Message ID | 20210615114741.151342-1-aldyh@redhat.com |
---|---|
State | New |
Headers | show |
Series | Add debugging helpers for ranger. | expand |
On Tue, Jun 15, 2021 at 01:47:41PM +0200, Aldy Hernandez via Gcc-patches wrote: > +// ========================================= > +// Debugging helpers. > +// ========================================= > + > +// Query all statements in the IL to precalculate computable ranges in RANGER. Not a review, just a random nit. The above comment doesn't match what the function is actually doing: > + > +static DEBUG_FUNCTION void > +debug_seed_ranger (gimple_ranger &ranger) > +{ > + // Recalculate SCEV to make sure the dump lists everything. > + if (scev_initialized_p ()) > + { > + scev_finalize (); > + scev_initialize (); > + } > + > + basic_block bb; > + int_range_max r; > + FOR_EACH_BB_FN (bb, cfun) > + { > + gimple *last = last_stmt (bb); > + if (last && gimple_get_lhs (last)) > + ranger.range_of_stmt (r, last); which is only doing it for the last stmts in the basic blocks if any. So e.g. in the common case of GIMPLE_COND at the end of a bb it does nothing. Jakub
On 6/15/21 7:55 AM, Jakub Jelinek via Gcc-patches wrote: > On Tue, Jun 15, 2021 at 01:47:41PM +0200, Aldy Hernandez via Gcc-patches wrote: >> +// ========================================= >> +// Debugging helpers. >> +// ========================================= >> + >> +// Query all statements in the IL to precalculate computable ranges in RANGER. > Not a review, just a random nit. > The above comment doesn't match what the function is actually doing: > >> + >> +static DEBUG_FUNCTION void >> +debug_seed_ranger (gimple_ranger &ranger) >> +{ >> + // Recalculate SCEV to make sure the dump lists everything. >> + if (scev_initialized_p ()) >> + { >> + scev_finalize (); >> + scev_initialize (); >> + } >> + >> + basic_block bb; >> + int_range_max r; >> + FOR_EACH_BB_FN (bb, cfun) >> + { >> + gimple *last = last_stmt (bb); >> + if (last && gimple_get_lhs (last)) >> + ranger.range_of_stmt (r, last); > which is only doing it for the last stmts in the basic blocks if any. > So e.g. in the common case of GIMPLE_COND at the end of a bb it does > nothing. > > Jakub > In fact, you can simply drop the gimple_get_lhs (last) part of the condition... range_of_stmt works just fine without a LHS, and will then calculate the GIMPLE_COND operands. Andrew
On Tue, Jun 15, 2021 at 09:43:33AM -0400, Andrew MacLeod wrote: > > > + basic_block bb; > > > + int_range_max r; > > > + FOR_EACH_BB_FN (bb, cfun) > > > + { > > > + gimple *last = last_stmt (bb); > > > + if (last && gimple_get_lhs (last)) > > > + ranger.range_of_stmt (r, last); > > which is only doing it for the last stmts in the basic blocks if any. > > So e.g. in the common case of GIMPLE_COND at the end of a bb it does > > nothing. > > > > Jakub > > > In fact, you can simply drop the gimple_get_lhs (last) part of the > condition... range_of_stmt works just fine without a LHS, and will then > calculate the GIMPLE_COND operands. But even then, not every stmt in the bb participates in the computation of the value (if any) of the last stmt in the block, some stmts could compute values only used in the PHIs or in the middle of other bbs, others could store to memory, or be calls without return value, ... If ranger does caching, can't it just call range_of_stmt on every non-debug stmt in the bb? Jakub
On 6/15/21 9:48 AM, Jakub Jelinek wrote: > On Tue, Jun 15, 2021 at 09:43:33AM -0400, Andrew MacLeod wrote: >>>> + basic_block bb; >>>> + int_range_max r; >>>> + FOR_EACH_BB_FN (bb, cfun) >>>> + { >>>> + gimple *last = last_stmt (bb); >>>> + if (last && gimple_get_lhs (last)) >>>> + ranger.range_of_stmt (r, last); >>> which is only doing it for the last stmts in the basic blocks if any. >>> So e.g. in the common case of GIMPLE_COND at the end of a bb it does >>> nothing. >>> >>> Jakub >>> >> In fact, you can simply drop the gimple_get_lhs (last) part of the >> condition... range_of_stmt works just fine without a LHS, and will then >> calculate the GIMPLE_COND operands. > But even then, not every stmt in the bb participates in the computation > of the value (if any) of the last stmt in the block, some stmts could > compute values only used in the PHIs or in the middle of other bbs, > others could store to memory, or be calls without return value, ... > If ranger does caching, can't it just call range_of_stmt on every non-debug > stmt in the bb? > > Jakub > Yes. Which is exactly what the old execute() routine use to do to fill the cache and dump state. I suspect this is driven by his threader work, and in that particular case, all he cares about is things that are exported so he isn't noticing the other things in the middle of the block :-) Should have noticed the GIMPLE_CONDS I would have thought. perhaps they were collected indirectly if they were relevant. Andrew
