| Message ID | 20220902143639.2464892-1-hjl.tools@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | Ignore debug insns with CONCAT and CONCATN for insn scheduling | expand |
On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: > CONCAT and CONCATN never appear in the insn chain. They are only used > in debug insn. Ignore debug insns with CONCAT and CONCATN for insn > scheduling to avoid different insn orders with and without debug insn. > > gcc/ > > PR rtl-optimization/106746 > * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT > and CONCATN. Shouldn't we be ignoring everything in a debug insn? I don't see why CONCAT/CONCATN are special here. jeff
On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > > > On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: > > CONCAT and CONCATN never appear in the insn chain. They are only used > > in debug insn. Ignore debug insns with CONCAT and CONCATN for insn > > scheduling to avoid different insn orders with and without debug insn. > > > > gcc/ > > > > PR rtl-optimization/106746 > > * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT > > and CONCATN. > Shouldn't we be ignoring everything in a debug insn? I don't see why > CONCAT/CONCATN are special here. Debug insns are processed by insn scheduling. I think it is to improve debug experiences. It is just that there are no matching usages of CONCAT/CONCATN in non-debug insns. -- H.J.
On 9/21/22 16:11, H.J. Lu wrote: > On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: >> >> >> On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: >>> CONCAT and CONCATN never appear in the insn chain. They are only used >>> in debug insn. Ignore debug insns with CONCAT and CONCATN for insn >>> scheduling to avoid different insn orders with and without debug insn. >>> >>> gcc/ >>> >>> PR rtl-optimization/106746 >>> * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT >>> and CONCATN. >> Shouldn't we be ignoring everything in a debug insn? I don't see why >> CONCAT/CONCATN are special here. > Debug insns are processed by insn scheduling. I think it is to improve debug > experiences. It is just that there are no matching usages of CONCAT/CONCATN > in non-debug insns. But from a dependency standpoint ISTM all debug insn can be ignored. I still don't see why concat/concatn should be special here. jeff
On Sat, Sep 24, 2022 at 1:37 PM Jeff Law <jeffreyalaw@gmail.com> wrote: > > > On 9/21/22 16:11, H.J. Lu wrote: > > On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches > > <gcc-patches@gcc.gnu.org> wrote: > >> > >> > >> On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: > >>> CONCAT and CONCATN never appear in the insn chain. They are only used > >>> in debug insn. Ignore debug insns with CONCAT and CONCATN for insn > >>> scheduling to avoid different insn orders with and without debug insn. > >>> > >>> gcc/ > >>> > >>> PR rtl-optimization/106746 > >>> * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT > >>> and CONCATN. > >> Shouldn't we be ignoring everything in a debug insn? I don't see why > >> CONCAT/CONCATN are special here. > > Debug insns are processed by insn scheduling. I think it is to improve debug > > experiences. It is just that there are no matching usages of CONCAT/CONCATN > > in non-debug insns. > > But from a dependency standpoint ISTM all debug insn can be ignored. I > still don't see why concat/concatn should be special here. > I tried to ignore everything in a debug insn. It caused many regressions in the GCC testsuite.
On 9/26/22 13:52, H.J. Lu wrote: > On Sat, Sep 24, 2022 at 1:37 PM Jeff Law <jeffreyalaw@gmail.com> wrote: >> >> On 9/21/22 16:11, H.J. Lu wrote: >>> On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches >>> <gcc-patches@gcc.gnu.org> wrote: >>>> >>>> On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: >>>>> CONCAT and CONCATN never appear in the insn chain. They are only used >>>>> in debug insn. Ignore debug insns with CONCAT and CONCATN for insn >>>>> scheduling to avoid different insn orders with and without debug insn. >>>>> >>>>> gcc/ >>>>> >>>>> PR rtl-optimization/106746 >>>>> * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT >>>>> and CONCATN. >>>> Shouldn't we be ignoring everything in a debug insn? I don't see why >>>> CONCAT/CONCATN are special here. >>> Debug insns are processed by insn scheduling. I think it is to improve debug >>> experiences. It is just that there are no matching usages of CONCAT/CONCATN >>> in non-debug insns. >> But from a dependency standpoint ISTM all debug insn can be ignored. I >> still don't see why concat/concatn should be special here. >> > I tried to ignore everything in a debug insn. It caused many regressions in > the GCC testsuite. Not terribly useful -- what failed and why? jeff
