The aarch64 PCS rules ignore user alignment for scalars and
vectors and use the "natural" alignment of the type. GCC tried
to calculate that natural alignment using:
TYPE_ALIGN (TYPE_MAIN_VARIANT (type))
But as discussed in the PR, it's possible that the main variant
of a pointer type is an overaligned type (although that's usually
accidental).
This isn't known to be a problem for other types, so this patch
changes the bare minimum. It might be that we need to ignore
TYPE_ALIGN in other cases too.
gcc/
PR target/108910
* config/aarch64/aarch64.cc (aarch64_function_arg_alignment): Do
not trust TYPE_ALIGN for pointer types; use POINTER_SIZE instead.
gcc/testsuite/
PR target/108910
* gcc.dg/torture/pr108910.c: New test.
(cherry picked from commit
66946624b96b762985de56444d726a0ebd4e0df5)
gcc_assert (TYPE_MODE (type) == mode);
if (!AGGREGATE_TYPE_P (type))
- return TYPE_ALIGN (TYPE_MAIN_VARIANT (type));
+ {
+ /* The ABI alignment is the natural alignment of the type, without
+ any attributes applied. Normally this is the alignment of the
+ TYPE_MAIN_VARIANT, but not always; see PR108910 for a counterexample.
+ For now we just handle the known exceptions explicitly. */
+ type = TYPE_MAIN_VARIANT (type);
+ if (POINTER_TYPE_P (type))
+ {
+ gcc_assert (known_eq (POINTER_SIZE, GET_MODE_BITSIZE (mode)));
+ return POINTER_SIZE;
+ }
+ return TYPE_ALIGN (type);
+ }
if (TREE_CODE (type) == ARRAY_TYPE)
return TYPE_ALIGN (TREE_TYPE (type));
--- /dev/null
+extern void foo (float, float *, float *);
+
+void
+bar (void *p)
+{
+ float *__attribute__((aligned (64))) q = __builtin_assume_aligned (p, 64);
+ foo (0.0f, q, q);
+}