Re: [PATCH] Fix expansion with COMPOUND_LITERAL_EXPR (PR middle-end/64067)
On Thu, Nov 27, 2014 at 11:03:20AM +0100, Richard Biener wrote: > > The testcase shows that when expanding ARRAY_REF from a const static var > > with initializer that contains COMPOUND_LITERAL_EXPR, we can try to expand > > COMPOUND_LITERAL_EXPR even when modifier is not EXPAND_INITIALIZER (it is > > EXPAND_SUM in that testcase, but could be many others). > > While gimplification should get rid of all the compound literal exprs > > for automatic compound literals, COMPOUND_LITERAL_EXPR for static compound > > literals can survive in the initializers, thus this patch handles them > > always. > > > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > Ok. (I wonder if the fallthru still makes sense or if we should > hit a gcc_unreachable () for other uses of COMPOUND_LITERAL_EXPR). I think the fallthru can't work, because it will hit the gcc_unreachable (); on unhandled COMPOUND_LITERAL_EXPR in expand_expr_real_*. So changing the if in the patch into gcc_assert and removing /* FALLTHRU */ might be ok too (and removing == COMPOUND_LITERAL_EXPR from the fallthru if). > > 2014-11-26 Jakub Jelinek > > > > PR middle-end/64067 > > * expr.c (expand_expr_addr_expr_1) : > > Handle it by returning address of COMPOUND_LITERAL_EXPR_DECL > > not only if modifier is EXPAND_INITIALIZER, but whenever > > COMPOUND_LITERAL_EXPR_DECL is non-NULL and TREE_STATIC. > > > > * gcc.c-torture/compile/pr64067.c: New test. > > > > --- gcc/expr.c.jj 2014-11-18 08:26:45.0 +0100 > > +++ gcc/expr.c 2014-11-26 17:38:52.877675088 +0100 > > @@ -7677,11 +7677,13 @@ expand_expr_addr_expr_1 (tree exp, rtx t > >break; > > > > case COMPOUND_LITERAL_EXPR: > > - /* Allow COMPOUND_LITERAL_EXPR in initializers, if e.g. > > -rtl_for_decl_init is called on DECL_INITIAL with > > -COMPOUNT_LITERAL_EXPRs in it, they aren't gimplified. */ > > - if (modifier == EXPAND_INITIALIZER > > - && COMPOUND_LITERAL_EXPR_DECL (exp)) > > + /* Allow COMPOUND_LITERAL_EXPR in initializers or coming from > > +initializers, if e.g. rtl_for_decl_init is called on DECL_INITIAL > > +with COMPOUND_LITERAL_EXPRs in it, or ARRAY_REF on a const static > > +array with address of COMPOUND_LITERAL_EXPR in DECL_INITIAL; > > +the initializers aren't gimplified. */ > > + if (COMPOUND_LITERAL_EXPR_DECL (exp) > > + && TREE_STATIC (COMPOUND_LITERAL_EXPR_DECL (exp))) > > return expand_expr_addr_expr_1 (COMPOUND_LITERAL_EXPR_DECL (exp), > > target, tmode, modifier, as); > >/* FALLTHRU */ Jakub
Re: [PATCH] Fix expansion with COMPOUND_LITERAL_EXPR (PR middle-end/64067)
On Wed, 26 Nov 2014, Jakub Jelinek wrote:
> Hi!
>
> The testcase shows that when expanding ARRAY_REF from a const static var
> with initializer that contains COMPOUND_LITERAL_EXPR, we can try to expand
> COMPOUND_LITERAL_EXPR even when modifier is not EXPAND_INITIALIZER (it is
> EXPAND_SUM in that testcase, but could be many others).
> While gimplification should get rid of all the compound literal exprs
> for automatic compound literals, COMPOUND_LITERAL_EXPR for static compound
> literals can survive in the initializers, thus this patch handles them
> always.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok. (I wonder if the fallthru still makes sense or if we should
hit a gcc_unreachable () for other uses of COMPOUND_LITERAL_EXPR).
Thanks,
Richard.
> 2014-11-26 Jakub Jelinek
>
> PR middle-end/64067
> * expr.c (expand_expr_addr_expr_1) :
> Handle it by returning address of COMPOUND_LITERAL_EXPR_DECL
> not only if modifier is EXPAND_INITIALIZER, but whenever
> COMPOUND_LITERAL_EXPR_DECL is non-NULL and TREE_STATIC.
>
> * gcc.c-torture/compile/pr64067.c: New test.
>
> --- gcc/expr.c.jj 2014-11-18 08:26:45.0 +0100
> +++ gcc/expr.c2014-11-26 17:38:52.877675088 +0100
> @@ -7677,11 +7677,13 @@ expand_expr_addr_expr_1 (tree exp, rtx t
>break;
>
> case COMPOUND_LITERAL_EXPR:
> - /* Allow COMPOUND_LITERAL_EXPR in initializers, if e.g.
