On September 18, 2018 5:47:34 AM GMT+02:00, Alexandre Oliva <ol...@adacore.com>
wrote:
>Richard, this is the patch you proposed in the PR a while ago, along
>with the testcase I'd posted. Please let me know if the commit-message
>paragraph below sounds good to you, and then I'll proceed to install
>it,
>or amend as requested. (do we really want the MEM_REF to carry
>stricter
>alignment than the base type? I'd have retained the object's alignment
>only for looser-than-type alignment)
I think so. We want to retain the alignment we'd extracted from the original
ref. As you said the indirection is spuirously introduced by the compiler.
>
>Building an ADDR_EXPR uses the canonical type to build the pointer
>type, but then, as we dereference it, we lose track of lax alignment
>known to apply to the dereferenced object. This might not be a
>problem in general, but it is when the compiler implicitly introduces
>address taking and dereferencing, as it does for asm statements, and
>as it may do in some loop optimizations.
>
>Regstrapped on x86_64-linux-gnu.
OK.
Richard.
>From: Richard Biener <rguent...@suse.de>
>for gcc/ChangeLog
>
> PR middle-end/87054
> * gimplify.c (gimplify_expr): Retain alignment of
> addressable lvalue in dereference.
>
>From: Alexandre Oliva <ol...@adacore.com>
>for gcc/testsuite/ChangeLog
>
> PR middle-end/87054
> * gcc.dg/pr87054.c: New.
>---
> gcc/gimplify.c | 8 +++++++-
> gcc/testsuite/gcc.dg/pr87054.c | 29 +++++++++++++++++++++++++++++
> 2 files changed, 36 insertions(+), 1 deletion(-)
> create mode 100644 gcc/testsuite/gcc.dg/pr87054.c
>
>diff --git a/gcc/gimplify.c b/gcc/gimplify.c
>index f0eb04a751ccc..509fc2f3f5be2 100644
>--- a/gcc/gimplify.c
>+++ b/gcc/gimplify.c
>@@ -12538,9 +12538,15 @@ gimplify_expr (tree *expr_p, gimple_seq
>*pre_p, gimple_seq *post_p,
> /* An lvalue will do. Take the address of the expression, store it
> in a temporary, and replace the expression with an INDIRECT_REF of
> that temporary. */
>+ tree ref_alias_type = reference_alias_ptr_type (*expr_p);
>+ unsigned int ref_align = get_object_alignment (*expr_p);
>+ tree ref_type = TREE_TYPE (*expr_p);
> tmp = build_fold_addr_expr_loc (input_location, *expr_p);
> gimplify_expr (&tmp, pre_p, post_p, is_gimple_reg, fb_rvalue);
>- *expr_p = build_simple_mem_ref (tmp);
>+ if (TYPE_ALIGN (ref_type) != ref_align)
>+ ref_type = build_aligned_type (ref_type, ref_align);
>+ *expr_p = build2 (MEM_REF, ref_type,
>+ tmp, build_zero_cst (ref_alias_type));
> }
>else if ((fallback & fb_rvalue) && is_gimple_reg_rhs_or_call (*expr_p))
> {
>diff --git a/gcc/testsuite/gcc.dg/pr87054.c
>b/gcc/testsuite/gcc.dg/pr87054.c
>new file mode 100644
>index 0000000000000..4ca2b62d2c7c9
>--- /dev/null
>+++ b/gcc/testsuite/gcc.dg/pr87054.c
>@@ -0,0 +1,29 @@
>+// { dg-do run }
>+// { dg-options "-O2" }
>+
>+#ifndef T
>+# ifdef __SSE__
>+# define T __int128
>+# else
>+# define T long
>+# endif
>+#endif
>+#ifndef R
>+# ifdef __SSE__
>+# define R "x"
>+# else
>+# define R "r"
>+# endif
>+#endif
>+
>+
>+typedef T A; // #define T to long or __int128
>+struct B { char d; A c; } __attribute__((packed));
>+struct B b[50]; // many elements to avoid loop unrolling
>+
>+int main () {
>+ int i;
>+ for (i = 0; i < sizeof(b) / sizeof(*b); i++) {
>+ asm ("" : "+" R (b[i].c)); // #define R to "r" on ppc or "x" on
>x86_64
>+ }
>+}
