Track TBAA metadata in jl_cgval_t

[yuyichao]

So that field load can be marked with the correct TBAA.

This fixes the issue @JeffBezanson noticed in #15402 (comment)

My repro of the issue now gives (yeah, it vectorizes).

define %jl_value_t* @"julia_collect_to2!_51446"(%jl_value_t*, %Gen*) #0 {
top:
  %2 = getelementptr inbounds %Gen, %Gen* %1, i64 0, i32 0, i32 0
  %3 = load i64, i64* %2, align 8
  %4 = getelementptr inbounds %Gen, %Gen* %1, i64 0, i32 0, i32 1
  %5 = load i64, i64* %4, align 8
  %6 = add i64 %5, 1
  %7 = icmp eq i64 %3, %6
  br i1 %7, label %L2, label %if.lr.ph

if.lr.ph:                                         ; preds = %top
  %8 = bitcast %jl_value_t* %0 to double**
  %9 = load double*, double** %8, align 8
  %10 = load i64, i64* %4, align 8
  %11 = sub i64 %10, %3
  %backedge.overflow = icmp eq i64 %11, -1
  %12 = add i64 %10, 1
  br i1 %backedge.overflow, label %scalar.ph, label %overflow.checked

overflow.checked:                                 ; preds = %if.lr.ph
  %13 = sub i64 %12, %3
  %n.vec = and i64 %13, -8
  %end.idx.rnd.down = add i64 %n.vec, %3
  %cmp.zero = icmp eq i64 %n.vec, 0
  %ind.end = or i64 %n.vec, 1
  br i1 %cmp.zero, label %middle.block, label %vector.ph

vector.ph:                                        ; preds = %overflow.checked
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %vector.ph
  %index = phi i64 [ %3, %vector.ph ], [ %index.next, %vector.body ]
  %normalized.idx = sub i64 %index, %3
  %broadcast.splatinsert8 = insertelement <4 x i64> undef, i64 %index, i32 0
  %broadcast.splat9 = shufflevector <4 x i64> %broadcast.splatinsert8, <4 x i64> undef, <4 x i32> zeroinitializer
  %induction10 = add <4 x i64> %broadcast.splat9, <i64 0, i64 1, i64 2, i64 3>
  %induction11 = add <4 x i64> %broadcast.splat9, <i64 4, i64 5, i64 6, i64 7>
  %14 = sitofp <4 x i64> %induction10 to <4 x double>
  %15 = sitofp <4 x i64> %induction11 to <4 x double>
  %16 = fdiv <4 x double> %14, <double 1.000000e+01, double 1.000000e+01, double 1.000000e+01, double 1.000000e+01>
  %17 = fdiv <4 x double> %15, <double 1.000000e+01, double 1.000000e+01, double 1.000000e+01, double 1.000000e+01>
  %18 = getelementptr double, double* %9, i64 %normalized.idx
  %19 = bitcast double* %18 to <4 x double>*
  store <4 x double> %16, <4 x double>* %19, align 8
  %20 = getelementptr double, double* %18, i64 4
  %21 = bitcast double* %20 to <4 x double>*
  store <4 x double> %17, <4 x double>* %21, align 8
  %index.next = add i64 %index, 8
  %22 = icmp eq i64 %index.next, %end.idx.rnd.down
  br i1 %22, label %middle.block, label %vector.body

middle.block:                                     ; preds = %vector.body, %overflow.checked
  %resume.val = phi i64 [ 1, %overflow.checked ], [ %ind.end, %vector.body ]
  %resume.val5 = phi i64 [ %3, %overflow.checked ], [ %end.idx.rnd.down, %vector.body ]
  %trunc.resume.val = phi i64 [ %3, %overflow.checked ], [ %end.idx.rnd.down, %vector.body ]
  %cmp.n = icmp eq i64 %12, %resume.val5
  br i1 %cmp.n, label %L2.loopexit, label %scalar.ph

scalar.ph:                                        ; preds = %middle.block, %if.lr.ph
  %bc.resume.val = phi i64 [ %resume.val, %middle.block ], [ 1, %if.lr.ph ]
  %bc.trunc.resume.val = phi i64 [ %trunc.resume.val, %middle.block ], [ %3, %if.lr.ph ]
  br label %if

L2.loopexit:                                      ; preds = %middle.block, %if
  br label %L2

L2:                                               ; preds = %L2.loopexit, %top
  ret %jl_value_t* %0

if:                                               ; preds = %scalar.ph, %if
  %i.04 = phi i64 [ %bc.resume.val, %scalar.ph ], [ %28, %if ]
  %st.03 = phi i64 [ %bc.trunc.resume.val, %scalar.ph ], [ %23, %if ]
  %23 = add i64 %st.03, 1
  %24 = sitofp i64 %st.03 to double
  %25 = fdiv double %24, 1.000000e+01
  %26 = add i64 %i.04, -1
  %27 = getelementptr double, double* %9, i64 %26
  store double %25, double* %27, align 8
  %28 = add i64 %i.04, 1
  %29 = icmp eq i64 %st.03, %10
  br i1 %29, label %L2.loopexit, label %if
}

Marked as WIP since I still want to double check if the tbaa assignment are all valid.

[yuyichao]

I took this chance to clean up the tbaa tree a little, add a few more nodes for memory accesses that we weren't marking before and delete a few nodes that are not used anymore. A summary of the difference,

1. (Maybe the most important one) array buffer is not allowed to alias jl_value_t* (why would you do that....)
2. Access of jl_datatype_t uses tbaa_const (I don't think we mutate it after type construction time which should be invisible to codegen)
3. Access of type tag are marked with tbaa_tag (It's still valid to unsafe_load it though, unsafe_store! is also valid from tbaa point of view but you'll run into other issues very quickly...)
4. Replace tbaa_user with tbaa_mutab for mutable types, both inherit from tbaa_value.

Local tests passed, (CI seems to be struggling with the long queue...)