* [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand
@ 2017-06-29 3:01 Peter Bergner
2017-06-29 8:34 ` Richard Biener
0 siblings, 1 reply; 4+ messages in thread
From: Peter Bergner @ 2017-06-29 3:01 UTC (permalink / raw)
To: GCC Patches; +Cc: Richard Biener, marc.mutz
With the fix to PR51513 and follow on fixes for PR80707, PR80775 and PR80823,
we can now end up with switch statements that contain nothing but a default
case statement. The expand_case() function contains code that assumes we
have at least one non-default case statement, leading to the ICE reported
in the PR81194.
This patch fixes the bug by expanding switch statements that contain only
a default case statement, as a GOTO to the default case's label.
This passed bootstrap and regtesting on x86_64-linux with no regressions.
Ok for trunk?
Peter
gcc/
PR middle-end/81194
* cfgexpand.c (expand_gimple_stmt_1): Handle switch statements
with only one label.
* stmt.c (expand_case): Assert NCASES is greater than one.
gcc/testsuite/
PR middle-end/81194
* g++.dg/pr81194.C: New test.
Index: gcc/cfgexpand.c
===================================================================
--- gcc/cfgexpand.c (revision 249747)
+++ gcc/cfgexpand.c (working copy)
@@ -3566,7 +3566,13 @@
case GIMPLE_PREDICT:
break;
case GIMPLE_SWITCH:
- expand_case (as_a <gswitch *> (stmt));
+ {
+ gswitch *swtch = as_a <gswitch *> (stmt);
+ if (gimple_switch_num_labels (swtch) == 1)
+ expand_goto (CASE_LABEL (gimple_switch_default_label (swtch)));
+ else
+ expand_case (swtch);
+ }
break;
case GIMPLE_ASM:
expand_asm_stmt (as_a <gasm *> (stmt));
Index: gcc/stmt.c
===================================================================
--- gcc/stmt.c (revision 249747)
+++ gcc/stmt.c (working copy)
@@ -1142,8 +1142,11 @@
/* cleanup_tree_cfg removes all SWITCH_EXPR with their index
expressions being INTEGER_CST. */
gcc_assert (TREE_CODE (index_expr) != INTEGER_CST);
-
+ /* Optimization of switch statements with only one label has already
+ occurred, so we should never see them at this point. */
+ gcc_assert (ncases > 1);
+
do_pending_stack_adjust ();
/* Find the default case target label. */
Index: gcc/testsuite/g++.dg/pr81194.C
===================================================================
--- gcc/testsuite/g++.dg/pr81194.C (nonexistent)
+++ gcc/testsuite/g++.dg/pr81194.C (working copy)
@@ -0,0 +1,60 @@
+// { dg-do compile }
+// { dg-options "-O2 -std=c++17 -fno-exceptions" }
+
+template <class a> struct b { typedef a *c; };
+class e {};
+template <typename a> class d {
+public:
+ typedef typename b<a>::c c;
+ c begin();
+ c end();
+};
+struct f {
+ enum { g } h;
+};
+struct i {
+ d<f *> j();
+};
+struct l {
+ d<i *> k();
+};
+class ac;
+class o {
+public:
+ o(int *, int *, int *, ac *);
+};
+class ac {
+public:
+ ac(e);
+ virtual o *ae(int *, int *, int *, int *);
+};
+class p {
+ void af(f *m) {
+ switch (m->h)
+ case f::g:
+ ag();
+ }
+
+public:
+ void n() {
+ l ah;
+ for (i *ai : ah.k())
+ for (f *m : ai->j())
+ af(m);
+ }
+ virtual void ag() { __builtin_unreachable(); }
+};
+template <typename = int> class an : o {
+public:
+ an(int *, int *, int *, int *, ac *);
+};
+class q : ac {
+public:
+ q() : ac([]() -> e {}()) {}
+ o *ae(int *ap, int *aq, int *ar, int *as) { an(ap, aq, ar, as, this); }
+};
+template <typename at>
+an<at>::an(int *, int *aq, int *ar, int *as, ac *au) : o(aq, ar, as, au) {
+ p().n();
+}
+void av() { new q; }
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand
2017-06-29 3:01 [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand Peter Bergner
@ 2017-06-29 8:34 ` Richard Biener
2017-06-29 8:51 ` Richard Biener
0 siblings, 1 reply; 4+ messages in thread
From: Richard Biener @ 2017-06-29 8:34 UTC (permalink / raw)
To: Peter Bergner; +Cc: GCC Patches, marc.mutz
On Thu, Jun 29, 2017 at 5:01 AM, Peter Bergner <bergner@vnet.ibm.com> wrote:
> With the fix to PR51513 and follow on fixes for PR80707, PR80775 and PR80823,
> we can now end up with switch statements that contain nothing but a default
> case statement. The expand_case() function contains code that assumes we
> have at least one non-default case statement, leading to the ICE reported
> in the PR81194.
