| Message ID | 5007C796.4020900@st.com |
|---|---|
| State | New |
| Headers | show |
On Thu, Jul 19, 2012 at 10:38 AM, Christian Bruel <christian.bruel@st.com> wrote: > This is a SH regression on the 4.7 and trunk while building Webkit > (pre-processed file size is about 2.2Mb :-) http://gcc.gnu.org/wiki/A_guide_to_testcase_reduction Ciao! Steven
Christian Bruel <christian.bruel@st.com> wrote: > This is a SH regression on the 4.7 and trunk while building Webkit > (pre-processed file size is about 2.2Mb :-) > > A far branch to a return rtx produces an ICE in > find_dead_or_set_registers at line resource.c:497: > > next = JUMP_LABEL (this_jump_insn); > if (ANY_RETURN_P (next)) <- next is null > next = NULL_RTX; > > Turns out that JUMP_LABEL was not set after gen_return in > sh.c:gen_far_branch. > > This patch fixes this. Tested for sh4-linux > > OK for 4.7 and trunk ? > + if (bp->far_label) > JUMP_LABEL (jump) = bp->far_label; The 2nd line should be indented. OK with that change. Thanks for fixing this! Regards, kaz
On 07/19/2012 11:14 AM, Steven Bosscher wrote: > On Thu, Jul 19, 2012 at 10:38 AM, Christian Bruel > <christian.bruel@st.com> wrote: >> This is a SH regression on the 4.7 and trunk while building Webkit >> (pre-processed file size is about 2.2Mb :-) > > http://gcc.gnu.org/wiki/A_guide_to_testcase_reduction The 2.2Mb file is already preprocessed and cleaned-up. It's c++ and the many inlined functions are necessary to contribute to the final IR in which the far jump exposes. Moving around code, by dichotomy or other core removal techniques is not enough to reduce the problem in a small enough test. no problem to attach it to a new bugzilla before committing the fix. Thanks Christian > > Ciao! > Steven >
On 07/19/2012 12:35 PM, Kaz Kojima wrote: > Christian Bruel <christian.bruel@st.com> wrote: >> This is a SH regression on the 4.7 and trunk while building Webkit >> (pre-processed file size is about 2.2Mb :-) >> >> A far branch to a return rtx produces an ICE in >> find_dead_or_set_registers at line resource.c:497: >> >> next = JUMP_LABEL (this_jump_insn); >> if (ANY_RETURN_P (next)) <- next is null >> next = NULL_RTX; >> >> Turns out that JUMP_LABEL was not set after gen_return in >> sh.c:gen_far_branch. >> >> This patch fixes this. Tested for sh4-linux >> >> OK for 4.7 and trunk ? > >> + if (bp->far_label) >> JUMP_LABEL (jump) = bp->far_label; > > The 2nd line should be indented. OK with that change. > Thanks for fixing this! thanks. the missing indentation was a diff artifact. I'll also add the PR reference to the changelog entry > > Regards, > kaz >
2012-07-19 Christian Bruel <christian.bruel@st.com> * config/sh/sh.c (gen_far_branch): Set JUMP_LABEL for return jumps. Index: gcc/config/sh/sh.c =================================================================== --- gcc/config/sh/sh.c (revision 189613) +++ gcc/config/sh/sh.c (working copy) @@ -5304,6 +5304,7 @@ } else jump = emit_jump_insn_after (gen_return (), insn); + /* Emit a barrier so that reorg knows that any following instructions are not reachable via a fall-through path. But don't do this when not optimizing, since we wouldn't suppress the @@ -5312,7 +5313,16 @@ if (optimize) emit_barrier_after (jump); emit_label_after (bp->near_label, insn); + + if (bp->far_label) JUMP_LABEL (jump) = bp->far_label; + else + { + rtx pat = PATTERN (jump); + gcc_assert (ANY_RETURN_P (pat)); + JUMP_LABEL (jump) = pat; + } + ok = invert_jump (insn, label, 1); gcc_assert (ok);