Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2015-02-11 18:18 GMT+00:00 Jiong Wang : > On 11/02/15 14:21, Kenneth Zadeck wrote: >> >> On 02/11/2015 06:20 AM, Jiong Wang wrote: >>> >>> 2014-12-19 15:21 GMT+00:00 Kenneth Zadeck : however, since i am a nice person loop-invariant solves the DF_UD_CHAIN which builds a data structure that connects each use with all of the defs that reach it. I believe that this is the opposite of what you want here. if you really need this, you need to also turn on the DF_DU_CHAIN which builds the opposite structure.Both structures can be space hogs, but they are both turned on in other places in the compiler so it is unlikely to be an issue. >>> >>> Exactly, Thanks, Kenneth. >>> >>> This approach works from my experiment and look much better than >>> previous REG_NOTE approach. >>> while it do have one problem. We need the UD/DU chain built before we >>> do insn re-shuffling. >>> While after re-shuffling, UD chain needs update, otherwise, the later >>> "check_dependecies" in find_invariant_insn may fail. >>> >>> although we have re-shuffle instruction 1 into 2, the later check >>> still using old UD info, thus >>> decide instruction 2 is not iv. >>> >>> 1: regA <- vfp + regB >>> 2: regA <- vfp + const >>> >>> my current fix is to insert those re-shuffled insn into a table named >>> "vfp_const_iv", then skip those >>> dependencies check for them as they don't have any dependencies. >> >> You now are beginning to understand why Mark Wegman and I invented SSA >> form almost 30 years ago > > > Indeed, thanks. > > attachment is the new draft patch, pass x86-84 bootstrap > (actually it will not affect x86-64, because of addressing mode), > while failed on aarch64 bootstrap during stage2/3 binary comparision, > there must be some glitches needs to be fixed. > > Will clean up later after I finished my upcoming long flight and what I > am seeking now is whether the general thoughts, code logic & framework, in > this > patch is acceptable to the community? > > personally, I think this version is much better than previous version. > new added code integrated with existed code in a more natural way. no use > of unsafe REG_NOTE info which is not updated in this pass. > > and from AArch64 gcc bootstrap, 239 new loop invariant found by this > re-shuffling. so, this small improvement on rtl loop invariant pass do have > some value. > > please review and give comments. > > Thanks. Ping ~ And for the bootstrap stage2/3 binary comparision failure, the different is in ira-costs.o. because for stage2, we actually add one extra compilation option "-gtoggle" while not for stage3, as we want to make sure debug info generation will not affect any real instruction generation. but, after some investigation I found gcc actually generate difference code when debug info enabled because DEBUG_INSN will affect data flow analysis. (insn 2556 2555 2776 257 (set (reg/f:DI 1473) (plus:DI (reg/f:DI 64 sfp) (reg:DI 1474))) ../../gcc/gcc/ira-costs.c:628 87 {*adddi3_aarch64} (expr_list:REG_DEAD (reg:DI 1474) (nil))) (debug_insn 2776 2556 2557 257 (var_location:SI D#105 (mem:SI (plus:DI (reg/f:DI 1473) (const_int -240 [0xff10])) [23 classes S4 A32])) -1 (nil)) (insn 2557 2776 2558 257 (set (reg:SI 591 [ D.69930 ]) (mem:SI (plus:DI (reg/f:DI 1473) (const_int -240 [0xff10])) [23 classes S4 A32])) ../../gcc/gcc/ira-costs.c:628 39 {*movsi_aarch64} (expr_list:REG_DEAD (reg/f:DI 1473) (nil))) without "debug_insn 2776", operands of insn 2556 and 2557 can be shuffled, so that insn 2556 becomes "reg 1473 = sfp - 240", thus hoisted out of the loop as invariant. while with debug_insn 2776, reg 1473 are used in both insn 2776 and insn 2557, thus the single def-use analysis returns false, thus we won't do any transformation on this which is correct. So I think this stage2/3 binary difference is acceptable? for compile time, I haven't found difference from -ftime-report. and ree.c and enabled "DF_UD_CHAIN + DF_DU_CHAIN" also. > > 2015-02-11 Jiong Wang > > gcc/ > * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. > (vfp_const_iv): New hash table. > (expensive_addr_check_p): New boolean. > (init_inv_motion_data): Initialize new variables. > (free_inv_motion_data): Release hash table. > (create_new_invariant): Set cheap_address to false for iv in > vfp_const_iv table. > (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv > table. > (use_for_single_du): New function. > (reshuffle_insn_with_vfp): Likewise. > (find_invariants_bb): Call reshuffle_insn_with_vfp. > gcc/testsuite/ > * gcc.dg/pr62173.c: New testcase.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Tue, Apr 14, 2015 at 5:06 PM, Jiong Wang wrote: > but, after some investigation I found gcc actually generate difference > code when debug info enabled because > DEBUG_INSN will affect data flow analysis. It should not, so that's a bug. > So I think this stage2/3 binary difference is acceptable? No, they should be identical. If there's a difference, then there's a bug - which, it seems, you've already found, too. Ciao! Steven
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 04/14/2015 10:48 AM, Steven Bosscher wrote: So I think this stage2/3 binary difference is acceptable? No, they should be identical. If there's a difference, then there's a bug - which, it seems, you've already found, too. RIght. And so the natural question is how to fix. At first glance it would seem like having this new code ignore dependencies rising from debug insns would work. Which then begs the question, what happens to the debug insn -- it's certainly not going to be correct anymore if the transformation is made. jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2015-04-14 18:24 GMT+01:00 Jeff Law : > On 04/14/2015 10:48 AM, Steven Bosscher wrote: >>> >>> So I think this stage2/3 binary difference is acceptable? >> >> >> No, they should be identical. If there's a difference, then there's a >> bug - which, it seems, you've already found, too. > > RIght. And so the natural question is how to fix. > > At first glance it would seem like having this new code ignore dependencies > rising from debug insns would work. > > Which then begs the question, what happens to the debug insn -- it's > certainly not going to be correct anymore if the transformation is made. Exactly. The debug_insn 2776 in my example is to record the base address of a local array. the new code is doing correctly here by not shuffling the operands of insn 2556 and 2557 as there is additional reference of reg:1473 from debug insn, although the code will still execute correctly if we do the transformation. my understanding to fix this: * delete the out-of-date mismatch debug_insn? as there is no guarantee to generate accurate debug info under -O2. IMO, this debug_insn may affect "DW_AT_location" field for variable descrption of "classes" in .debug_info section, but it's omitted in the final output already. <3><38a4d>: Abbrev Number: 137 (DW_TAG_variable) <38a4f> DW_AT_name : (indirect string, offset: 0x18db): classes <38a53> DW_AT_decl_file : 1 <38a54> DW_AT_decl_line : 548 <38a56> DW_AT_type: <0x38cb4> * update the debug_insn? if the following change is OK with dwarf standard from insn0: reg0 = fp + reg1 debug_insn: var_loc = reg0 + const_off insn1: reg2 = reg0 + const_off to insn0: reg0 = fp + const_off debug_insn: var_loc = reg0 + reg1 insn1: reg2 = reg0 + reg1 Thanks, Regards, Jiong > > jeff >
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2015-02-11 18:18 GMT+00:00 Jiong Wang : > > 2015-02-11 Jiong Wang > > gcc/ > * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. > (vfp_const_iv): New hash table. > (expensive_addr_check_p): New boolean. > (init_inv_motion_data): Initialize new variables. > (free_inv_motion_data): Release hash table. > (create_new_invariant): Set cheap_address to false for iv in > vfp_const_iv table. > (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv > table. > (use_for_single_du): New function. > (reshuffle_insn_with_vfp): Likewise. > (find_invariants_bb): Call reshuffle_insn_with_vfp. For performance measurement: on AArch64: * this patch finds several hundreds new loop invariants across speck2k6. * one bench in spec2k6 float get +4.5% performance improvement. I guess similar improvements could be achieved on other RISC backend. Regards, Jiong
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Jiong Wang writes: > 2015-04-14 18:24 GMT+01:00 Jeff Law : >> On 04/14/2015 10:48 AM, Steven Bosscher wrote: So I think this stage2/3 binary difference is acceptable? >>> >>> >>> No, they should be identical. If there's a difference, then there's a >>> bug - which, it seems, you've already found, too. >> >> RIght. And so the natural question is how to fix. >> >> At first glance it would seem like having this new code ignore dependencies >> rising from debug insns would work. >> >> Which then begs the question, what happens to the debug insn -- it's >> certainly not going to be correct anymore if the transformation is made. > > Exactly. > > The debug_insn 2776 in my example is to record the base address of a > local array. the new code is doing correctly here by not shuffling the > operands of insn 2556 and 2557 as there is additional reference of > reg:1473 from debug insn, although the code will still execute correctly > if we do the transformation. > > my understanding to fix this: > > * delete the out-of-date mismatch debug_insn? as there is no guarantee > to generate accurate debug info under -O2. > > IMO, this debug_insn may affect "DW_AT_location" field for variable > descrption of "classes" in .debug_info section, but it's omitted in > the final output already. > > <3><38a4d>: Abbrev Number: 137 (DW_TAG_variable) > <38a4f> DW_AT_name : (indirect string, offset: 0x18db): classes > <38a53> DW_AT_decl_file : 1 > <38a54> DW_AT_decl_line : 548 > <38a56> DW_AT_type: <0x38cb4> > > * update the debug_insn? if the following change is OK with dwarf standard > >from > > insn0: reg0 = fp + reg1 > debug_insn: var_loc = reg0 + const_off > insn1: reg2 = reg0 + const_off > >to > > insn0: reg0 = fp + const_off > debug_insn: var_loc = reg0 + reg1 > insn1: reg2 = reg0 + reg1 > > Thanks, > And attachment is the new patch which will update debug_insn as described in the second solution above. Now the stage2/3 binary differences on AArch64 gone away. Bootstrap OK. On AArch64, this patch give 600+ new rtl loop invariants found across spec2k6 float. +4.5% perf improvement on 436.cactusADM because four new invariants found in the critical function "regex_compile". The similar improvements may be achieved on other RISC backends like powerpc/mips I guess. One thing to mention, for AArch64, one minor glitch in aarch64_legitimize_address needs to be fixed to let this patch take effect, I will send out that patch later as it's a seperate issue. Powerpc/Mips don't have this glitch in LEGITIMIZE_ADDRESS hook, so should be OK, and I verified the base address of local array in the testcase given by Seb on pr62173 do hoisted on ppc64 now. I think pr62173 is fixed on those 64bit arch by this patch. Thoughts? Thanks. 2015-04-21 Jiong Wang gcc/ * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. (vfp_const_iv): New hash table. (expensive_addr_check_p): New boolean. (init_inv_motion_data): Initialize new variables.> (free_inv_motion_data): Release hash table. (create_new_invariant): Set cheap_address to false for iv in vfp_const_iv table. (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv table. (use_for_single_du): New function. (reshuffle_insn_with_vfp): Likewise. (find_invariants_bb): Call reshuffle_insn_with_vfp. gcc/testsuite/ * gcc.dg/pr62173.c: New testcase. -- Regards, Jiong diff --git a/gcc/loop-invariant.c b/gcc/loop-invariant.c index f79b497..f70dfb0 100644 --- a/gcc/loop-invariant.c +++ b/gcc/loop-invariant.c @@ -203,6 +203,8 @@ typedef struct invariant *invariant_p; /* The invariants. */ static vec invariants; +static hash_table > *vfp_const_iv; +static bool need_expensive_addr_check_p; /* Check the size of the invariant table and realloc if necessary. */ @@ -695,7 +697,7 @@ find_defs (struct loop *loop) df_remove_problem (df_chain); df_process_deferred_rescans (); - df_chain_add_problem (DF_UD_CHAIN); + df_chain_add_problem (DF_UD_CHAIN + DF_DU_CHAIN); df_set_flags (DF_RD_PRUNE_DEAD_DEFS); df_analyze_loop (loop); check_invariant_table_size (); @@ -742,6 +744,9 @@ create_new_invariant (struct def *def, rtx_insn *insn, bitmap depends_on, See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ inv->cheap_address = address_cost (SET_SRC (set), word_mode, ADDR_SPACE_GENERIC, speed) < 3; + + if (need_expensive_addr_check_p && vfp_const_iv->find (insn)) + inv->cheap_address = false; } else { @@ -952,7 +957,8 @@ find_invariant_insn (rtx_insn *insn, bool always_reached, bool always_executed) return; depends_on = BITMAP_ALLOC (NULL); - if (!check_dependencies (insn, depends_on)) + if (!vfp_const_iv->find (insn) + && !check_dependencies (insn, depends_on)) { BITMAP_FREE (depends_on); return; @@ -1007,6 +1013,180 @@ find_invariants_insn (rtx_insn *in