On 6/15/21 4:05 PM, Andrew MacLeod wrote: > On 6/15/21 9:48 AM, Jakub Jelinek wrote: >> On Tue, Jun 15, 2021 at 09:43:33AM -0400, Andrew MacLeod wrote: >>>>> + basic_block bb; >>>>> + int_range_max r; >>>>> + FOR_EACH_BB_FN (bb, cfun) >>>>> + { >>>>> + gimple *last = last_stmt (bb); >>>>> + if (last && gimple_get_lhs (last)) >>>>> + ranger.range_of_stmt (r, last); >>>> which is only doing it for the last stmts in the basic blocks if any. >>>> So e.g. in the common case of GIMPLE_COND at the end of a bb it does >>>> nothing. >>>> >>>> Jakub >>>> >>> In fact, you can simply drop the gimple_get_lhs (last) part of the >>> condition... range_of_stmt works just fine without a LHS, and will then >>> calculate the GIMPLE_COND operands. >> But even then, not every stmt in the bb participates in the computation >> of the value (if any) of the last stmt in the block, some stmts could >> compute values only used in the PHIs or in the middle of other bbs, >> others could store to memory, or be calls without return value, ... >> If ranger does caching, can't it just call range_of_stmt on every >> non-debug >> stmt in the bb? >> >> Jakub >> > Yes. Which is exactly what the old execute() routine use to do to fill > the cache and dump state. > > I suspect this is driven by his threader work, and in that particular > case, all he cares about is things that are exported so he isn't > noticing the other things in the middle of the block :-) Should have > noticed the GIMPLE_CONDS I would have thought. perhaps they were > collected indirectly if they were relevant. Bah. Indeed. Fixed. Aldy
Attached is the final version of the patch I have pushed. Thanks. Aldy On Tue, Jun 15, 2021 at 4:26 PM Aldy Hernandez <aldyh@redhat.com> wrote: > > > > On 6/15/21 4:05 PM, Andrew MacLeod wrote: > > On 6/15/21 9:48 AM, Jakub Jelinek wrote: > >> On Tue, Jun 15, 2021 at 09:43:33AM -0400, Andrew MacLeod wrote: > >>>>> + basic_block bb; > >>>>> + int_range_max r; > >>>>> + FOR_EACH_BB_FN (bb, cfun) > >>>>> + { > >>>>> + gimple *last = last_stmt (bb); > >>>>> + if (last && gimple_get_lhs (last)) > >>>>> + ranger.range_of_stmt (r, last); > >>>> which is only doing it for the last stmts in the basic blocks if any. > >>>> So e.g. in the common case of GIMPLE_COND at the end of a bb it does > >>>> nothing. > >>>> > >>>> Jakub > >>>> > >>> In fact, you can simply drop the gimple_get_lhs (last) part of the > >>> condition... range_of_stmt works just fine without a LHS, and will then > >>> calculate the GIMPLE_COND operands. > >> But even then, not every stmt in the bb participates in the computation > >> of the value (if any) of the last stmt in the block, some stmts could > >> compute values only used in the PHIs or in the middle of other bbs, > >> others could store to memory, or be calls without return value, ... > >> If ranger does caching, can't it just call range_of_stmt on every > >> non-debug > >> stmt in the bb? > >> > >> Jakub > >> > > Yes. Which is exactly what the old execute() routine use to do to fill > > the cache and dump state. > > > > I suspect this is driven by his threader work, and in that particular > > case, all he cares about is things that are exported so he isn't > > noticing the other things in the middle of the block :-) Should have > > noticed the GIMPLE_CONDS I would have thought. perhaps they were > > collected indirectly if they were relevant. > > Bah. Indeed. Fixed. > > Aldy >
diff --git a/gcc/gimple-range.cc b/gcc/gimple-range.cc index b534b8e0a2c..2a0da417708 100644 --- a/gcc/gimple-range.cc +++ b/gcc/gimple-range.cc @@ -1661,4 +1661,75 @@ disable_ranger (struct function *fun) fun->x_range_query = &global_ranges; } +// ========================================= +// Debugging helpers. +// ========================================= + +// Query all statements in the IL to precalculate computable ranges in RANGER. + +static DEBUG_FUNCTION void +debug_seed_ranger (gimple_ranger &ranger) +{ + // Recalculate SCEV to make sure the dump lists everything. + if (scev_initialized_p ()) + { + scev_finalize (); + scev_initialize (); + } + + basic_block bb; + int_range_max r; + FOR_EACH_BB_FN (bb, cfun) + { + gimple *last = last_stmt (bb); + if (last && gimple_get_lhs (last)) + ranger.range_of_stmt (r, last); + } +} + +// Dump all that ranger knows for the current function. + +DEBUG_FUNCTION void +dump_ranger (FILE *out) +{ + gimple_ranger ranger; + debug_seed_ranger (ranger); + ranger.dump (out); +} + +DEBUG_FUNCTION void +debug_ranger () +{ + dump_ranger (stderr); +} + +// Dump all that ranger knows on a path of BBs. + +DEBUG_FUNCTION void +dump_ranger (FILE *dump_file, const vec<basic_block> &path) +{ + if (path.length () == 0) + { + fprintf (dump_file, "empty\n"); + return; + } + + gimple_ranger ranger; + debug_seed_ranger (ranger); + + unsigned i = path.length (); + do + { + i--; + ranger.dump_bb (dump_file, path[i]); + } + while (i > 0); +} + +DEBUG_FUNCTION void +debug_ranger (const vec<basic_block> &path) +{ + dump_ranger (stderr, path); +} + #include "gimple-range-tests.cc"