, 0
   %t = icmp eq i8 %1, 0
   tail call void @g(i1 %t)
   br label %f
 }