> - rtl_for_decl_init is called on DECL_INITIAL with
> - COMPOUNT_LITERAL_EXPRs in it, they aren't gimplified. */
> - if (modifier == EXPAND_INITIALIZER
> - && COMPOUND_LITERAL_EXPR_DECL (exp))
> + /* Allow COMPOUND_LITERAL_EXPR in initializers or coming from
> + initializers, if e.g. rtl_for_decl_init is called on DECL_INITIAL
> + with COMPOUND_LITERAL_EXPRs in it, or ARRAY_REF on a const static
> + array with address of COMPOUND_LITERAL_EXPR in DECL_INITIAL;
> + the initializers aren't gimplified. */
> + if (COMPOUND_LITERAL_EXPR_DECL (exp)
> + && TREE_STATIC (COMPOUND_LITERAL_EXPR_DECL (exp)))
> return expand_expr_addr_expr_1 (COMPOUND_LITERAL_EXPR_DECL (exp),
> target, tmode, modifier, as);
>/* FALLTHRU */
> --- gcc/testsuite/gcc.c-torture/compile/pr64067.c.jj 2014-11-26
> 17:37:37.835865797 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr64067.c 2014-11-26
> 17:37:02.0 +0100
> @@ -0,0 +1,10 @@
> +/* PR middle-end/64067 */
> +
> +struct S { int s; };
> +int *const v[1] = { &((struct S) { .s = 42 }).s };
> +
> +int *
> +foo (void)
> +{
> + return v[0];
> +}
>
> Jakub
>
>
--
Richard Biener
SUSE LINUX GmbH, GF: Jeff Hawn, Jennifer Guild, Felix Imendoerffer, HRB 21284
(AG Nuernberg)
Maxfeldstrasse 5, 90409 Nuernberg, Germany
[PATCH] Fix expansion with COMPOUND_LITERAL_EXPR (PR middle-end/64067)
Hi!
The testcase shows that when expanding ARRAY_REF from a const static var
with initializer that contains COMPOUND_LITERAL_EXPR, we can try to expand
COMPOUND_LITERAL_EXPR even when modifier is not EXPAND_INITIALIZER (it is
EXPAND_SUM in that testcase, but could be many others).
While gimplification should get rid of all the compound literal exprs
for automatic compound literals, COMPOUND_LITERAL_EXPR for static compound
literals can survive in the initializers, thus this patch handles them
always.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2014-11-26 Jakub Jelinek
PR middle-end/64067
* expr.c (expand_expr_addr_expr_1) :
Handle it by returning address of COMPOUND_LITERAL_EXPR_DECL
not only if modifier is EXPAND_INITIALIZER, but whenever
COMPOUND_LITERAL_EXPR_DECL is non-NULL and TREE_STATIC.
* gcc.c-torture/compile/pr64067.c: New test.
--- gcc/expr.c.jj 2014-11-18 08:26:45.0 +0100
+++ gcc/expr.c 2014-11-26 17:38:52.877675088 +0100
@@ -7677,11 +7677,13 @@ expand_expr_addr_expr_1 (tree exp, rtx t
break;
case COMPOUND_LITERAL_EXPR:
- /* Allow COMPOUND_LITERAL_EXPR in initializers, if e.g.
-rtl_for_decl_init is called on DECL_INITIAL with
-COMPOUNT_LITERAL_EXPRs in it, they aren't gimplified. */
- if (modifier == EXPAND_INITIALIZER
- && COMPOUND_LITERAL_EXPR_DECL (exp))
+ /* Allow COMPOUND_LITERAL_EXPR in initializers or coming from
+initializers, if e.g. rtl_for_decl_init is called on DECL_INITIAL
+with COMPOUND_LITERAL_EXPRs in it, or ARRAY_REF on a const static
+array with address of COMPOUND_LITERAL_EXPR in DECL_INITIAL;
+the initializers aren't gimplified. */
+ if (COMPOUND_LITERAL_EXPR_DECL (exp)
+ && TREE_STATIC (COMPOUND_LITERAL_EXPR_DECL (exp)))
return expand_expr_addr_expr_1 (COMPOUND_LITERAL_EXPR_DECL (exp),
target, tmode, modifier, as);
/* FALLTHRU */
--- gcc/testsuite/gcc.c-torture/compile/pr64067.c.jj2014-11-26
17:37:37.835865797 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr64067.c 2014-11-26
17:37:02.0 +0100
@@ -0,0 +1,10 @@
+/* PR middle-end/64067 */
+
+struct S { int s; };
+int *const v[1] = { &((struct S) { .s = 42 }).s };
+
+int *
+foo (void)
+{
+ return v[0];
+}
Jakub