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 04/21/2015 08:24 AM, Jiong Wang wrote: Jiong Wang writes: 2015-04-14 18:24 GMT+01:00 Jeff Law : On 04/14/2015 10:48 AM, Steven Bosscher wrote: So I think this stage2/3 binary difference is acceptable? No, they should be identical. If there's a difference, then there's a bug - which, it seems, you've already found, too. RIght. And so the natural question is how to fix. At first glance it would seem like having this new code ignore dependencies rising from debug insns would work. Which then begs the question, what happens to the debug insn -- it's certainly not going to be correct anymore if the transformation is made. Exactly. The debug_insn 2776 in my example is to record the base address of a local array. the new code is doing correctly here by not shuffling the operands of insn 2556 and 2557 as there is additional reference of reg:1473 from debug insn, although the code will still execute correctly if we do the transformation. my understanding to fix this: * delete the out-of-date mismatch debug_insn? as there is no guarantee to generate accurate debug info under -O2. IMO, this debug_insn may affect "DW_AT_location" field for variable descrption of "classes" in .debug_info section, but it's omitted in the final output already. <3><38a4d>: Abbrev Number: 137 (DW_TAG_variable) <38a4f> DW_AT_name : (indirect string, offset: 0x18db): classes <38a53> DW_AT_decl_file : 1 <38a54> DW_AT_decl_line : 548 <38a56> DW_AT_type: <0x38cb4> * update the debug_insn? if the following change is OK with dwarf standard from insn0: reg0 = fp + reg1 debug_insn: var_loc = reg0 + const_off insn1: reg2 = reg0 + const_off to insn0: reg0 = fp + const_off debug_insn: var_loc = reg0 + reg1 insn1: reg2 = reg0 + reg1 Thanks, And attachment is the new patch which will update debug_insn as described in the second solution above. Now the stage2/3 binary differences on AArch64 gone away. Bootstrap OK. On AArch64, this patch give 600+ new rtl loop invariants found across spec2k6 float. +4.5% perf improvement on 436.cactusADM because four new invariants found in the critical function "regex_compile". The similar improvements may be achieved on other RISC backends like powerpc/mips I guess. One thing to mention, for AArch64, one minor glitch in aarch64_legitimize_address needs to be fixed to let this patch take effect, I will send out that patch later as it's a seperate issue. Powerpc/Mips don't have this glitch in LEGITIMIZE_ADDRESS hook, so should be OK, and I verified the base address of local array in the testcase given by Seb on pr62173 do hoisted on ppc64 now. I think pr62173 is fixed on those 64bit arch by this patch. Thoughts? Thanks. 2015-04-21 Jiong Wang gcc/ * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. (vfp_const_iv): New hash table. (expensive_addr_check_p): New boolean. (init_inv_motion_data): Initialize new variables.> (free_inv_motion_data): Release hash table. (create_new_invariant): Set cheap_address to false for iv in vfp_const_iv table. (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv table. (use_for_single_du): New function. (reshuffle_insn_with_vfp): Likewise. (find_invariants_bb): Call reshuffle_insn_with_vfp. gcc/testsuite/ * gcc.dg/pr62173.c: New testcase. So ultimately isn't this just a specialized version of reassociation of pointer arithmetic? And if so, don't we have all the usual problems around introducing overflow? ISTM if this is going to go forward (ie, we somehow convince ourselves that this kind of reassociation is OK), then should it be made to apply on pointers in general rather than restricting to stack, frame, virtual-frame? I wish I'd looked more closely at the patch the first time around so that I could have raised these questions sooner. Jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Jeff Law writes: > On 04/21/2015 08:24 AM, Jiong Wang wrote: >> >> Jiong Wang writes: >> >>> 2015-04-14 18:24 GMT+01:00 Jeff Law : On 04/14/2015 10:48 AM, Steven Bosscher wrote: >> >> So I think this stage2/3 binary difference is acceptable? > > > No, they should be identical. If there's a difference, then there's a > bug - which, it seems, you've already found, too. RIght. And so the natural question is how to fix. At first glance it would seem like having this new code ignore dependencies rising from debug insns would work. Which then begs the question, what happens to the debug insn -- it's certainly not going to be correct anymore if the transformation is made. >>> >>> Exactly. >>> >>> The debug_insn 2776 in my example is to record the base address of a >>> local array. the new code is doing correctly here by not shuffling the >>> operands of insn 2556 and 2557 as there is additional reference of >>> reg:1473 from debug insn, although the code will still execute correctly >>> if we do the transformation. >>> >>> my understanding to fix this: >>> >>>* delete the out-of-date mismatch debug_insn? as there is no guarantee >>> to generate accurate debug info under -O2. >>> >>> IMO, this debug_insn may affect "DW_AT_location" field for variable >>> descrption of "classes" in .debug_info section, but it's omitted in >>> the final output already. >>> >>> <3><38a4d>: Abbrev Number: 137 (DW_TAG_variable) >>> <38a4f> DW_AT_name : (indirect string, offset: 0x18db): classes >>> <38a53> DW_AT_decl_file : 1 >>> <38a54> DW_AT_decl_line : 548 >>> <38a56> DW_AT_type: <0x38cb4> >>> >>>* update the debug_insn? if the following change is OK with dwarf >>> standard >>> >>> from >>> >>> insn0: reg0 = fp + reg1 >>> debug_insn: var_loc = reg0 + const_off >>> insn1: reg2 = reg0 + const_off >>> >>> to >>> >>> insn0: reg0 = fp + const_off >>> debug_insn: var_loc = reg0 + reg1 >>> insn1: reg2 = reg0 + reg1 >>> >>> Thanks, >>> >> >> And attachment is the new patch which will update debug_insn as >> described in the second solution above. >> >> Now the stage2/3 binary differences on AArch64 gone away. Bootstrap OK. >> >> On AArch64, this patch give 600+ new rtl loop invariants found across >> spec2k6 float. +4.5% perf improvement on 436.cactusADM because four new >> invariants found in the critical function "regex_compile". >> >> The similar improvements may be achieved on other RISC backends like >> powerpc/mips I guess. >> >> One thing to mention, for AArch64, one minor glitch in >> aarch64_legitimize_address needs to be fixed to let this patch take >> effect, I will send out that patch later as it's a seperate issue. >> Powerpc/Mips don't have this glitch in LEGITIMIZE_ADDRESS hook, so >> should be OK, and I verified the base address of local array in the >> testcase given by Seb on pr62173 do hoisted on ppc64 now. I think >> pr62173 is fixed on those 64bit arch by this patch. >> >> Thoughts? >> >> Thanks. >> >> 2015-04-21 Jiong Wang >> >> gcc/ >>* loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. >>(vfp_const_iv): New hash table. >>(expensive_addr_check_p): New boolean. >>(init_inv_motion_data): Initialize new variables.> >>(free_inv_motion_data): Release hash table. >>(create_new_invariant): Set cheap_address to false for iv in >>vfp_const_iv table. >>(find_invariant_insn): Skip dependencies check for iv in vfp_const_iv >>table. >>(use_for_single_du): New function. >>(reshuffle_insn_with_vfp): Likewise. >>(find_invariants_bb): Call reshuffle_insn_with_vfp. >> >> gcc/testsuite/ >> * gcc.dg/pr62173.c: New testcase. > So ultimately isn't this just a specialized version of reassociation of > pointer arithmetic? And if so, don't we have all the usual problems > around introducing overflow? > > > ISTM if this is going to go forward (ie, we somehow convince ourselves > that this kind of reassociation is OK), then should it be made to apply > on pointers in general rather than restricting to stack, frame, > virtual-frame? > Jeff, Thanks for the review. This transformation is not reassociation of pointer arithmetic. The idea of this patch is, given two instructions with variable value, we may get new instruction sequences with fixed value by reassociating their operands. And currently GCC only generate such instruction sequences for local array accessing as far as I known. for the statement "D.2707 = A[i]", gcc generate the following instruction sequence: (insn 92 91 93 6 (set (reg/f:DI 148) (plus:DI (reg/f:DI 64 sfp) (reg:DI 147 [ i ]))) (expr_list:REG_DEAD (reg:DI 147 [ i ]) (nil))) (insn 93 92 94 6 (set (reg:SI 149 [ D.2707 ]) (zero_extend:SI (mem/j:QI (plus:DI (reg/f:DI 148) (const_in
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Hi Matthew, 2015-04-21 15:24 GMT+01:00 Jiong Wang : > > 2015-04-21 Jiong Wang > > gcc/ > * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. > (vfp_const_iv): New hash table. > (expensive_addr_check_p): New boolean. > (init_inv_motion_data): Initialize new variables.> > (free_inv_motion_data): Release hash table. > (create_new_invariant): Set cheap_address to false for iv in > vfp_const_iv table. > (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv > table. > (use_for_single_du): New function. > (reshuffle_insn_with_vfp): Likewise. > (find_invariants_bb): Call reshuffle_insn_with_vfp. Is it possible for you to test this patch on spec2k6 on MIPS? especially 436.cactusADM. I am interested in the performance impact on other load/store machines. Thanks. Regards, Jiong
RE: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
> Hi Matthew, > > 2015-04-21 15:24 GMT+01:00 Jiong Wang : > > > > > 2015-04-21 Jiong Wang > > > > gcc/ > > * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. > > (vfp_const_iv): New hash table. > > (expensive_addr_check_p): New boolean. > > (init_inv_motion_data): Initialize new variables.> > > (free_inv_motion_data): Release hash table. > > (create_new_invariant): Set cheap_address to false for iv in > > vfp_const_iv table. > > (find_invariant_insn): Skip dependencies check for iv in > vfp_const_iv > > table. > > (use_for_single_du): New function. > > (reshuffle_insn_with_vfp): Likewise. > > (find_invariants_bb): Call reshuffle_insn_with_vfp. > > Is it possible for you to test this patch on spec2k6 on MIPS? > especially 436.cactusADM. > I am interested in the performance impact on other load/store machines. I will certainly try and fit it in. It may take me a week to get back to you though as I'm away on vacation for a few days. Matthew
