Problem about memcpy elimination

IR & Optimizations
1

Hello everybody, I have some questions about the optimization of the following code

struct T {
    long a, b;
};
__attribute__((noinline))long test1(const T *t) {
    long x = t->a + t->b;
    return x;
}
__attribute__((noinline))void func(T &t) {
    T t1 = t;
    T t2 = t;
    long x1 = test1(&t1);
    long x2 = test1(&t2);
    std::cout << x1 + x2 << "\n";
}

with clang++ -O2, the IR of func is as follows

define dso_local void @func(T&)(ptr nocapture noundef nonnull readonly align 8 dereferenceable(16) %t) local_unnamed_addr #2 {
entry:
  %t1 = alloca %struct.T, align 8, !DIAssignID !894
  %t2 = alloca %struct.T, align 8, !DIAssignID !896
  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %t1) #8
  call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 8 dereferenceable(16) %t1, ptr noundef nonnull align 8 dereferenceable(16) %t, i64 16, i1 false), !tbaa.struct !899, !DIAssignID !901
  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %t2) #8
  call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 8 dereferenceable(16) %t2, ptr noundef nonnull align 8 dereferenceable(16) %t, i64 16, i1 false), !tbaa.struct !899, !DIAssignID !904
  %call = call noundef i64 @test1(T const*)(ptr noundef nonnull %t1)
  %call1 = call noundef i64 @test1(T const*)(ptr noundef nonnull %t2)
  %add = add nsw i64 %call1, %call
  ...
  ret void
}

the attributes of test1 is as follows

attributes #0 = { mustprogress nofree noinline norecurse nosync nounwind willreturn memory(argmem: read) uwtable "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }

why not just use %t as the argument to the call test1 in func, knowing that test1 has no side effects and doesn’t change its arguments

Does anyone know why it’s not optimized like this, thank you very much

2

I believe this would usually be handled by MemCpyOptPass::processImmutArgument(), but it specifically looks for a readonly noalias nocapture argument. The noalias part is missing in your case. You can see that the optimization works if you add a __restrict qualifier to the argument.

However, as the function is memory(argmem: read), I think it would still be valid to perform the optimization even without noalias, so it should be easy to extend support to this case.

1 Like
3

I tried to add __restrict to the test1 parameter. This worked. The %t1 and %t2 argument of call test1 were successfully replaced with %t by MemCpyOpt. Thanks.

4

[MemCpyOpt] Allow memcpy elision for non-noalias arguments by nikic · Pull Request #107860 · llvm/llvm-project · GitHub makes the transform work without noalias.

1 Like