Fix layout flags for types that have oddly sized primitive type fields (#58435)

This is caused because for LLVMs sake we have to say that the oddly
typed field is smaller than it actually is. (I wonder if we could
represent it as an iN field in a struct and have it work but the result
would be the same for now)

Fix #58434, fix #49318, close #49362.

---------

Co-authored-by: Mosè Giordano <[email protected]>
Co-authored-by: Sukera <[email protected]>
(cherry picked from commit 1b0b028)

Author: 3 people

src/datatype.c

@@ -702,13 +702,19 @@ void jl_compute_field_offsets(jl_datatype_t *st)
                     // Should never happen
                     throw_ovf(should_malloc, desc, st, fsz);
                 desc[i].isptr = 0;
+
                 if (jl_is_uniontype(fld)) {
                     fsz += 1; // selector byte
                     zeroinit = 1;
                     // TODO: Some unions could be bits comparable.
                     isbitsegal = 0;
                 }
                 else {
+                    if (fsz > jl_datatype_size(fld)) {
+                        // We have to pad the size to integer size class, but it means this has some padding
+                        isbitsegal = 0;
+                        haspadding = 1;
+                    }
                     uint32_t fld_npointers = ((jl_datatype_t*)fld)->layout->npointers;
                     if (((jl_datatype_t*)fld)->layout->flags.haspadding)
                         haspadding = 1;

test/core.jl

@@ -8518,3 +8518,9 @@ module GlobalBindingMulti
     using .M.C
 end
 @test GlobalBindingMulti.S === GlobalBindingMulti.M.C.S
+
+#58434 bitsegal comparison of oddly sized fields
+primitive type ByteString58434 (18 * 8) end
+
+@test Base.datatype_isbitsegal(Tuple{ByteString58434}) == false
+@test Base.datatype_haspadding(Tuple{ByteString58434}) == (length(Base.padding(Tuple{ByteString58434})) > 0)