| Message ID | ZB1oZcqg/ujxE+D+@tucnak |
|---|---|
| State | New |
| Headers | show |
| Series | builtins: Fix up ICE in inline_string_cmp [PR109258] | expand |
On Fri, 24 Mar 2023, Jakub Jelinek wrote: > Hi! > > The PR109086 r13-6690 inline_string_cmp change to > if (diff != result) > emit_move_insn (result, diff); > regressed > FAIL: go.test/test/fixedbugs/bug207.go, -O2 -g (internal compiler error: in emit_move_insn, at expr.cc:4224) > The problem is the Go FE doesn't mark __builtin_memcmp as pure (I'll also > send patch for that) and so result is const0_rtx when the call lost its lhs > and the above move ICEs because moving something into const0_rtx is obviously > invalid. > I think it is better not to rely on all FEs having these *cmp functions > pure anD DCE being performed. The following patch just punts from the > inline expansion in that case, so we just emit normal library call. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? OK. > 2023-03-24 Jakub Jelinek <jakub@redhat.com> > > PR middle-end/109258 > * builtins.cc (inline_expand_builtin_bytecmp): Return NULL_RTX early > if target == const0_rtx. > > --- gcc/builtins.cc.jj 2023-03-23 10:00:58.308100548 +0100 > +++ gcc/builtins.cc 2023-03-23 11:05:38.308135309 +0100 > @@ -7178,8 +7178,8 @@ inline_expand_builtin_bytecmp (tree exp, > bool is_ncmp = (fcode == BUILT_IN_STRNCMP || fcode == BUILT_IN_MEMCMP); > > /* Do NOT apply this inlining expansion when optimizing for size or > - optimization level below 2. */ > - if (optimize < 2 || optimize_insn_for_size_p ()) > + optimization level below 2 or if unused *cmp hasn't been DCEd. */ > + if (optimize < 2 || optimize_insn_for_size_p () || target == const0_rtx) > return NULL_RTX; > > gcc_checking_assert (fcode == BUILT_IN_STRCMP > > Jakub > >
--- gcc/builtins.cc.jj 2023-03-23 10:00:58.308100548 +0100 +++ gcc/builtins.cc 2023-03-23 11:05:38.308135309 +0100 @@ -7178,8 +7178,8 @@ inline_expand_builtin_bytecmp (tree exp, bool is_ncmp = (fcode == BUILT_IN_STRNCMP || fcode == BUILT_IN_MEMCMP); /* Do NOT apply this inlining expansion when optimizing for size or - optimization level below 2. */ - if (optimize < 2 || optimize_insn_for_size_p ()) + optimization level below 2 or if unused *cmp hasn't been DCEd. */ + if (optimize < 2 || optimize_insn_for_size_p () || target == const0_rtx) return NULL_RTX; gcc_checking_assert (fcode == BUILT_IN_STRCMP
Hi! The PR109086 r13-6690 inline_string_cmp change to if (diff != result) emit_move_insn (result, diff); regressed FAIL: go.test/test/fixedbugs/bug207.go, -O2 -g (internal compiler error: in emit_move_insn, at expr.cc:4224) The problem is the Go FE doesn't mark __builtin_memcmp as pure (I'll also send patch for that) and so result is const0_rtx when the call lost its lhs and the above move ICEs because moving something into const0_rtx is obviously invalid. I think it is better not to rely on all FEs having these *cmp functions pure anD DCE being performed. The following patch just punts from the inline expansion in that case, so we just emit normal library call. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2023-03-24 Jakub Jelinek <jakub@redhat.com> PR middle-end/109258 * builtins.cc (inline_expand_builtin_bytecmp): Return NULL_RTX early if target == const0_rtx. Jakub