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2015-04-28 14:56 GMT+01:00 Matthew Fortune : >> Hi Matthew, >> >> 2015-04-21 15:24 GMT+01:00 Jiong Wang : >> >> > >> > 2015-04-21 Jiong Wang >> > >> > gcc/ >> > * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. >> > (vfp_const_iv): New hash table. >> > (expensive_addr_check_p): New boolean. >> > (init_inv_motion_data): Initialize new variables.> >> > (free_inv_motion_data): Release hash table. >> > (create_new_invariant): Set cheap_address to false for iv in >> > vfp_const_iv table. >> > (find_invariant_insn): Skip dependencies check for iv in >> vfp_const_iv >> > table. >> > (use_for_single_du): New function. >> > (reshuffle_insn_with_vfp): Likewise. >> > (find_invariants_bb): Call reshuffle_insn_with_vfp. >> >> Is it possible for you to test this patch on spec2k6 on MIPS? >> especially 436.cactusADM. >> I am interested in the performance impact on other load/store machines. > > I will certainly try and fit it in. It may take me a week to get back to > you though as I'm away on vacation for a few days. Thanks very much. I appreciate that. And you can add "-fdump-rtl-loop2_invariant", then we can see how many new loop invariants found at RTL level across spec2k6 on MIPS. Thanks. Regards, Jiong
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2014-12-19 15:21 GMT+00:00 Kenneth Zadeck : > > however, since i am a nice person > > loop-invariant solves the DF_UD_CHAIN which builds a data structure that > connects each use with all of the defs that reach it. I believe that this > is the opposite of what you want here. > > if you really need this, you need to also turn on the DF_DU_CHAIN which > builds the opposite structure.Both structures can be space hogs, but > they are both turned on in other places in the compiler so it is unlikely to > be an issue. Exactly, Thanks, Kenneth. This approach works from my experiment and look much better than previous REG_NOTE approach. while it do have one problem. We need the UD/DU chain built before we do insn re-shuffling. While after re-shuffling, UD chain needs update, otherwise, the later "check_dependecies" in find_invariant_insn may fail. although we have re-shuffle instruction 1 into 2, the later check still using old UD info, thus decide instruction 2 is not iv. 1: regA <- vfp + regB 2: regA <- vfp + const my current fix is to insert those re-shuffled insn into a table named "vfp_const_iv", then skip those dependencies check for them as they don't have any dependencies. > > > >> > LOG_LINKs have nothing to do with single use; they point from the > _first_ > use to its corresponding def. > > You might want to look at what fwprop does instead. Pass rtl fwprop uses df information in single-definition way, it doesn't really take into consideration if register is a single use. This often corrupts other optimizations like post-increment and load/store pair. For example: add r2, r1, r0 ldr rx, [r2] add r2, r2, #4 is transformed into below form: add r2, r1, r0 ldr rx, [r1, r0] add r2, r2, #4 As a result, post-increment opportunity is corrupted, also definition of r2 can't be deleted because it's not single use. Thanks, bin >>> >>> thanks for all these suggestion. >>> >>> Have look at the LOG_LINK build function, a simple reverse scan, while >>> needs to allocate big map array for all pseudo regs. Haven't looked at >>> similar code in fwprop, >>> >>> actually, when found the first matched insn pattern, I just want to >>> scan several insns next, then abort quickly if nothing meet >>> requirement. there is no need to build full single-use information. >>> >>> still can anyone confirm that it is safe to re-use REG_DEAD info there >>> without calling df_note_add_problem and df_analysis first? or I am >>> using those info passed down from the previous pass which calculated >>> these info and maybe broken? >> >> It's not safe to use REG_DEAD info without re-computing it. > > not sure that reg_dead is the right answer even if you did re-compute it. > I believe you can have two parallel uses (on both sides of an if-then-else) > for a single def (above the if then else) and have two REG_DEAD notes. > >> Richard. > >
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 02/11/2015 06:20 AM, Jiong Wang wrote: 2014-12-19 15:21 GMT+00:00 Kenneth Zadeck : however, since i am a nice person loop-invariant solves the DF_UD_CHAIN which builds a data structure that connects each use with all of the defs that reach it. I believe that this is the opposite of what you want here. if you really need this, you need to also turn on the DF_DU_CHAIN which builds the opposite structure.Both structures can be space hogs, but they are both turned on in other places in the compiler so it is unlikely to be an issue. Exactly, Thanks, Kenneth. This approach works from my experiment and look much better than previous REG_NOTE approach. while it do have one problem. We need the UD/DU chain built before we do insn re-shuffling. While after re-shuffling, UD chain needs update, otherwise, the later "check_dependecies" in find_invariant_insn may fail. although we have re-shuffle instruction 1 into 2, the later check still using old UD info, thus decide instruction 2 is not iv. 1: regA <- vfp + regB 2: regA <- vfp + const my current fix is to insert those re-shuffled insn into a table named "vfp_const_iv", then skip those dependencies check for them as they don't have any dependencies. You now are beginning to understand why Mark Wegman and I invented SSA form almost 30 years ago There is no good way to keep the information up to data as you are changing the program.To a large extent you are on your own. If what you are suggesting works, then this is likely much faster than rebuilding the chains. LOG_LINKs have nothing to do with single use; they point from the _first_ use to its corresponding def. You might want to look at what fwprop does instead. Pass rtl fwprop uses df information in single-definition way, it doesn't really take into consideration if register is a single use. This often corrupts other optimizations like post-increment and load/store pair. For example: add r2, r1, r0 ldr rx, [r2] add r2, r2, #4 is transformed into below form: add r2, r1, r0 ldr rx, [r1, r0] add r2, r2, #4 As a result, post-increment opportunity is corrupted, also definition of r2 can't be deleted because it's not single use. Thanks, bin thanks for all these suggestion. Have look at the LOG_LINK build function, a simple reverse scan, while needs to allocate big map array for all pseudo regs. Haven't looked at similar code in fwprop, actually, when found the first matched insn pattern, I just want to scan several insns next, then abort quickly if nothing meet requirement. there is no need to build full single-use information. still can anyone confirm that it is safe to re-use REG_DEAD info there without calling df_note_add_problem and df_analysis first? or I am using those info passed down from the previous pass which calculated these info and maybe broken? It's not safe to use REG_DEAD info without re-computing it. not sure that reg_dead is the right answer even if you did re-compute it. I believe you can have two parallel uses (on both sides of an if-then-else) for a single def (above the if then else) and have two REG_DEAD notes. Richard.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 11/02/15 14:21, Kenneth Zadeck wrote: On 02/11/2015 06:20 AM, Jiong Wang wrote: 2014-12-19 15:21 GMT+00:00 Kenneth Zadeck : however, since i am a nice person loop-invariant solves the DF_UD_CHAIN which builds a data structure that connects each use with all of the defs that reach it. I believe that this is the opposite of what you want here. if you really need this, you need to also turn on the DF_DU_CHAIN which builds the opposite structure.Both structures can be space hogs, but they are both turned on in other places in the compiler so it is unlikely to be an issue. Exactly, Thanks, Kenneth. This approach works from my experiment and look much better than previous REG_NOTE approach. while it do have one problem. We need the UD/DU chain built before we do insn re-shuffling. While after re-shuffling, UD chain needs update, otherwise, the later "check_dependecies" in find_invariant_insn may fail. although we have re-shuffle instruction 1 into 2, the later check still using old UD info, thus decide instruction 2 is not iv. 1: regA <- vfp + regB 2: regA <- vfp + const my current fix is to insert those re-shuffled insn into a table named "vfp_const_iv", then skip those dependencies check for them as they don't have any dependencies. You now are beginning to understand why Mark Wegman and I invented SSA form almost 30 years ago Indeed, thanks. attachment is the new draft patch, pass x86-84 bootstrap (actually it will not affect x86-64, because of addressing mode), while failed on aarch64 bootstrap during stage2/3 binary comparision, there must be some glitches needs to be fixed. Will clean up later after I finished my upcoming long flight and what I am seeking now is whether the general thoughts, code logic & framework, in this patch is acceptable to the community? personally, I think this version is much better than previous version. new added code integrated with existed code in a more natural way. no use of unsafe REG_NOTE info which is not updated in this pass. and from AArch64 gcc bootstrap, 239 new loop invariant found by this re-shuffling. so, this small improvement on rtl loop invariant pass do have some value. please review and give comments. Thanks. 2015-02-11 Jiong Wang gcc/ * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. (vfp_const_iv): New hash table. (expensive_addr_check_p): New boolean. (init_inv_motion_data): Initialize new variables. (free_inv_motion_data): Release hash table. (create_new_invariant): Set cheap_address to false for iv in vfp_const_iv table. (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv table. (use_for_single_du): New function. (reshuffle_insn_with_vfp): Likewise. (find_invariants_bb): Call reshuffle_insn_with_vfp. gcc/testsuite/ * gcc.dg/pr62173.c: New testcase. diff --git a/gcc/loop-invariant.c b/gcc/loop-invariant.c index f79b497..b1c4760 100644 --- a/gcc/loop-invariant.c +++ b/gcc/loop-invariant.c @@ -203,6 +203,8 @@ typedef struct invariant *invariant_p; /* The invariants. */ static vec invariants; +static hash_table > *vfp_const_iv; +static bool need_expensive_addr_check_p; /* Check the size of the invariant table and realloc if necessary. */ @@ -695,7 +697,7 @@ find_defs (struct loop *loop) df_remove_problem (df_chain); df_process_deferred_rescans (); - df_chain_add_problem (DF_UD_CHAIN); + df_chain_add_problem (DF_UD_CHAIN + DF_DU_CHAIN); df_set_flags (DF_RD_PRUNE_DEAD_DEFS); df_analyze_loop (loop); check_invariant_table_size (); @@ -742,6 +744,9 @@ create_new_invariant (struct def *def, rtx_insn *insn, bitmap depends_on, See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ inv->cheap_address = address_cost (SET_SRC (set), word_mode, ADDR_SPACE_GENERIC, speed) < 3; + + if (need_expensive_addr_check_p && vfp_const_iv->find (insn)) + inv->cheap_address = false; } else { @@ -952,7 +957,8 @@ find_invariant_insn (rtx_insn *insn, bool always_reached, bool always_executed) return; depends_on = BITMAP_ALLOC (NULL); - if (!check_dependencies (insn, depends_on)) + if (!vfp_const_iv->find (insn) + && !check_dependencies (insn, depends_on)) { BITMAP_FREE (depends_on); return; @@ -1007,6 +1013,143 @@ find_invariants_insn (rtx_insn *insn, bool always_reached, bool always_executed) record_uses (insn); } +/* Find the single use of def of insn. At the same time, +make sure def of insn is the single def which reach the use. */ + +static rtx_insn * +use_for_single_du (rtx_insn *insn) +{ + struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); + df_ref def; + + FOR_EACH_INSN_INFO_DEF (def, insn_info) +{ + struct df_link *chain = DF_REF_CHAIN (def); + + if (!chain + || chain->next) + return NULL; + + df_ref use = chain->ref; + chain = DF_REF_CHAIN (use); + + if (!chain + || chain->next) + return NULL; + + return DF
