Conversation
…leteConstantOffsets check fails
clang/lib/Sema/SemaBounds.cpp
Outdated
| Expr *LowerOffsetVariable; | ||
| Expr *UpperOffsetVariable; | ||
| bool LowerConstantIncomplete; | ||
| bool UpperConstantIncomplete; |
There was a problem hiding this comment.
It may suffice to have just one boolean field IsConstantIncomplete.
In CreateBaseRange, we get LowerKind (returned by SplitIntoBaseAndOffset on Lower) and UpperKind (returned by SplitIntoBaseAndOffset on Upper) Then, IsConstantIncomplete is set to true if either LowerKind or UpperKind is IncompleteConstantSized.
If we do this, the function CheckIncompleteConstantOffsets will not be required (this is causing the function EqualTypes to be called twice unnecessarily). Then we can directly call CheckIncompleteOffset.
| int f34(_Array_ptr<struct S> p : bounds(p, p), _Array_ptr<struct S> q : count(0)) { // expected-note {{(expanded) declared bounds are 'bounds(p, p)'}} | ||
| // The upper bound q + 0 is an incomplete constant sized upper offset, | ||
| // but the upper bound p is not an incomplete constant sized upper offset. | ||
| p = q; // expected-warning {{cannot prove declared bounds for 'p' are valid after assignment}} \ |
There was a problem hiding this comment.
I think the expected warning on line 310 in this test case is not really correct. The lower/upper bounds of p are not marked as Lower/UpperConstantIncomplete because of the way these fields are set. Currently, these fields are set only when the call to GetReferentSizeInChars fails. In the case of the bounds expression (p, p), I guess the control is going to the exit label in SplitIntoBaseAndOffset where the lower and upper bounds are split as p + 0 without calling GetReferentSizeInChars, and therefore the incompleteness of the type of p is not recorded.
I suggest the following:
Try checking if the type of Base is incomplete in the code under the exit label in SplitIntoBaseAndOffset. If the type is incomplete, return Kind::IncompleteConstantSized. If this gets further complicated or causes more issues/failures, file an issue for this test case and adding a TODO here with a reference to the filed issue.
Ideally, a BaseRange object must be marked as having an incomplete type if the type of the base (in this case p) is incomplete.
…mplete referent type
…into base-range-incomplete-type
…th HasIncompleteConstant
3 participants
See issue #1107
This PR modifies
BaseRangeto be able to compare constant-sized ranges whose base expressions have the same incomplete type.Consider source bounds
bounds(p, p + a)and target boundsbounds(q, q + b), whereaandbare constants. Ifpandqhave typesT1andT2, whereT1andT2are incomplete types, then we cannot normalizeaandbby multiplying them by the size of the referent types ofT1andT2. However, we can prove thatbounds(p, p + a)implybounds(q, q + b)ifT1 == T2anda >= b.