On Mon, Sep 26, 2022 at 05:23:45PM -0600, Jeff Law via Gcc-patches wrote: > > On 9/26/22 13:52, H.J. Lu wrote: > > On Sat, Sep 24, 2022 at 1:37 PM Jeff Law <jeffreyalaw@gmail.com> wrote: > > > > > > On 9/21/22 16:11, H.J. Lu wrote: > > > > On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches > > > > <gcc-patches@gcc.gnu.org> wrote: > > > > > > > > > > On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: > > > > > > CONCAT and CONCATN never appear in the insn chain. They are only used > > > > > > in debug insn. Ignore debug insns with CONCAT and CONCATN for insn > > > > > > scheduling to avoid different insn orders with and without debug insn. > > > > > > > > > > > > gcc/ > > > > > > > > > > > > PR rtl-optimization/106746 > > > > > > * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT > > > > > > and CONCATN. > > > > > Shouldn't we be ignoring everything in a debug insn? I don't see why > > > > > CONCAT/CONCATN are special here. > > > > Debug insns are processed by insn scheduling. I think it is to improve debug > > > > experiences. It is just that there are no matching usages of CONCAT/CONCATN > > > > in non-debug insns. > > > But from a dependency standpoint ISTM all debug insn can be ignored. I > > > still don't see why concat/concatn should be special here. > > > > > I tried to ignore everything in a debug insn. It caused many regressions in > > the GCC testsuite. > Not terribly useful -- what failed and why? I think the design for debug insns in the scheduler is that they do affect scheduling decisions, but what is in debug insns should only affect actual scheduling of the debug insns and not the rest. So it wouldn't surprise me if ignoring everything in a debug insn broke a lot. But I admit I never fully understood how it works, hopefully Alex or Vlad do. Jakub
On 9/26/22 17:43, Jakub Jelinek wrote: > On Mon, Sep 26, 2022 at 05:23:45PM -0600, Jeff Law via Gcc-patches wrote: >> On 9/26/22 13:52, H.J. Lu wrote: >>> On Sat, Sep 24, 2022 at 1:37 PM Jeff Law <jeffreyalaw@gmail.com> wrote: >>>> On 9/21/22 16:11, H.J. Lu wrote: >>>>> On Wed, Sep 7, 2022 at 10:03 AM Jeff Law via Gcc-patches >>>>> <gcc-patches@gcc.gnu.org> wrote: >>>>>> On 9/2/2022 8:36 AM, H.J. Lu via Gcc-patches wrote: >>>>>>> CONCAT and CONCATN never appear in the insn chain. They are only used >>>>>>> in debug insn. Ignore debug insns with CONCAT and CONCATN for insn >>>>>>> scheduling to avoid different insn orders with and without debug insn. >>>>>>> >>>>>>> gcc/ >>>>>>> >>>>>>> PR rtl-optimization/106746 >>>>>>> * sched-deps.cc (sched_analyze_2): Ignore debug insns with CONCAT >>>>>>> and CONCATN. >>>>>> Shouldn't we be ignoring everything in a debug insn? I don't see why >>>>>> CONCAT/CONCATN are special here. >>>>> Debug insns are processed by insn scheduling. I think it is to improve debug >>>>> experiences. It is just that there are no matching usages of CONCAT/CONCATN >>>>> in non-debug insns. >>>> But from a dependency standpoint ISTM all debug insn can be ignored. I >>>> still don't see why concat/concatn should be special here. >>>> >>> I tried to ignore everything in a debug insn. It caused many regressions in >>> the GCC testsuite. >> Not terribly useful -- what failed and why? > I think the design for debug insns in the scheduler is that they do affect > scheduling decisions, but what is in debug insns should only affect actual > scheduling of the debug insns and not the rest. > So it wouldn't surprise me if ignoring everything in a debug insn broke a > lot. But I admit I never fully understood how it works, hopefully Alex or > Vlad do. But if we look at debug insns in scheduling decisions, wouldn't that likely trigger a variety of compare-debug failures. Something here doesn't make sense. jeff
diff --git a/gcc/sched-deps.cc b/gcc/sched-deps.cc index 948aa0c3b60..b472e4fbb09 100644 --- a/gcc/sched-deps.cc +++ b/gcc/sched-deps.cc @@ -2794,6 +2794,20 @@ sched_analyze_2 (class deps_desc *deps, rtx x, rtx_insn *insn) return; + case VAR_LOCATION: + if (GET_CODE (PAT_VAR_LOCATION_LOC (x)) == CONCAT + || GET_CODE (PAT_VAR_LOCATION_LOC (x)) == CONCATN) + { + /* CONCAT and CONCATN never appear in the insn chain. They + are only used in debug insn. Ignore insns with CONCAT and + CONCATN for insn scheduling to avoid different insn orders + with and without debug insn. */ + if (cslr_p && sched_deps_info->finish_rhs) + sched_deps_info->finish_rhs (); + return; + } + break; + default: break; } diff --git a/gcc/testsuite/gcc.dg/pr106746.c b/gcc/testsuite/gcc.dg/pr106746.c new file mode 100644 index 00000000000..1fc29de28c3 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr106746.c @@ -0,0 +1,30 @@ +/* { dg-do compile } */ +/* { dg-options "-Wno-psabi -O2 -fsched2-use-superblocks -fcompare-debug" } */ +typedef char __attribute__((__vector_size__ (64))) U; +typedef short __attribute__((__vector_size__ (64))) V; +typedef int __attribute__((__vector_size__ (64))) W; + +char c; +U a; +U *r; +W foo0_v512u32_0; + +void +foo (W) +{ + U u; + V v; + W w = __builtin_shuffle (foo0_v512u32_0, foo0_v512u32_0); + u = __builtin_shufflevector (a, u, 3, 0, 4, 9, 9, 6, + 7, 8, 5, 0, 6, 1, 8, 1, + 2, 8, 6, 1, 8, 4, 9, 3, + 8, 4, 6, 0, 9, 0, 1, 8, + 2, 3, 3, 0, 4, 9, 9, 6, + 7, 8, 5, 0, 6, 1, 8, 1, + 2, 8, 6, 1, 8, 4, 9, 3, + 8, 4, 6, 0, 9, 0, 1, 8, + 2, 3); + v *= c; + w &= c; + *r = (U) v + (U) w; +}