>
> This patch fixes the bug by expanding switch statements that contain only
> a default case statement, as a GOTO to the default case's label.
>
> This passed bootstrap and regtesting on x86_64-linux with no regressions.
> Ok for trunk?
Ok.
Btw, I'm curious -- which pass ends up with just the default case? IMHO
it would be a cfgcleanup task but eventually we fail to run it? That is
cleanup_control_expr_graph calls find_taken_edge () and maybe that
doesn't handle the case where there's just the default label and thus
the value of the switch var doesn't matter ...
Thanks,
Richard.
> Peter
>
> gcc/
> PR middle-end/81194
> * cfgexpand.c (expand_gimple_stmt_1): Handle switch statements
> with only one label.
> * stmt.c (expand_case): Assert NCASES is greater than one.
>
> gcc/testsuite/
> PR middle-end/81194
> * g++.dg/pr81194.C: New test.
>
> Index: gcc/cfgexpand.c
> ===================================================================
> --- gcc/cfgexpand.c (revision 249747)
> +++ gcc/cfgexpand.c (working copy)
> @@ -3566,7 +3566,13 @@
> case GIMPLE_PREDICT:
> break;
> case GIMPLE_SWITCH:
> - expand_case (as_a <gswitch *> (stmt));
> + {
> + gswitch *swtch = as_a <gswitch *> (stmt);
> + if (gimple_switch_num_labels (swtch) == 1)
> + expand_goto (CASE_LABEL (gimple_switch_default_label (swtch)));
> + else
> + expand_case (swtch);
> + }
> break;
> case GIMPLE_ASM:
> expand_asm_stmt (as_a <gasm *> (stmt));
> Index: gcc/stmt.c
> ===================================================================
> --- gcc/stmt.c (revision 249747)
> +++ gcc/stmt.c (working copy)
> @@ -1142,8 +1142,11 @@
> /* cleanup_tree_cfg removes all SWITCH_EXPR with their index
> expressions being INTEGER_CST. */
> gcc_assert (TREE_CODE (index_expr) != INTEGER_CST);
> -
>
> + /* Optimization of switch statements with only one label has already
> + occurred, so we should never see them at this point. */
> + gcc_assert (ncases > 1);
> +
> do_pending_stack_adjust ();
>
> /* Find the default case target label. */
> Index: gcc/testsuite/g++.dg/pr81194.C
> ===================================================================
> --- gcc/testsuite/g++.dg/pr81194.C (nonexistent)
> +++ gcc/testsuite/g++.dg/pr81194.C (working copy)
> @@ -0,0 +1,60 @@
> +// { dg-do compile }
> +// { dg-options "-O2 -std=c++17 -fno-exceptions" }
> +
> +template <class a> struct b { typedef a *c; };
> +class e {};
> +template <typename a> class d {
> +public:
> + typedef typename b<a>::c c;
> + c begin();
> + c end();
> +};
> +struct f {
> + enum { g } h;
> +};
> +struct i {
> + d<f *> j();
> +};
> +struct l {
> + d<i *> k();
> +};
> +class ac;
> +class o {
> +public:
> + o(int *, int *, int *, ac *);
