Do not cache AliasSetTracker even in the old PM

[sanjoy]

Summary

Fixes PR32129. Test case contributed by Igor Laevsky.

I can't easily think of other ways to safely cache the ASTs. Given
we're going to move to recomputing them in the new PM anyway (unless
there are future plans to not to?), I think it is uncontroversial to
bring the old PM in sync.

If this change is accepted, I'll also delete the code that keeps track
of the Loop->AST mapping. But first I want to know if this change is
acceptable or not.

[hfinkel]

We're only caching to save on compile time, right? Is there any noticeable compile-time impact?

[sanjoy]

In D30726#695160, @hfinkel wrote:

We're only caching to save on compile time, right? Is there any noticeable compile-time impact?

I'll measure.

[chandlerc]

In D30726#695172, @sanjoy wrote:

In D30726#695160, @hfinkel wrote:

We're only caching to save on compile time, right? Is there any noticeable compile-time impact?

I'll measure.

I think this is the critical thing.

My expectation had been to have to add caching back in the new PM to address compile time issues. I think we need to understand how much this impacts compile time.

Certainly, not caching things seems much easier to get correct. ;]

[• dberlin]

We definitely have testcases where ASTs (in LICM and elsewhere) take ... forever (hundreds of seconds).

If this makes that worse, it'd be bad, but Alina's working on moving LICM to MemorySSA anyway (and i expect will continue when she gets back).
So if it's a temporary hit, and we need it for correctness, ....

[sanjoy]

Fix implemented as D30849

lib/Transforms/Scalar/LICM.cpp

bool runOnLoop(Loop *L, LPPassManager &LPM) override {
     // pass.  Function analyses need to be preserved across loop transformations
     // but ORE cannot be preserved (see comment before the pass definition).
     OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
     return LICM.runOnLoop(L,
                           &getAnalysis<AAResultsWrapperPass>().getAAResults(),
                           &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
                           &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
                           &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
-                          SE ? &SE->getSE() : nullptr, &ORE, false);
+                          SE ? &SE->getSE() : nullptr, &ORE, true);
   }
 
   /// This transformation requires natural loop information & requires that
   /// loop preheaders be inserted into the CFG...
   ///
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
     AU.addRequired<TargetLibraryInfoWrapperPass>();

test/Transforms/LICM/pr32129.ll

+; RUN: opt -S -licm -loop-unswitch -licm < %s | FileCheck %s
+
+declare void @llvm.experimental.guard(i1, ...)
+
+define void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NOT: guard
+entry:
+  br label %header
+
+header:
+  br label %loop
+
+loop:
+  %0 = icmp ult i32 0, 400
+  call void (i1, ...) @llvm.experimental.guard(i1 %0, i32 9) [ "deopt"() ]
+  br i1 undef, label %header, label %loop
+}