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 04/24/2015 10:18 AM, Jiong Wang wrote: 2015-04-21 Jiong Wang gcc/ * loop-invariant.c (find_defs): Enable DF_DU_CHAIN build. (vfp_const_iv): New hash table. (expensive_addr_check_p): New boolean. (init_inv_motion_data): Initialize new variables.> (free_inv_motion_data): Release hash table. (create_new_invariant): Set cheap_address to false for iv in vfp_const_iv table. (find_invariant_insn): Skip dependencies check for iv in vfp_const_iv table. (use_for_single_du): New function. (reshuffle_insn_with_vfp): Likewise. (find_invariants_bb): Call reshuffle_insn_with_vfp. gcc/testsuite/ * gcc.dg/pr62173.c: New testcase. So ultimately isn't this just a specialized version of reassociation of pointer arithmetic? And if so, don't we have all the usual problems around introducing overflow? ISTM if this is going to go forward (ie, we somehow convince ourselves that this kind of reassociation is OK), then should it be made to apply on pointers in general rather than restricting to stack, frame, virtual-frame? Jeff, Thanks for the review. This transformation is not reassociation of pointer arithmetic. ^^^ The idea of this patch is, given two instructions with variable value, we may get new instruction sequences with fixed value by reassociating their operands. And currently GCC only generate such instruction sequences for local array accessing as far as I known. Which is precisely reassociation of pointer arithmetic. Given: x = a + b y = *(x + c) You're generating: x = a + c y = *(x + b) That's reassociation of pointer arithmetic. For all kinds of reassociation we have to concern ourselves with adding overflow where it didn't already occur. Assuming a 32 bit architecture we could get overflow if A is 0x7fff, b is -4 and and c = 3 0x7fff + -4 = 0x7ffb 0x7ffb + 3 = 0x7ffe If you make the transformation you're suggesting we get 0x7fff + 3 = 0x8002 OVERFLOW 0x8002 - 4 = 0x7ffe Now if you always know pointers are unsigned, then the overflow is defined and you'd be OK. But that's a property of the target and one that's not well modeled within GCC (we have POINTER_EXTEND_UNSIGNED which kind of tells us something in this space). In addition to worrying about overflow, you have to worry about segmented architectures with implicit segment selection -- especially if the segment selection comes from the base register than the entire effective address. On such architectures a + b != b + a when used in a memory reference. The HPPA and mn103 ports immediately come to mind. I'm sure there's at least one more since I recall helping someone with these issues in the past. Anyway So given x = a + b; y = *(x + c) Can only be transformed into x = a + c y = *(x + b) If and only if a + b would choose the same segment as a + c. On the PA that wouldn't be true for something like a = 0x400 b = 0x10 c = -4 a + b is 0x4010 a + c is 0x3ffc Those select different segments. Even though the full address is the same (0x400c). Sadly we don't expose this aspect of target characteristics at all. To proceed, you'd probably need ways to conditionalize these transformations. ie, target hooks. Jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Jeff Law writes: > For all kinds of reassociation we have to concern ourselves with adding > overflow where it didn't already occur. Assuming a 32 bit architecture > we could get overflow if A is 0x7fff, b is -4 and and c = 3 > > 0x7fff + -4 = 0x7ffb > 0x7ffb + 3 = 0x7ffe > > > If you make the transformation you're suggesting we get > > 0x7fff + 3 = 0x8002 OVERFLOW > 0x8002 - 4 = 0x7ffe > > Now if you always know pointers are unsigned, then the overflow is > defined and you'd be OK. But that's a property of the target and one > that's not well modeled within GCC (we have POINTER_EXTEND_UNSIGNED > which kind of tells us something in this space). I see, understood, cool! Thanks for such detailed explanation. Above scenario do may happen for general pointer arith reassociation. One thing may make life easier as my reassociation is restricted within frame pointer. the "(plus (plus fp, index_reg) + const_off)" pattern was to address some variable on stack. index_reg, const_off were part of the stack offset of the variable. Reassociate them means reorder two parts of the stack offset. There may be way to prove the transformation will not add extra overflow risk, especially when the index_reg is unsigned. I understand for general pointer arith reassociation, there do have big risk, as the involved operands largely come from irrelevant instruction, no relationship between the values from those operands, we can deduce nothing. > > In addition to worrying about overflow, you have to worry about > segmented architectures with implicit segment selection -- especially if > the segment selection comes from the base register than the entire > effective address. > Hmm, understood! This let me recall something as dark as x86 segment descriptor in protecting mode... > > Sadly we don't expose this aspect of target characteristics at all. > > > To proceed, you'd probably need ways to conditionalize these > transformations. ie, target hooks. > > Jeff -- Regards, Jiong
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 05/14/2015 03:13 PM, Jiong Wang wrote: Jeff Law writes: For all kinds of reassociation we have to concern ourselves with adding overflow where it didn't already occur. Assuming a 32 bit architecture we could get overflow if A is 0x7fff, b is -4 and and c = 3 0x7fff + -4 = 0x7ffb 0x7ffb + 3 = 0x7ffe If you make the transformation you're suggesting we get 0x7fff + 3 = 0x8002 OVERFLOW 0x8002 - 4 = 0x7ffe Now if you always know pointers are unsigned, then the overflow is defined and you'd be OK. But that's a property of the target and one that's not well modeled within GCC (we have POINTER_EXTEND_UNSIGNED which kind of tells us something in this space). I see, understood, cool! Thanks for such detailed explanation. Above scenario do may happen for general pointer arith reassociation. One thing may make life easier as my reassociation is restricted within frame pointer. the "(plus (plus fp, index_reg) + const_off)" pattern was to address some variable on stack. index_reg, const_off were part of the stack offset of the variable. Reassociate them means reorder two parts of the stack offset. There may be way to prove the transformation will not add extra overflow risk, especially when the index_reg is unsigned. I understand for general pointer arith reassociation, there do have big risk, as the involved operands largely come from irrelevant instruction, no relationship between the values from those operands, we can deduce nothing. Given the special status of SP, FP and ARGP and a known constant part, we can probably do something here. More below... In addition to worrying about overflow, you have to worry about segmented architectures with implicit segment selection -- especially if the segment selection comes from the base register than the entire effective address. Hmm, understood! This let me recall something as dark as x86 segment descriptor in protecting mode... Possibly, I've actually never studied the segmented aspects of the x86. But I'm painfully familiar with the others mentioned :( My recollection for the segmented stuff on the PA is we only had a single guard page at both ends of the segment. So we only allowed an offset of +-4k when doing address reassociations in legitimize_address. This was possible because we had callouts from the right places in the RTL generators/optimizers to allow targets to rewrite address arithmetic. So we could naturally bury the target details away from the code generator/optimizers. So we could possibly parameterize the transformation around similar concepts. The design issue here is it's introducing more target dependencies in places where we've really wanted to avoid them. In theory the gimple optimizers are supposed to be target independent. Reality is some stuff bleeds into them (the one that's mentioned the most often is branch costing, but there's others). *If* we decide to go forward with using some target hooks here. I'd be tempted to do 2. One that's effective a tri-state. Full reassociation, limited reassociation, no reassociation. The second would bound the constants in the limited reassociation case. Thoughts? Jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Jeff Law writes: > On 05/14/2015 03:13 PM, Jiong Wang wrote: >> >> Jeff Law writes: >> >>> For all kinds of reassociation we have to concern ourselves with adding >>> overflow where it didn't already occur. Assuming a 32 bit architecture >>> we could get overflow if A is 0x7fff, b is -4 and and c = 3 >>> >>> 0x7fff + -4 = 0x7ffb >>> 0x7ffb + 3 = 0x7ffe >>> >>> >>> If you make the transformation you're suggesting we get >>> >>> 0x7fff + 3 = 0x8002 OVERFLOW >>> 0x8002 - 4 = 0x7ffe >>> >>> Now if you always know pointers are unsigned, then the overflow is >>> defined and you'd be OK. But that's a property of the target and one >>> that's not well modeled within GCC (we have POINTER_EXTEND_UNSIGNED >>> which kind of tells us something in this space). >> >> I see, understood, cool! Thanks for such detailed explanation. >> >> Above scenario do may happen for general pointer arith >> reassociation. >> >> One thing may make life easier as my reassociation is restricted within >> frame pointer. the "(plus (plus fp, index_reg) + const_off)" pattern was >> to address some variable on stack. index_reg, const_off were part of >> the stack offset of the variable. Reassociate them means reorder two >> parts of the stack offset. There may be way to prove the transformation >> will not add extra overflow risk, especially when the index_reg is >> unsigned. >> >> I understand for general pointer arith reassociation, there do have big >> risk, as the involved operands largely come from irrelevant instruction, >> no relationship between the values from those operands, we can deduce >> nothing. > Given the special status of SP, FP and ARGP and a known constant part, > we can probably do something here. More below... > > > >> >>> >>> In addition to worrying about overflow, you have to worry about >>> segmented architectures with implicit segment selection -- especially if >>> the segment selection comes from the base register than the entire >>> effective address. >>> >> >> Hmm, understood! >> >> This let me recall something as dark as x86 segment descriptor in protecting >> mode... > Possibly, I've actually never studied the segmented aspects of the x86. > But I'm painfully familiar with the others mentioned :( > > My recollection for the segmented stuff on the PA is we only had a > single guard page at both ends of the segment. So we only allowed an > offset of +-4k when doing address reassociations in legitimize_address. > This was possible because we had callouts from the right places in the > RTL generators/optimizers to allow targets to rewrite address > arithmetic. So we could naturally bury the target details away from the > code generator/optimizers. > > So we could possibly parameterize the transformation around similar > concepts. The design issue here is it's introducing more target > dependencies in places where we've really wanted to avoid them. In > theory the gimple optimizers are supposed to be target independent. > Reality is some stuff bleeds into them (the one that's mentioned the > most often is branch costing, but there's others). > > *If* we decide to go forward with using some target hooks here. I'd be > tempted to do 2. One that's effective a tri-state. Full reassociation, > limited reassociation, no reassociation. The second would bound the > constants in the limited reassociation case. > > Thoughts? Thanks for these thoughts. I tried but still can't prove this transformation will not introduce extra pointer overflow even given it's reassociation with vfp, although my first impression is it do will not introduce extra risk in real application. Have done a quick check on hppa's legitimize_address. I see for (plus sym_ref, const_int), if const_int is beyond +-4K, then that hook will force them into register, then (plus reg, reg) is always OK. So for target hooks, my understanding of your idea is something like: new hook targetm.pointer_arith_reassociate (), if return -1 then support full reassociation, 0 for limited, 1 for should not do any reassociation. the default version return -1 as most targets are OK to do reassociation given we can prove there is no introducing of overflow risk. While for target like HPPA, we should define this hook to return 0 for limited support. Then, if targetm.pointer_arith_reassociate () return 1, we should further invoke the second hook targetm.limited_reassociate_p (rtx x), to check the reassociated rtx 'x' meets any restrictions, for example for HPPA, constants part shouldn't beyond +-4K. not sure whether my understanding is correct. -- Regards, Jiong
