Fix AAResults::callCapturesBefore for operand bundles

[carnaval]

Summary

AAResults::callCapturesBefore would previously ignore operand bundles. It was possible for a later instruction to miss its memory dependency on a call site that would only access the pointer through a bundle.

[sanjoy]

Can you please add a line or two in the commit message describing the problem and the fix?

test/CodeGen/X86/statepoint-memdep.ll

1 ↗

(On Diff #60514)

You don't need the 2>&1.

5 ↗

(On Diff #60514)

Add a CHECK-LABEL: @f( to make it easier to add new tests later.

[carnaval]

address comments

[sanjoy]

lgtm, do you want me to commit this for you?

[carnaval]

yep, I don't have commit access. Thanks for the review !

llvm/trunk/lib/Analysis/AliasAnalysis.cpp

ModRefInfo AAResults::callCapturesBefore(const Instruction *I,
   if (llvm::PointerMayBeCapturedBefore(Object, /* ReturnCaptures */ true,
                                        /* StoreCaptures */ true, I, DT,
                                        /* include Object */ true,
                                        /* OrderedBasicBlock */ OBB))
     return MRI_ModRef;
 
   unsigned ArgNo = 0;
   ModRefInfo R = MRI_NoModRef;
-  for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
+  for (auto CI = CS.data_operands_begin(), CE = CS.data_operands_end();
        CI != CE; ++CI, ++ArgNo) {
     // Only look at the no-capture or byval pointer arguments.  If this
     // pointer were passed to arguments that were neither of these, then it
     // couldn't be no-capture.
     if (!(*CI)->getType()->isPointerTy() ||
         (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
       continue;
 

llvm/trunk/test/CodeGen/X86/statepoint-memdep.ll

+; RUN: opt -S -dse < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @f() {
+  ; CHECK-LABEL: @f(
+  %s = alloca i64
+  ; Verify that this first store is not considered killed by the second one
+  ; since it could be observed from the deopt continuation.
+  ; CHECK: store i64 1, i64* %s
+  store i64 1, i64* %s
+  call void @g() [ "deopt"(i64* %s) ]
+  store i64 0, i64* %s
+  ret void
+}
+
+declare void @g()