[llvm-branch-commits] [clang] [openmp] [Clang][OpenMP][Tile] Allow non-constant tile sizes. (PR #91345)

[Alexey Bataev via llvm-branch-commits]

================
@@ -17432,16 +17457,54 @@ OMPClause *SemaOpenMP::ActOnOpenMPSizesClause(ArrayRef<Expr *> SizeExprs,
                                               SourceLocation StartLoc,
                                               SourceLocation LParenLoc,
                                               SourceLocation EndLoc) {
-  for (Expr *SizeExpr : SizeExprs) {
-    ExprResult NumForLoopsResult = VerifyPositiveIntegerConstantInClause(
-        SizeExpr, OMPC_sizes, /*StrictlyPositive=*/true);
-    if (!NumForLoopsResult.isUsable())
-      return nullptr;
+  SmallVector<Expr *> SanitizedSizeExprs;
+  llvm::append_range(SanitizedSizeExprs, SizeExprs);
+
+  for (Expr *&SizeExpr : SanitizedSizeExprs) {
+    // Skip if already sanitized, e.g. during a partial template instantiation.
+    if (!SizeExpr)
+      continue;
+
+    bool IsValid = isNonNegativeIntegerValue(SizeExpr, SemaRef, OMPC_sizes,
+                                             /*StrictlyPositive=*/true);
+
+    // isNonNegativeIntegerValue returns true for non-integral types (but still
+    // emits error diagnostic), so check for the expected type explicitly.
+    QualType SizeTy = SizeExpr->getType();
+    if (!SizeTy->isIntegerType())
+      IsValid = false;
+
+    // Handling in templates is tricky. There are four possibilities to
+    // consider:
+    //
+    // 1a. The expression is valid and we are in a instantiated template or not
+    //     in a template:
+    //       Pass valid expression to be further analysed later in Sema.
+    // 1b. The expression is valid and we are in a template (including partial
+    //     instantiation):
+    //       isNonNegativeIntegerValue skipped any checks so there is no
+    //       guarantee it will be correct after instantiation.
+    //       ActOnOpenMPSizesClause will be called again at instantiation when
+    //       it is not in a dependent context anymore. This may cause warnings
+    //       to be emitted multiple times.
+    // 2a. The expression is invalid and we are in an instantiated template or
+    //     not in a template:
+    //       Invalidate the expression with a clearly wrong value (nullptr) so
+    //       later in Sema we do not have to do the same validity analysis again
+    //       or crash from unexpected data. Error diagnostics have already been
+    //       emitted.
+    // 2b. The expression is invalid and we are in a template (including partial
+    //     instantiation):
+    //       Pass the invalid expression as-is, template instantiation may
+    //       replace unexpected types/values with valid ones. The directives
+    //       with this clause must not try to use these expressions in dependent
+    //       contexts.
----------------
alexey-bataev wrote:

Sure, delaying is a good approach. If the analysis is delayed till actual instantiation, it does not mean that the templates are not handled.

https://github.com/llvm/llvm-project/pull/91345