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 05/21/2015 02:46 PM, Jiong Wang wrote: Thanks for these thoughts. I tried but still can't prove this transformation will not introduce extra pointer overflow even given it's reassociation with vfp, although my first impression is it do will not introduce extra risk in real application. Have done a quick check on hppa's legitimize_address. I see for (plus sym_ref, const_int), if const_int is beyond +-4K, then that hook will force them into register, then (plus reg, reg) is always OK. I'm virtually certain the PA's legitimize_address is not overflow safe. It was written long before we started worrying about overflows in address computations. It was mostly concerned with trying generate good addressing modes without running afoul of the implicit space register selection issues. A SYMBOL_REF is always a valid base register. However, as the comment in hppa_legitimize_address notes, we might be given a MEM for something like: x[n-10]. We don't want to rewrite that as (x-10) + n, even though doing so would be beneficial for LICM. So for target hooks, my understanding of your idea is something like: new hook targetm.pointer_arith_reassociate (), if return -1 then support full reassociation, 0 for limited, 1 for should not do any reassociation. the default version return -1 as most targets are OK to do reassociation given we can prove there is no introducing of overflow risk. While for target like HPPA, we should define this hook to return 0 for limited support. Right. Rather than use magic constants, I'd suggest an enum for the tri-state. FULL_PTR_REASSOCIATION, PARTIAL_PTR_REASSOCIATION, NO_PTR_REASSOCIATION. Then, if targetm.pointer_arith_reassociate () return 1, we should further invoke the second hook targetm.limited_reassociate_p (rtx x), to check the reassociated rtx 'x' meets any restrictions, for example for HPPA, constants part shouldn't beyond +-4K. Right. Jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
Jeff Law writes: > On 05/21/2015 02:46 PM, Jiong Wang wrote: >> >> Thanks for these thoughts. >> >> I tried but still can't prove this transformation will not introduce >> extra pointer overflow even given it's reassociation with vfp, although >> my first impression is it do will not introduce extra risk in real >> application. >> >> Have done a quick check on hppa's legitimize_address. I see for (plus >> sym_ref, const_int), if const_int is beyond +-4K, then that hook will >> force them into register, then (plus reg, reg) is always OK. > I'm virtually certain the PA's legitimize_address is not overflow safe. > It was written long before we started worrying about overflows in > address computations. It was mostly concerned with trying generate good > addressing modes without running afoul of the implicit space register > selection issues. > > A SYMBOL_REF is always a valid base register. However, as the comment > in hppa_legitimize_address notes, we might be given a MEM for something > like: x[n-10]. > > We don't want to rewrite that as (x-10) + n, even though doing so > would be beneficial for LICM. > > >> >> So for target hooks, my understanding of your idea is something like: >> >> new hook targetm.pointer_arith_reassociate (), if return -1 then >> support full reassociation, 0 for limited, 1 for should not do any >> reassociation. the default version return -1 as most targets are OK to >> do reassociation given we can prove there is no introducing of overflow >> risk. While for target like HPPA, we should define this hook to return >> 0 for limited support. > Right. Rather than use magic constants, I'd suggest an enum for the > tri-state. FULL_PTR_REASSOCIATION, PARTIAL_PTR_REASSOCIATION, > NO_PTR_REASSOCIATION. > > >> >> Then, if targetm.pointer_arith_reassociate () return 1, we should >> further invoke the second hook targetm.limited_reassociate_p (rtx x), >> to check the reassociated rtx 'x' meets any restrictions, for example >> for HPPA, constants part shouldn't beyond +-4K. > Right. > > Jeff For the record, after Bin's recent tree-ssa-ivopt improvement originated from PR62173, this patch is not benefitial anymore. I can't see such re-shuffling opportunites in RTL level anymore. This patch was trying to hoist those RTL sequences generated for local array base address which haven't been hoisted out of the loop at tree level, while now they are handled quite well by tree-ssa-ivopt. During both aarch64 and mips64 bootstrapping, optimization in this patch haven't been triggered while there were quite a few before Bin's tree level fix. I have stopped working on this patch. Thanks for those time spent on reviewing and discussing on this. -- Regards, Jiong
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 09/02/2015 07:36 AM, Jiong Wang wrote: For the record, after Bin's recent tree-ssa-ivopt improvement originated from PR62173, this patch is not benefitial anymore. It happens sometimes. I have stopped working on this patch. Thanks for those time spent on reviewing and discussing on this. No problem, thanks for testing after Bin's changes and letting me know that this has become a non-issue. Jeff
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: > For PR62173, the ideal solution is to resolve the problem on tree level > ivopt pass. > > While, apart from the tree level issue, PR 62173 also exposed another two > RTL level issues. > one of them is looks like we could improve RTL level loop invariant hoisting > by re-shuffle insns. > > for Seb's testcase > > void bar(int i) { > char A[10]; > int d = 0; > while (i > 0) > A[d++] = i--; > > while (d > 0) > foo(A[d--]); > } > > the insn sequences to calculate A[I]'s address looks like: > > (insn 76 75 77 22 (set (reg/f:DI 109) > (plus:DI (reg/f:DI 64 sfp) > (reg:DI 108 [ i ]))) seb-pop.c:8 84 {*adddi3_aarch64} > (expr_list:REG_DEAD (reg:DI 108 [ i ]) > (nil))) > (insn 77 76 78 22 (set (reg:SI 110 [ D.2633 ]) > (zero_extend:SI (mem/j:QI (plus:DI (reg/f:DI 109) > (const_int -16 [0xfff0])) [0 A S1 A8]))) seb-pop.c:8 76 > {*zero_extendqisi2_aarch64} > (expr_list:REG_DEAD (reg/f:DI 109) > (nil))) > > while for most RISC archs, reg + reg addressing is typical, so if we > re-shuffle > the instruction sequences into the following: > > (insn 96 94 97 22 (set (reg/f:DI 129) > (plus:DI (reg/f:DI 64 sfp) > (const_int -16 [0xfff0]))) seb-pop.c:8 84 {*adddi3_aarch64} > (nil)) > (insn 97 96 98 22 (set (reg:DI 130 [ i ]) > (sign_extend:DI (reg/v:SI 97 [ i ]))) seb-pop.c:8 70 > {*extendsidi2_aarch64} > (expr_list:REG_DEAD (reg/v:SI 97 [ i ]) > (nil))) > (insn 98 97 99 22 (set (reg:SI 131 [ D.2633 ]) > (zero_extend:SI (mem/j:QI (plus:DI (reg/f:DI 129) > (reg:DI 130 [ i ])) [0 A S1 A8]))) seb-pop.c:8 76 > {*zero_extendqisi2_aarch64} > (expr_list:REG_DEAD (reg:DI 130 [ i ]) > (expr_list:REG_DEAD (reg/f:DI 129) > (nil > > which means re-associate the constant imm with the virtual frame pointer. > > transform > > RA <- fixed_reg + RC > RD <- MEM (RA + const_offset) > > into: > > RA <- fixed_reg + const_offset > RD <- MEM (RA + RC) > > then RA <- fixed_reg + const_offset is actually loop invariant, so the later > RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate > what tree > level ivopts could not sort out. There is a LIM pass after gimple ivopts - if the invariantness is already visible there why not handle it there similar to the special-cases in rewrite_bittest and rewrite_reciprocal? And of course similar tricks could be applied on the RTL level to RTL invariant motion? Thanks, Richard. > and this patch only tries to re-shuffle instructions within single basic > block which > is a inner loop which is perf critical. > > I am reusing the loop info in fwprop because there is loop info and it's run > before > GCSE. > > verified on aarch64 and mips64, the array base address hoisted out of loop. > > bootstrap ok on x86-64 and aarch64. > > comments? > > thanks. > > gcc/ > PR62173 > fwprop.c (prepare_for_gcse_pre): New function. > (fwprop_done): Call it.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: > On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: >> For PR62173, the ideal solution is to resolve the problem on tree level >> ivopt pass. >> >> While, apart from the tree level issue, PR 62173 also exposed another two >> RTL level issues. >> one of them is looks like we could improve RTL level loop invariant hoisting >> by re-shuffle insns. >> >> for Seb's testcase >> >> void bar(int i) { >> char A[10]; >> int d = 0; >> while (i > 0) >> A[d++] = i--; >> >> while (d > 0) >> foo(A[d--]); >> } >> >> the insn sequences to calculate A[I]'s address looks like: >> >> (insn 76 75 77 22 (set (reg/f:DI 109) >> (plus:DI (reg/f:DI 64 sfp) >> (reg:DI 108 [ i ]))) seb-pop.c:8 84 {*adddi3_aarch64} >> (expr_list:REG_DEAD (reg:DI 108 [ i ]) >> (nil))) >> (insn 77 76 78 22 (set (reg:SI 110 [ D.2633 ]) >> (zero_extend:SI (mem/j:QI (plus:DI (reg/f:DI 109) >> (const_int -16 [0xfff0])) [0 A S1 A8]))) seb-pop.c:8 76 >> {*zero_extendqisi2_aarch64} >> (expr_list:REG_DEAD (reg/f:DI 109) >> (nil))) >> >> while for most RISC archs, reg + reg addressing is typical, so if we >> re-shuffle >> the instruction sequences into the following: >> >> (insn 96 94 97 22 (set (reg/f:DI 129) >> (plus:DI (reg/f:DI 64 sfp) >> (const_int -16 [0xfff0]))) seb-pop.c:8 84 {*adddi3_aarch64} >> (nil)) >> (insn 97 96 98 22 (set (reg:DI 130 [ i ]) >> (sign_extend:DI (reg/v:SI 97 [ i ]))) seb-pop.c:8 70 >> {*extendsidi2_aarch64} >> (expr_list:REG_DEAD (reg/v:SI 97 [ i ]) >> (nil))) >> (insn 98 97 99 22 (set (reg:SI 131 [ D.2633 ]) >> (zero_extend:SI (mem/j:QI (plus:DI (reg/f:DI 129) >> (reg:DI 130 [ i ])) [0 A S1 A8]))) seb-pop.c:8 76 >> {*zero_extendqisi2_aarch64} >> (expr_list:REG_DEAD (reg:DI 130 [ i ]) >> (expr_list:REG_DEAD (reg/f:DI 129) >> (nil >> >> which means re-associate the constant imm with the virtual frame pointer. >> >> transform >> >> RA <- fixed_reg + RC >> RD <- MEM (RA + const_offset) >> >> into: >> >> RA <- fixed_reg + const_offset >> RD <- MEM (RA + RC) >> >> then RA <- fixed_reg + const_offset is actually loop invariant, so the later >> RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate >> what tree >> level ivopts could not sort out. > > There is a LIM pass after gimple ivopts - if the invariantness is already > visible there why not handle it there similar to the special-cases in > rewrite_bittest and rewrite_reciprocal? > > And of course similar tricks could be applied on the RTL level to > RTL invariant motion? Oh, and the patch misses a testcase. > Thanks, > Richard. > >> and this patch only tries to re-shuffle instructions within single basic >> block which >> is a inner loop which is perf critical. >> >> I am reusing the loop info in fwprop because there is loop info and it's run >> before >> GCSE. >> >> verified on aarch64 and mips64, the array base address hoisted out of loop. >> >> bootstrap ok on x86-64 and aarch64. >> >> comments? >> >> thanks. >> >> gcc/ >> PR62173 >> fwprop.c (prepare_for_gcse_pre): New function. >> (fwprop_done): Call it.