> +};
> +class ac {
> +public:
> + ac(e);
> + virtual o *ae(int *, int *, int *, int *);
> +};
> +class p {
> + void af(f *m) {
> + switch (m->h)
> + case f::g:
> + ag();
> + }
> +
> +public:
> + void n() {
> + l ah;
> + for (i *ai : ah.k())
> + for (f *m : ai->j())
> + af(m);
> + }
> + virtual void ag() { __builtin_unreachable(); }
> +};
> +template <typename = int> class an : o {
> +public:
> + an(int *, int *, int *, int *, ac *);
> +};
> +class q : ac {
> +public:
> + q() : ac([]() -> e {}()) {}
> + o *ae(int *ap, int *aq, int *ar, int *as) { an(ap, aq, ar, as, this); }
> +};
> +template <typename at>
> +an<at>::an(int *, int *aq, int *ar, int *as, ac *au) : o(aq, ar, as, au) {
> + p().n();
> +}
> +void av() { new q; }
>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand
2017-06-29 8:34 ` Richard Biener
@ 2017-06-29 8:51 ` Richard Biener
2017-06-29 13:00 ` Peter Bergner
0 siblings, 1 reply; 4+ messages in thread
From: Richard Biener @ 2017-06-29 8:51 UTC (permalink / raw)
To: Peter Bergner; +Cc: GCC Patches, marc.mutz
On Thu, Jun 29, 2017 at 10:34 AM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Thu, Jun 29, 2017 at 5:01 AM, Peter Bergner <bergner@vnet.ibm.com> wrote:
>> With the fix to PR51513 and follow on fixes for PR80707, PR80775 and PR80823,
>> we can now end up with switch statements that contain nothing but a default
>> case statement. The expand_case() function contains code that assumes we
>> have at least one non-default case statement, leading to the ICE reported
>> in the PR81194.
>>
>> This patch fixes the bug by expanding switch statements that contain only
>> a default case statement, as a GOTO to the default case's label.
>>
>> This passed bootstrap and regtesting on x86_64-linux with no regressions.
>> Ok for trunk?
>
> Ok.
>
> Btw, I'm curious -- which pass ends up with just the default case? IMHO
> it would be a cfgcleanup task but eventually we fail to run it? That is
> cleanup_control_expr_graph calls find_taken_edge () and maybe that
> doesn't handle the case where there's just the default label and thus
> the value of the switch var doesn't matter ...
To answer myself the unreachable case vanishes at
execute_cleanup_cfg_post_optimizing
via group_case_labels. find_taken_edge wouldn't handle this case either.
I am testing a patch fixing both - your patch should still go in.
Thanks,
Richard.
> Thanks,
> Richard.
>
>> Peter
>>
>> gcc/
>> PR middle-end/81194
>> * cfgexpand.c (expand_gimple_stmt_1): Handle switch statements
>> with only one label.
>> * stmt.c (expand_case): Assert NCASES is greater than one.
>>
>> gcc/testsuite/
>> PR middle-end/81194
>> * g++.dg/pr81194.C: New test.
