Properly update AA metadata when performing call slot optimization

[hfinkel]

LGTM.

lib/Transforms/Scalar/MemCpyOptimizer.cpp

754 ↗

(On Diff #19532)

Please at least add a FIXME here about properly combining MD_tbaa_struct and MD_mem_parallel_loop_access.

[dotdash]

Added a FIXME for MD_tbaa_struct and MD_mem_parallel_loop_access

llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp

if (!isDestSufficientlyAligned) {
     assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
     cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
   }
 
   // Drop any cached information about the call, because we may have changed
   // its dependence information by changing its parameter.
   MD->removeInstruction(C);
 
+  // Update AA metadata
+  // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
+  // handled here, but combineMetadata doesn't support them yet
+  unsigned KnownIDs[] = {
+    LLVMContext::MD_tbaa,
+    LLVMContext::MD_alias_scope,
+    LLVMContext::MD_noalias,
+  };
+  combineMetadata(C, cpy, KnownIDs);
+
   // Remove the memcpy.
   MD->removeInstruction(cpy);
   ++NumMemCpyInstr;
 
   return true;
 }
 
 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)

llvm/trunk/test/Transforms/MemCpyOpt/callslot_aa.ll

+; RUN: opt < %s -S -basicaa -memcpyopt | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+%T = type { i64, i64 }
+
+define void @test(i8* %src) {
+  %tmp = alloca i8
+  %dst = alloca i8
+; CHECK:   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 1, i32 8, i1 false)
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp, i8* %src, i64 1, i32 8, i1 false), !noalias !2
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %tmp, i64 1, i32 8, i1 false)
+
+  ret void
+}
+
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8*, i8*, i64, i32, i1)
+
+; Check that the noalias for "dst" was removed by checking that the metadata is gone
+; CHECK-NOT: "dst"
+!0 = !{!0}
+!1 = distinct !{!1, !0, !"dst"}
+!2 = distinct !{!1}