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 04/12/14 11:07, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: which means re-associate the constant imm with the virtual frame pointer. transform RA <- fixed_reg + RC RD <- MEM (RA + const_offset) into: RA <- fixed_reg + const_offset RD <- MEM (RA + RC) then RA <- fixed_reg + const_offset is actually loop invariant, so the later RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate what tree level ivopts could not sort out. There is a LIM pass after gimple ivopts - if the invariantness is already visible there why not handle it there similar to the special-cases in rewrite_bittest and rewrite_reciprocal? maybe, needs further check. And of course similar tricks could be applied on the RTL level to RTL invariant motion? Thanks. I just checked the code, yes, loop invariant motion pass is the natural place to integrate such multi-insns invariant analysis trick. those code could be integrated into loop-invariant.c cleanly, but then I found although the invariant detected successfully but it's not moved out of loop because of cost issue, and looks like the patch committed to fix PR33928 is trying to prevent such cheap address be hoisted to reduce register pressure. 805 /* ??? Try to determine cheapness of address computation. Unfortunately 806 the address cost is only a relative measure, we can't really compare 807 it with any absolute number, but only with other address costs. 808 But here we don't have any other addresses, so compare with a magic 809 number anyway. It has to be large enough to not regress PR33928 810 (by avoiding to move reg+8,reg+16,reg+24 invariants), but small 811 enough to not regress 410.bwaves either (by still moving reg+reg 812 invariants). 813 See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ 814 inv->cheap_address = address_cost (SET_SRC (set), word_mode, 815 ADDR_SPACE_GENERIC, speed) < 3; I think that maybe necessary for x86 which is short of register, while for RISC, it may not be that necessary, especially the whole register pressure is not big. currently, what I could think of is for this transformation below, we should increase the costs: A == RA <- virtual_stack_var + RC RD <- MEM (RA + const_offset) into: B == RA <- virtual_stack_var + const_offset <--B RD <- MEM (RA + RC) because the cost is not that cheap, if there is not re-assocation of virtual_stack_var with const_offset, then lra elimination will create another instruction to hold the elimination result, so format A will actually be RT <- real_stack_pointer + elimination_offset RA <- RT + RC RD <- MEM (RA + const_offset) so, the re-assocation and later hoisting of invariant B could actually save two instructions in the loop, this is why there are 15% perf gap for bzip2 under some situation. On aarch64, for Seb's testcase, before IV hoisting === b .L4 .L37: sub w19, w19, #1 .L4: add x1, x29, 48 add x0, x1, x0, sxtw ldrbw0, [x0, -16] bl foo mov w0, w19 cbnzw19, .L37 ldr x19, [sp, 16] ldp x29, x30, [sp], 48 ret after this transformation, and the IV hoisting (need one extra callee saved, x20) === .L3: add x20, x29, 32 <-- IV hoisted b .L4 .L37: sub w19, w19, #1 .L4: ldrbw0, [x20, w0, sxtw] bl foo mov w0, w19 cbnzw19, .L37 ldp x19, x20, [sp, 16] ldp x29, x30, [sp], 48 ret Oh, and the patch misses a testcase. It's at the start of the email, will include in the patch. void bar(int i) { char A[10]; int d = 0; while (i > 0) A[d++] = i--; while (d > 0) foo(A[d--]); } Thanks, Richard. and this patch only tries to re-shuffle instructions within single basic block which is a inner loop which is perf critical. I am reusing the loop info in fwprop because there is loop info and it's run before GCSE. verified on aarch64 and mips64, the array base address hoisted out of loop. bootstrap ok on x86-64 and aarch64. comments? thanks. gcc/ PR62173 fwprop.c (prepare_for_gcse_pre): New function. (fwprop_done): Call it.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 04/12/14 19:32, Jiong Wang wrote: On 04/12/14 11:07, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: which means re-associate the constant imm with the virtual frame pointer. transform RA <- fixed_reg + RC RD <- MEM (RA + const_offset) into: RA <- fixed_reg + const_offset RD <- MEM (RA + RC) then RA <- fixed_reg + const_offset is actually loop invariant, so the later RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate what tree level ivopts could not sort out. There is a LIM pass after gimple ivopts - if the invariantness is already visible there why not handle it there similar to the special-cases in rewrite_bittest and rewrite_reciprocal? maybe, needs further check. And of course similar tricks could be applied on the RTL level to RTL invariant motion? Thanks. I just checked the code, yes, loop invariant motion pass is the natural place to integrate such multi-insns invariant analysis trick. those code could be integrated into loop-invariant.c cleanly, but then I found although the invariant detected successfully but it's not moved out of loop because of cost issue, and looks like the patch committed to fix PR33928 is trying to prevent such cheap address be hoisted to reduce register pressure. 805 /* ??? Try to determine cheapness of address computation. Unfortunately 806 the address cost is only a relative measure, we can't really compare 807 it with any absolute number, but only with other address costs. 808 But here we don't have any other addresses, so compare with a magic 809 number anyway. It has to be large enough to not regress PR33928 810 (by avoiding to move reg+8,reg+16,reg+24 invariants), but small 811 enough to not regress 410.bwaves either (by still moving reg+reg 812 invariants). 813 See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ 814 inv->cheap_address = address_cost (SET_SRC (set), word_mode, 815 ADDR_SPACE_GENERIC, speed) < 3; I think that maybe necessary for x86 which is short of register, while for RISC, it may not be that necessary, especially the whole register pressure is not big. currently, what I could think of is for this transformation below, we should increase the costs: A == RA <- virtual_stack_var + RC RD <- MEM (RA + const_offset) into: B == RA <- virtual_stack_var + const_offset <--B RD <- MEM (RA + RC) because the cost is not that cheap, if there is not re-assocation of virtual_stack_var with const_offset, then lra elimination will create another instruction to hold the elimination result, so format A will actually be RT <- real_stack_pointer + elimination_offset RA <- RT + RC RD <- MEM (RA + const_offset) so, the re-assocation and later hoisting of invariant B could actually save two instructions in the loop, this is why there are 15% perf gap for bzip2 under some situation. updated patch. moved this instruction shuffling trick to rtl loop invariant pass. as described above, this patch tries to transform A to B, so that after the transformation: * RA <- virtual_stack_var + const_offset could be hoisted out of the loop * easy the work of lra elimination as virtual_stack_var is associated with const_offset that the elimination offset could be combined with const_offset automatically. current rtl loop invariant pass treat "reg <- reg + off" as cheap address, while although "reg <- virtual_stack_var + offset" fall into the same format, but it's not that cheap as we could also save one lra elimination instruction. so this patch will mark "reg <- virtual_stack_var + offset" transformed from A to be expensive, so that it could be hoisted later. after patch, pr62173 is fixed on powerpc64, while *still not on aarch64*. because there are one glitch in aarch64_legitimize_address which cause unnecessary complex instructions sequences generated when legitimize some addresses. and if we fix that, we will get cheaper address for those cases which is generally good, and the cheaper address will cause tree level IVOPT do more IVOPT optimization which is generally good also, but from the speck2k result, there are actually around 1% code size regression on two cases, the reason is for target support post-index address, doing IVOPT may not always be the best choice because we lost the advantage of using post-index addressing. on aarch64, for the following testcase, the ivopted version is complexer than not ivopted version. while (oargc--) *(nargv++) = *(oargv++); so, I sent the generic fix here only, as it's an independent patch, and could benefit other targets like powerpc64 although the issue on aarch64 is still not resolved before we figure out h
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 15/12/14 15:28, Jiong Wang wrote: On 04/12/14 19:32, Jiong Wang wrote: On 04/12/14 11:07, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: which means re-associate the constant imm with the virtual frame pointer. transform RA <- fixed_reg + RC RD <- MEM (RA + const_offset) into: RA <- fixed_reg + const_offset RD <- MEM (RA + RC) then RA <- fixed_reg + const_offset is actually loop invariant, so the later RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate what tree level ivopts could not sort out. There is a LIM pass after gimple ivopts - if the invariantness is already visible there why not handle it there similar to the special-cases in rewrite_bittest and rewrite_reciprocal? maybe, needs further check. And of course similar tricks could be applied on the RTL level to RTL invariant motion? Thanks. I just checked the code, yes, loop invariant motion pass is the natural place to integrate such multi-insns invariant analysis trick. those code could be integrated into loop-invariant.c cleanly, but then I found although the invariant detected successfully but it's not moved out of loop because of cost issue, and looks like the patch committed to fix PR33928 is trying to prevent such cheap address be hoisted to reduce register pressure. 805 /* ??? Try to determine cheapness of address computation. Unfortunately 806 the address cost is only a relative measure, we can't really compare 807 it with any absolute number, but only with other address costs. 808 But here we don't have any other addresses, so compare with a magic 809 number anyway. It has to be large enough to not regress PR33928 810 (by avoiding to move reg+8,reg+16,reg+24 invariants), but small 811 enough to not regress 410.bwaves either (by still moving reg+reg 812 invariants). 