>>
>> Index: gcc/cfgexpand.c
>> ===================================================================
>> --- gcc/cfgexpand.c (revision 249747)
>> +++ gcc/cfgexpand.c (working copy)
>> @@ -3566,7 +3566,13 @@
>> case GIMPLE_PREDICT:
>> break;
>> case GIMPLE_SWITCH:
>> - expand_case (as_a <gswitch *> (stmt));
>> + {
>> + gswitch *swtch = as_a <gswitch *> (stmt);
>> + if (gimple_switch_num_labels (swtch) == 1)
>> + expand_goto (CASE_LABEL (gimple_switch_default_label (swtch)));
>> + else
>> + expand_case (swtch);
>> + }
>> break;
>> case GIMPLE_ASM:
>> expand_asm_stmt (as_a <gasm *> (stmt));
>> Index: gcc/stmt.c
>> ===================================================================
>> --- gcc/stmt.c (revision 249747)
>> +++ gcc/stmt.c (working copy)
>> @@ -1142,8 +1142,11 @@
>> /* cleanup_tree_cfg removes all SWITCH_EXPR with their index
>> expressions being INTEGER_CST. */
>> gcc_assert (TREE_CODE (index_expr) != INTEGER_CST);
>> -
>>
>> + /* Optimization of switch statements with only one label has already
>> + occurred, so we should never see them at this point. */
>> + gcc_assert (ncases > 1);
>> +
>> do_pending_stack_adjust ();
>>
>> /* Find the default case target label. */
>> Index: gcc/testsuite/g++.dg/pr81194.C
>> ===================================================================
>> --- gcc/testsuite/g++.dg/pr81194.C (nonexistent)
>> +++ gcc/testsuite/g++.dg/pr81194.C (working copy)
>> @@ -0,0 +1,60 @@
>> +// { dg-do compile }
>> +// { dg-options "-O2 -std=c++17 -fno-exceptions" }
>> +
>> +template <class a> struct b { typedef a *c; };
>> +class e {};
>> +template <typename a> class d {
>> +public:
>> + typedef typename b<a>::c c;
>> + c begin();
>> + c end();
>> +};
>> +struct f {
>> + enum { g } h;
>> +};
>> +struct i {
>> + d<f *> j();
>> +};
>> +struct l {
>> + d<i *> k();
>> +};
>> +class ac;
>> +class o {
>> +public:
>> + o(int *, int *, int *, ac *);
>> +};
>> +class ac {
>> +public:
>> + ac(e);
>> + virtual o *ae(int *, int *, int *, int *);
>> +};
>> +class p {
>> + void af(f *m) {
>> + switch (m->h)
>> + case f::g:
>> + ag();
>> + }
>> +
>> +public:
>> + void n() {
>> + l ah;
>> + for (i *ai : ah.k())
>> + for (f *m : ai->j())
>> + af(m);
>> + }
>> + virtual void ag() { __builtin_unreachable(); }
>> +};
>> +template <typename = int> class an : o {
>> +public:
>> + an(int *, int *, int *, int *, ac *);
>> +};
>> +class q : ac {
>> +public:
>> + q() : ac([]() -> e {}()) {}
>> + o *ae(int *ap, int *aq, int *ar, int *as) { an(ap, aq, ar, as, this); }
>> +};
>> +template <typename at>
>> +an<at>::an(int *, int *aq, int *ar, int *as, ac *au) : o(aq, ar, as, au) {
>> + p().n();
>> +}
>> +void av() { new q; }
>>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand
2017-06-29 8:51 ` Richard Biener
@ 2017-06-29 13:00 ` Peter Bergner
0 siblings, 0 replies; 4+ messages in thread
From: Peter Bergner @ 2017-06-29 13:00 UTC (permalink / raw)
To: Richard Biener; +Cc: GCC Patches, marc.mutz
On 6/29/17 3:51 AM, Richard Biener wrote:
> On Thu, Jun 29, 2017 at 10:34 AM, Richard Biener
> To answer myself the unreachable case vanishes at
> execute_cleanup_cfg_post_optimizing
> via group_case_labels. find_taken_edge wouldn't handle this case either.
>
> I am testing a patch fixing both - your patch should still go in.
Ok, committed as revision 249783. Thanks.
Peter
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2017-06-29 13:00 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-06-29 3:01 [PATCH] Fix PR middle-end/81194, ICE during RTL pass: expand Peter Bergner
2017-06-29 8:34 ` Richard Biener
2017-06-29 8:51 ` Richard Biener
2017-06-29 13:00 ` Peter Bergner
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).