Re: [PING][PATCH 2/4] remove %G and %K from calls in front end and middle end (PR 98512)

[Martin Sebor <msebor@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2016 bytes --]

On 7/2/21 2:52 PM, David Malcolm wrote:
...
>>>> @@ -11425,10 +11425,10 @@ expand_expr_real_1 (tree exp, rtx
>>>> target, machine_mode tmode,
>>>>                                           DECL_ATTRIBUTES
>>>> (fndecl))) != NULL)
>>>>            {
>>>>              const char *ident = lang_hooks.decl_printable_name
>>>> (fndecl, 1);
>>>> -           warning_at (tree_nonartificial_location (exp),
>>>> +           warning_at (EXPR_LOCATION (exp),
>>>
>>> Are we preserving the existing behavior for
>>> __attribute__((__artificial__)) here?
>>> Is this behavior handled somewhere else in the patch kit?
>>
>> Yes.  The warning infrastructure (set_inlining_locations) uses
>> the location of the site into which the statement has been inlined
>> regardless of whether the inlined function is artificial.
> 
> Do we have test coverage for this, though?
> 
> [...]
>> Attached is the revised patch for reference.  Since it just removes
>> the uses of the %K and %G directives made redundant by the first
>> patch in the series I'll go ahead and commit it as obvious in a day
>> or so after patch 1 unless someone has further questions or requests
>> for changes.
> 
> Please can you look into the "__artificial__" test coverage, and
> address the line number thing above.

There are a bunch of tests for it but I couldn't find one that
verifies the inlining stack.  I always struggle with the DejaGnu
directives for these things (I couldn't get dg-multiline-output
to do what I wanted) so I cribbed the approach from
plugin/diagnostic-test-inlining-*.c.

Since I was changing the directives I got rid of the line numbers
from their comments but instead hardcoded them in the search pattern,
making the tests tighter.  That's a good thing but it does of course
mean that when new lines are added to the tests the directives will
fail and the patterns will need to be updated.

Attached is a patch with just the test I'm adding.

Martin

[-- Attachment #2: Wfree-nonheap-object-6.c.diff --]
[-- Type: text/x-patch, Size: 1834 bytes --]

diff --git a/gcc/testsuite/gcc.dg/Wfree-nonheap-object-6.c b/gcc/testsuite/gcc.dg/Wfree-nonheap-object-6.c
new file mode 100644
index 00000000000..c109558838b
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/Wfree-nonheap-object-6.c
@@ -0,0 +1,49 @@
+/* Similar to Wfree-nonheap-object-5.c but with attribute artificial:
+   verify that warnings for the same call site from distinct callers
+   include the correct function names in the inlining stack.
+   { dg-do compile }
+   { dg-options "-O1 -Wall" } */
+
+struct A
+{
+  void *p;
+};
+
+__attribute__ ((always_inline, artificial))
+inline void f0 (struct A *p)
+{
+  __builtin_free (p->p);      // { dg-warning "\\\[-Wfree-nonheap-object" }
+}
+
+// Expect two instances of the text below:
+// { dg-regexp "In function 'f0'," "first f0 prefix" { target *-*-* } 0 }
+// { dg-regexp "In function 'f0'," "second f0 prefix" { target *-*-* } 0 }
+
+__attribute__ ((always_inline, artificial))
+inline void f1 (struct A *p) { f0 (p); }
+__attribute__ ((always_inline, artificial))
+inline void f2 (struct A *p) { f1 (p); }
+
+extern int array[];
+// Also expect two instances of the note:
+// { dg-regexp "declared here" "first note for array" { target *-*-* } .-2 }
+// { dg-regexp "declared here" "second note for array" { target *-*-* } .-3 }
+
+void foo (struct A *p)
+{
+  p->p = array + 1;
+  f0 (p);
+}
+
+// { dg-regexp " +inlined from 'foo' at \[^:\]+Wfree-nonheap-object-6.c:35:\\d+:" "inlined from foo" }
+
+
+void bar (struct A *p)
+{
+  p->p = array + 2;
+  f2 (p);
+}
+
+// { dg-regexp " +inlined from 'f1' at \[^:\]+Wfree-nonheap-object-6.c:23:\\d+," "inlined from f1" }
+// { dg-regexp " +inlined from 'f2' at \[^:\]+Wfree-nonheap-object-6.c:25:\\d+," "inlined from f2" }
+// { dg-regexp " +inlined from 'bar' at \[^:\]+Wfree-nonheap-object-6.c:44:\\d+:" "inlined from bar" }