813 See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ 814 inv->cheap_address = address_cost (SET_SRC (set), word_mode, 815 ADDR_SPACE_GENERIC, speed) < 3; I think that maybe necessary for x86 which is short of register, while for RISC, it may not be that necessary, especially the whole register pressure is not big. currently, what I could think of is for this transformation below, we should increase the costs: A == RA <- virtual_stack_var + RC RD <- MEM (RA + const_offset) into: B == RA <- virtual_stack_var + const_offset <--B RD <- MEM (RA + RC) because the cost is not that cheap, if there is not re-assocation of virtual_stack_var with const_offset, then lra elimination will create another instruction to hold the elimination result, so format A will actually be RT <- real_stack_pointer + elimination_offset RA <- RT + RC RD <- MEM (RA + const_offset) so, the re-assocation and later hoisting of invariant B could actually save two instructions in the loop, this is why there are 15% perf gap for bzip2 under some situation. updated patch. moved this instruction shuffling trick to rtl loop invariant pass. as described above, this patch tries to transform A to B, so that after the transformation: * RA <- virtual_stack_var + const_offset could be hoisted out of the loop * easy the work of lra elimination as virtual_stack_var is associated with const_offset that the elimination offset could be combined with const_offset automatically. current rtl loop invariant pass treat "reg <- reg + off" as cheap address, while although "reg <- virtual_stack_var + offset" fall into the same format, but it's not that cheap as we could also save one lra elimination instruction. so this patch will mark "reg <- virtual_stack_var + offset" transformed from A to be expensive, so that it could be hoisted later. after patch, pr62173 is fixed on powerpc64, while *still not on aarch64*. because there are one glitch in aarch64_legitimize_address which cause unnecessary complex instructions sequences generated when legitimize some addresses. and if we fix that, we will get cheaper address for those cases which is generally good, and the cheaper address will cause tree level IVOPT do more IVOPT optimization which is generally good also, but from the speck2k result, there are actually around 1% code size regression on two cases, the reason is for target support post-index address, doing IVOPT may not always be the best choice because we lost the advantage of using post-index addressing. on aarch64, for the following testcase, the ivopted version is complexer than not ivopted version. while (oargc--) *(nargv++) = *(oargv++); so, I sent the generic fix here only, as it's an independent patch, and could benefit other targets like powerpc64 a
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Mon, Dec 15, 2014 at 4:29 PM, Jiong Wang wrote: > > On 15/12/14 15:28, Jiong Wang wrote: >> >> On 04/12/14 19:32, Jiong Wang wrote: >>> >>> On 04/12/14 11:07, Richard Biener wrote: >>> On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: > > On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: > > > which means re-associate the constant imm with the virtual frame > pointer. > > transform > > RA <- fixed_reg + RC > RD <- MEM (RA + const_offset) > > into: > > RA <- fixed_reg + const_offset > RD <- MEM (RA + RC) > > then RA <- fixed_reg + const_offset is actually loop invariant, so the > later > RTL GCSE PRE pass could catch it and do the hoisting, and thus > ameliorate > what tree > level ivopts could not sort out. > There is a LIM pass after gimple ivopts - if the invariantness is > already > visible there why not handle it there similar to the special-cases in > rewrite_bittest and rewrite_reciprocal? >>> >>> maybe, needs further check. >>> > And of course similar tricks could be applied on the RTL level to > RTL invariant motion? >>> >>> Thanks. I just checked the code, yes, loop invariant motion pass >>> is the natural place to integrate such multi-insns invariant analysis >>> trick. >>> >>> those code could be integrated into loop-invariant.c cleanly, but >>> then I found although the invariant detected successfully but it's not >>> moved >>> out of loop because of cost issue, and looks like the patch committed to >>> fix PR33928 >>> is trying to prevent such cheap address be hoisted to reduce register >>> pressure. >>> >>> >>> 805 /* ??? Try to determine cheapness of address computation. >>> Unfortunately >>> 806 the address cost is only a relative measure, we can't >>> really compare >>> 807 it with any absolute number, but only with other address >>> costs. >>> 808 But here we don't have any other addresses, so compare >>> with a magic >>> 809 number anyway. It has to be large enough to not regress >>> PR33928 >>> 810 (by avoiding to move reg+8,reg+16,reg+24 invariants), >>> but small >>> 811 enough to not regress 410.bwaves either (by still moving >>> reg+reg >>> 812 invariants). >>> 813 See >>> http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ >>> 814 inv->cheap_address = address_cost (SET_SRC (set), >>> word_mode, >>> 815 ADDR_SPACE_GENERIC, >>> speed) < 3; >>> >>> >>> I think that maybe necessary for x86 which is short of register, while >>> for RISC, >>> it may not be that necessary, especially the whole register pressure is >>> not big. >>> >>> currently, what I could think of is for this transformation below, we >>> should >>> increase the costs: >>> A >>> == >>> RA <- virtual_stack_var + RC >>> RD <- MEM (RA + const_offset) >>> >>> into: >>> >>> B >>> == >>> RA <- virtual_stack_var + const_offset <--B >>> RD <- MEM (RA + RC) >>> >>> because the cost is not that cheap, if there is not re-assocation of >>> virtual_stack_var >>> with const_offset, then lra elimination will create another instruction >>> to hold the >>> elimination result, so format A will actually be >>> >>> RT <- real_stack_pointer + elimination_offset >>> RA <- RT + RC >>> RD <- MEM (RA + const_offset) >>> >>> so, the re-assocation and later hoisting of invariant B could actually >>> save two instructions >>> in the loop, this is why there are 15% perf gap for bzip2 under some >>> situation. >> >> updated patch. >> >> moved this instruction shuffling trick to rtl loop invariant pass. >> as described above, this patch tries to transform A to B, so that >> after the transformation: >> >> * RA <- virtual_stack_var + const_offset could be hoisted out of the >> loop >> * easy the work of lra elimination as virtual_stack_var is associated >> with const_offset >> that the elimination offset could be combined with const_offset >> automatically. >> >> current rtl loop invariant pass treat "reg <- reg + off" as cheap address, >> while although >> "reg <- virtual_stack_var + offset" fall into the same format, but it's >> not that cheap as >> we could also save one lra elimination instruction. so this patch will >> mark >> "reg <- virtual_stack_var + offset" transformed from A to be expensive, so >> that it could be >> hoisted later. >> >> after patch, pr62173 is fixed on powerpc64, while *still not on aarch64*. >> because there are one >> glitch in aarch64_legitimize_address which cause unnecessary complex >> instructions sequences >> generated when legitimize some addresses. and if we fix that, we will get >> cheaper address for >> those cases which is generally good, and the cheaper address will cause >> tree le
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 17/12/14 15:54, Richard Biener wrote: On Mon, Dec 15, 2014 at 4:29 PM, Jiong Wang wrote: On 15/12/14 15:28, Jiong Wang wrote: On 04/12/14 19:32, Jiong Wang wrote: On 04/12/14 11:07, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:07 PM, Richard Biener wrote: On Thu, Dec 4, 2014 at 12:00 PM, Jiong Wang wrote: which means re-associate the constant imm with the virtual frame pointer. transform RA <- fixed_reg + RC RD <- MEM (RA + const_offset) into: RA <- fixed_reg + const_offset RD <- MEM (RA + RC) then RA <- fixed_reg + const_offset is actually loop invariant, so the later RTL GCSE PRE pass could catch it and do the hoisting, and thus ameliorate what tree level ivopts could not sort out. There is a LIM pass after gimple ivopts - if the invariantness is already visible there why not handle it there similar to the special-cases in rewrite_bittest and rewrite_reciprocal? maybe, needs further check. And of course similar tricks could be applied on the RTL level to RTL invariant motion? Thanks. I just checked the code, yes, loop invariant motion pass is the natural place to integrate such multi-insns invariant analysis trick. those code could be integrated into loop-invariant.c cleanly, but then I found although the invariant detected successfully but it's not moved out of loop because of cost issue, and looks like the patch committed to fix PR33928 is trying to prevent such cheap address be hoisted to reduce register pressure. 805 /* ??? Try to determine cheapness of address computation. Unfortunately 806 the address cost is only a relative measure, we can't really compare 807 it with any absolute number, but only with other address costs. 808 But here we don't have any other addresses, so compare with a magic 809 number anyway. It has to be large enough to not regress PR33928 810 (by avoiding to move reg+8,reg+16,reg+24 invariants), but small 811 enough to not regress 410.bwaves either (by still moving reg+reg 812 invariants). 813 See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ 814 inv->cheap_address = address_cost (SET_SRC (set), word_mode, 815 ADDR_SPACE_GENERIC, speed) < 3; I think that maybe necessary for x86 which is short of register, while for RISC, it may not be that necessary, especially the whole register pressure is not big. currently, what I could think of is for this transformation below, we should increase the costs: A == RA <- virtual_stack_var + RC RD <- MEM (RA + const_offset) into: B == RA <- virtual_stack_var + const_offset <--B RD <- MEM (RA + RC) because the cost is not that cheap, if there is not re-assocation of virtual_stack_var with const_offset, then lra elimination will create another instruction to hold the elimination result, so format A will actually be RT <- real_stack_pointer + elimination_offset RA <- RT + RC RD <- MEM (RA + const_offset) so, the re-assocation and later hoisting of invariant B could actually save two instructions in the loop, this is why there are 15% perf gap for bzip2 under some situation. updated patch. moved this instruction shuffling trick to rtl loop invariant pass. as described above, this patch tries to transform A to B, so that after the transformation: * RA <- virtual_stack_var + const_offset could be hoisted out of the loop * easy the work of lra elimination as virtual_stack_var is associated with const_offset that the elimination offset could be combined with const_offset automatically. current rtl loop invariant pass treat "reg <- reg + off" as cheap address, while although "reg <- virtual_stack_var + offset" fall into the same format, but it's not that cheap as we could also save one lra elimination instruction. so this patch will mark "reg <- virtual_stack_var + offset" transformed from A to be expensive, so that it could be hoisted later. after patch, pr62173 is fixed on powerpc64, while *still not on aarch64*. because there are one glitch in aarch64_legitimize_address which cause unnecessary complex instructions sequences generated when legitimize some addresses. and if we fix that, we will get cheaper address for those cases which is generally good, and the cheaper address will cause tree level IVOPT do more IVOPT optimization which is generally good also, but from the speck2k result, there are actually around 1% code size regression on two cases, the reason is for target support post-index address, doing IVOPT may not always be the best choice because we lost the advantage of using post-index addressing. on aarch64, for the following testcase, the ivopted version is complexer than not ivopted version. while (oargc--) *(nargv++) = *(oargv++); so, I sent the gene
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2014-12-18 17:00 GMT+00:00 Jiong Wang : ok for trunk? gcc/ PR62173 loop-invariant.c.c (expensive_addr): New hash_table. (need_expensive_addr_check_p): New bool. (find_exits): Rename to "find_exists_and_reshuffle. Support re-shuffle instructions for better loop invariant hoisting. another question is, is it safe to re-use REG_DEAD info here without calling df_note_add_problem and df_analysis first? am I using those info passed down from the previous pass which calculated these info and maybe broken?
