| Message ID | CAEwic4bkmzw1ivAp=pto2gMnbGGRq6x8nwrEoYL6mg6CsQY4XQ@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On 06/20/2014 10:52 AM, Kai Tietz wrote: > 2014-06-20 Kai Tietz <ktietz@redhat.com> > > PR target/39284 > * passes.def (peephole2): Add second peephole2 pass before > split before sched2 pass. > * config/i386/i386.md (peehole2): To combine > indirect jump with memory. > (split2): Likewise. Why are we adding a second pass instead of just moving the one? r~
2014-06-20 19:55 GMT+02:00 Richard Henderson <rth@redhat.com>: > On 06/20/2014 10:52 AM, Kai Tietz wrote: >> 2014-06-20 Kai Tietz <ktietz@redhat.com> >> >> PR target/39284 >> * passes.def (peephole2): Add second peephole2 pass before >> split before sched2 pass. >> * config/i386/i386.md (peehole2): To combine >> indirect jump with memory. >> (split2): Likewise. > > Why are we adding a second pass instead of just moving the one? > > > r~ I told that in a prior mail in that thread to Jeff. IIRC there are some conversion of impossible pushes then done too late, additional some patterns about split & dieing register too. Means we produce weaker code. Kai
On 06/20/14 12:07, Kai Tietz wrote: > 2014-06-20 19:55 GMT+02:00 Richard Henderson <rth@redhat.com>: >> On 06/20/2014 10:52 AM, Kai Tietz wrote: >>> 2014-06-20 Kai Tietz <ktietz@redhat.com> >>> >>> PR target/39284 >>> * passes.def (peephole2): Add second peephole2 pass before >>> split before sched2 pass. >>> * config/i386/i386.md (peehole2): To combine >>> indirect jump with memory. >>> (split2): Likewise. >> >> Why are we adding a second pass instead of just moving the one? >> >> >> r~ > > I told that in a prior mail in that thread to Jeff. IIRC there are > some conversion of impossible pushes then done too late, additional > some patterns about split & dieing register too. Means we produce > weaker code. So let's dig into this deeper. Examples & explanations would help. I know it feels like a bit of a runaround, but avoiding adding the pass would be good. jeff
Index: testsuite/gcc.target/i386/indjmp-1.c =================================================================== --- testsuite/gcc.target/i386/indjmp-1.c (Revision 0) +++ testsuite/gcc.target/i386/indjmp-1.c (Arbeitskopie) @@ -0,0 +1,23 @@ +/* { dg-do compile { target ia32 } } */ +/* { dg-options "-O2" } */ + +#define ADVANCE_AND_DISPATCH() goto *addresses[*pc++] + +void +Interpret(const unsigned char *pc) +{ + static const void *const addresses[] = { + &&l0, &&l1, &&l2 + }; + +l0: + ADVANCE_AND_DISPATCH(); + +l1: + ADVANCE_AND_DISPATCH(); + +l2: + return; +} + +/* { dg-final { scan-assembler-not "jmp\[ \t\]*.%eax" } } */ Index: passes.def =================================================================== --- passes.def (Revision 211850) +++ passes.def (Arbeitskopie) @@ -393,6 +393,7 @@ along with GCC; see the file COPYING3. If not see NEXT_PASS (pass_reorder_blocks); NEXT_PASS (pass_branch_target_load_optimize2); NEXT_PASS (pass_leaf_regs); + NEXT_PASS (pass_peephole2); NEXT_PASS (pass_split_before_sched2); NEXT_PASS (pass_sched2); NEXT_PASS (pass_stack_regs); Index: config/i386/i386.md =================================================================== --- config/i386/i386.md (Revision 211850) +++ config/i386/i386.md (Arbeitskopie) @@ -11466,6 +11466,15 @@ (match_dup 4))) (unspec [(const_int 0)] UNSPEC_PEEPSIB)])]) +;; Combining simple memory jump instruction + +(define_peephole2 + [(set (match_operand:W 0 "register_operand") + (match_operand:W 1 "memory_operand")) + (set (pc) (match_dup 0))] + "!TARGET_X32 && peep2_reg_dead_p (2, operands[0])" + [(set (pc) (match_dup 1))]) + ;; Call subroutine, returning value in operand 0 (define_expand "call_value"
So revert to use a second peephole2 pass before final split before sched2 pass. Ok for apply ChangeLog 2014-06-20 Kai Tietz <ktietz@redhat.com> PR target/39284 * passes.def (peephole2): Add second peephole2 pass before split before sched2 pass. * config/i386/i386.md (peehole2): To combine indirect jump with memory. (split2): Likewise. 2014-06-20 Kai Tietz <ktietz@redhat.com> * gcc.target/i386/indjmp-1.c: New test. Tested for i686-pc-cygwin, and x86_64-unknown-linux-gnu running. Ok for apply when boostraps are passing? Regards, Kai