Hi,
I don't think this is the right fix.
On that file, lines 1512-1515, and after calling isAllocSiteRemovable(), the
objectsize intrinsic is replaced with the "don't know" value. Which is
fine, semantics wise, but we can do better as you expect.
I think that at that point we could try to infer the size of the object, and
only resort to the don't know value if we fail. Take a look at the
getObjectSize() function, and InstCombineCalls.cpp for an example usage.
Nuno
P.S.: BTW, this patch is not for clang, so please send the new version to
llvm-commits.
----- Original Message -----
From: "MAYUR PANDEY" <[email protected]>
To: <[email protected]>
Sent: Tuesday, March 18, 2014 8:19 AM
Subject: [PATCH] Wrong output from built-in size function
Hi,
Please find attached the fix for wrong output from built-in size function.
The Testcase attached is a standard gcc testsuite TC.
extern void abort (void);
extern inline __attribute__ ((__always_inline__)) int
foo (char *dest)
{
return __builtin_object_size (dest, 1);
}
struct S
{
union
{
struct { int a, b; char c, d; } f;
struct { struct { int a, b; char c, d[255]; } e; } g;
} u;
};
int
main (void)
{
struct S s;
if (foo (s.u.g.e.d) != 255)
abort ();
return 0;
}
In the TC above the abort function is called as clang is unable to return
correct output on calling built-in size function. the reason seems to be
that in Instruction Combination pass for optimization,
Clang was not able to properly handle built-in size function as per new
DFS implementation. So Moving it back to follow iterative way specifically
for built-in size function.
Please Review.
Thanks,
Mayur
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
