| Message ID | 20180307215108.GN5867@tucnak |
|---|---|
| State | New |
| Headers | show |
| Series | Fix switchconv ICE (PR tree-optimization/84740) | expand |
On March 7, 2018 10:51:08 PM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote: >Hi! > >On the following testcase info.final_bb has only one PHI, the virtual >one, >so info.phi_count is 0. For info.phi_count == 0, we don't really >create any >arrays etc.; just returning early (before create_temp_arrays) would >work, >but would leave the useless GIMPLE_SWITCH in the IL and it would take >many >passes to optimize it properly. We know that all but the default: bbs >are >really empty and the default: can do something arbitrary, if all the >ranges >are consecutive, it is better to just let gen_inbound_check do its work >and convert the switch into if (cond) { former_default_bb }. >All but build_constructors function can handle 0 info.phi_count, but >build_constructors can't, on the other side doesn't need to do anything >if it is 0. > >Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? OK. Richard. >2018-03-07 Jakub Jelinek <jakub@redhat.com> > > PR tree-optimization/84740 > * tree-switch-conversion.c (process_switch): Call build_constructors > only if info.phi_count is non-zero. > > * gcc.dg/torture/pr84740.c: New test. > >--- gcc/tree-switch-conversion.c.jj 2018-01-09 08:58:15.000000000 +0100 >+++ gcc/tree-switch-conversion.c 2018-03-07 12:03:42.228246408 +0100 >@@ -1563,7 +1563,8 @@ process_switch (gswitch *swtch) > gather_default_values (info.default_case_nonstandard > ? gimple_switch_label (swtch, 1) > : gimple_switch_default_label (swtch), &info); >- build_constructors (swtch, &info); >+ if (info.phi_count) >+ build_constructors (swtch, &info); > >build_arrays (swtch, &info); /* Build the static arrays and >assignments. */ > gen_inbound_check (swtch, &info); /* Build the bounds check. */ >--- gcc/testsuite/gcc.dg/torture/pr84740.c.jj 2018-03-07 >12:13:02.153271986 +0100 >+++ gcc/testsuite/gcc.dg/torture/pr84740.c 2018-03-07 >12:12:42.643270368 +0100 >@@ -0,0 +1,24 @@ >+/* PR tree-optimization/84740 */ >+ >+void >+frobulate_for_gcc (unsigned int v) >+{ >+ const char *s; >+ switch (v) >+ { >+ case 0: >+ s = "foo"; >+ break; >+ case 1: >+ s = "bar"; >+ break; >+ case 2: >+ s = "spam"; >+ break; >+ default: >+ s = (const char *) 0; >+ break; >+ } >+ if (!s) >+ __builtin_printf ("%s\n", s); >+} > > Jakub
--- gcc/tree-switch-conversion.c.jj 2018-01-09 08:58:15.000000000 +0100 +++ gcc/tree-switch-conversion.c 2018-03-07 12:03:42.228246408 +0100 @@ -1563,7 +1563,8 @@ process_switch (gswitch *swtch) gather_default_values (info.default_case_nonstandard ? gimple_switch_label (swtch, 1) : gimple_switch_default_label (swtch), &info); - build_constructors (swtch, &info); + if (info.phi_count) + build_constructors (swtch, &info); build_arrays (swtch, &info); /* Build the static arrays and assignments. */ gen_inbound_check (swtch, &info); /* Build the bounds check. */ --- gcc/testsuite/gcc.dg/torture/pr84740.c.jj 2018-03-07 12:13:02.153271986 +0100 +++ gcc/testsuite/gcc.dg/torture/pr84740.c 2018-03-07 12:12:42.643270368 +0100 @@ -0,0 +1,24 @@ +/* PR tree-optimization/84740 */ + +void +frobulate_for_gcc (unsigned int v) +{ + const char *s; + switch (v) + { + case 0: + s = "foo"; + break; + case 1: + s = "bar"; + break; + case 2: + s = "spam"; + break; + default: + s = (const char *) 0; + break; + } + if (!s) + __builtin_printf ("%s\n", s); +}
Hi! On the following testcase info.final_bb has only one PHI, the virtual one, so info.phi_count is 0. For info.phi_count == 0, we don't really create any arrays etc.; just returning early (before create_temp_arrays) would work, but would leave the useless GIMPLE_SWITCH in the IL and it would take many passes to optimize it properly. We know that all but the default: bbs are really empty and the default: can do something arbitrary, if all the ranges are consecutive, it is better to just let gen_inbound_check do its work and convert the switch into if (cond) { former_default_bb }. All but build_constructors function can handle 0 info.phi_count, but build_constructors can't, on the other side doesn't need to do anything if it is 0. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2018-03-07 Jakub Jelinek <jakub@redhat.com> PR tree-optimization/84740 * tree-switch-conversion.c (process_switch): Call build_constructors only if info.phi_count is non-zero. * gcc.dg/torture/pr84740.c: New test. Jakub