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Thu, Dec 18, 2014 at 05:00:01PM +, Jiong Wang wrote: > On 17/12/14 15:54, Richard Biener wrote: > >ick. I realize we don't have SSA form on RTL but doesn't DF provide > >at least some help in looking up definition statements for pseudos? > >In fact we want to restrict the transform to single-use pseudos, thus > >hopefully it can at least tell us that... (maybe not and this is what > >LOG_LINKS are for in combine...?) At least loop-invariant alreadly > >computes df_chain with DF_UD_CHAIN which seems exactly what > >is needed (apart from maybe getting at single-use info). > > thanks very much for these inspiring questions. > > yes, we want to restrict the transformation on single-use pseudo only, > and it's better the transformation could re-use existed info and helper > function to avoid increase compile time. but I haven't found anything I > can reuse at the stage the transformation happen. > > the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study > the code about build LOG_LINKS, and try to see if we can do some factor out. LOG_LINKs in combine are just historical. combine should be converted to use DF fully. LOG_LINKs have nothing to do with single use; they point from the _first_ use to its corresponding def. You might want to look at what fwprop does instead. Segher
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Fri, Dec 19, 2014 at 6:09 AM, Segher Boessenkool wrote: > On Thu, Dec 18, 2014 at 05:00:01PM +, Jiong Wang wrote: >> On 17/12/14 15:54, Richard Biener wrote: >> >ick. I realize we don't have SSA form on RTL but doesn't DF provide >> >at least some help in looking up definition statements for pseudos? >> >In fact we want to restrict the transform to single-use pseudos, thus >> >hopefully it can at least tell us that... (maybe not and this is what >> >LOG_LINKS are for in combine...?) At least loop-invariant alreadly >> >computes df_chain with DF_UD_CHAIN which seems exactly what >> >is needed (apart from maybe getting at single-use info). >> >> thanks very much for these inspiring questions. >> >> yes, we want to restrict the transformation on single-use pseudo only, >> and it's better the transformation could re-use existed info and helper >> function to avoid increase compile time. but I haven't found anything I >> can reuse at the stage the transformation happen. >> >> the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study >> the code about build LOG_LINKS, and try to see if we can do some factor out. > > LOG_LINKs in combine are just historical. combine should be converted > to use DF fully. > > LOG_LINKs have nothing to do with single use; they point from the _first_ > use to its corresponding def. > > You might want to look at what fwprop does instead. Pass rtl fwprop uses df information in single-definition way, it doesn't really take into consideration if register is a single use. This often corrupts other optimizations like post-increment and load/store pair. For example: add r2, r1, r0 ldr rx, [r2] add r2, r2, #4 is transformed into below form: add r2, r1, r0 ldr rx, [r1, r0] add r2, r2, #4 As a result, post-increment opportunity is corrupted, also definition of r2 can't be deleted because it's not single use. Thanks, bin
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
2014-12-19 3:51 GMT+00:00 Bin.Cheng : > On Fri, Dec 19, 2014 at 6:09 AM, Segher Boessenkool > wrote: >> On Thu, Dec 18, 2014 at 05:00:01PM +, Jiong Wang wrote: >>> On 17/12/14 15:54, Richard Biener wrote: >>> >ick. I realize we don't have SSA form on RTL but doesn't DF provide >>> >at least some help in looking up definition statements for pseudos? >>> >In fact we want to restrict the transform to single-use pseudos, thus >>> >hopefully it can at least tell us that... (maybe not and this is what >>> >LOG_LINKS are for in combine...?) At least loop-invariant alreadly >>> >computes df_chain with DF_UD_CHAIN which seems exactly what >>> >is needed (apart from maybe getting at single-use info). >>> >>> thanks very much for these inspiring questions. >>> >>> yes, we want to restrict the transformation on single-use pseudo only, >>> and it's better the transformation could re-use existed info and helper >>> function to avoid increase compile time. but I haven't found anything I >>> can reuse at the stage the transformation happen. >>> >>> the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study >>> the code about build LOG_LINKS, and try to see if we can do some factor out. >> >> LOG_LINKs in combine are just historical. combine should be converted >> to use DF fully. >> >> LOG_LINKs have nothing to do with single use; they point from the _first_ >> use to its corresponding def. >> >> You might want to look at what fwprop does instead. > Pass rtl fwprop uses df information in single-definition way, it > doesn't really take into consideration if register is a single use. > This often corrupts other optimizations like post-increment and > load/store pair. For example: > > add r2, r1, r0 > ldr rx, [r2] > add r2, r2, #4 > is transformed into below form: > add r2, r1, r0 > ldr rx, [r1, r0] > add r2, r2, #4 > > As a result, post-increment opportunity is corrupted, also definition > of r2 can't be deleted because it's not single use. > > Thanks, > bin thanks for all these suggestion. Have look at the LOG_LINK build function, a simple reverse scan, while needs to allocate big map array for all pseudo regs. Haven't looked at similar code in fwprop, actually, when found the first matched insn pattern, I just want to scan several insns next, then abort quickly if nothing meet requirement. there is no need to build full single-use information. still can anyone confirm that it is safe to re-use REG_DEAD info there without calling df_note_add_problem and df_analysis first? or I am using those info passed down from the previous pass which calculated these info and maybe broken?
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Fri, Dec 19, 2014 at 11:28 AM, Jiong Wang wrote: > 2014-12-19 3:51 GMT+00:00 Bin.Cheng : >> On Fri, Dec 19, 2014 at 6:09 AM, Segher Boessenkool >> wrote: >>> On Thu, Dec 18, 2014 at 05:00:01PM +, Jiong Wang wrote: On 17/12/14 15:54, Richard Biener wrote: >ick. I realize we don't have SSA form on RTL but doesn't DF provide >at least some help in looking up definition statements for pseudos? >In fact we want to restrict the transform to single-use pseudos, thus >hopefully it can at least tell us that... (maybe not and this is what >LOG_LINKS are for in combine...?) At least loop-invariant alreadly >computes df_chain with DF_UD_CHAIN which seems exactly what >is needed (apart from maybe getting at single-use info). thanks very much for these inspiring questions. yes, we want to restrict the transformation on single-use pseudo only, and it's better the transformation could re-use existed info and helper function to avoid increase compile time. but I haven't found anything I can reuse at the stage the transformation happen. the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study the code about build LOG_LINKS, and try to see if we can do some factor out. >>> >>> LOG_LINKs in combine are just historical. combine should be converted >>> to use DF fully. As noted loop-invariant already computes DF use-def chains. The question is whether it does compute single-use info when doing that (should be trivially possible). Kenny? >>> LOG_LINKs have nothing to do with single use; they point from the _first_ >>> use to its corresponding def. >>> >>> You might want to look at what fwprop does instead. >> Pass rtl fwprop uses df information in single-definition way, it >> doesn't really take into consideration if register is a single use. >> This often corrupts other optimizations like post-increment and >> load/store pair. For example: >> >> add r2, r1, r0 >> ldr rx, [r2] >> add r2, r2, #4 >> is transformed into below form: >> add r2, r1, r0 >> ldr rx, [r1, r0] >> add r2, r2, #4 >> >> As a result, post-increment opportunity is corrupted, also definition >> of r2 can't be deleted because it's not single use. >> >> Thanks, >> bin > > thanks for all these suggestion. > > Have look at the LOG_LINK build function, a simple reverse scan, while > needs to allocate big map array for all pseudo regs. Haven't looked at > similar code in fwprop, > > actually, when found the first matched insn pattern, I just want to > scan several insns next, then abort quickly if nothing meet > requirement. there is no need to build full single-use information. > > still can anyone confirm that it is safe to re-use REG_DEAD info there > without calling df_note_add_problem and df_analysis first? or I am > using those info passed down from the previous pass which calculated > these info and maybe broken? It's not safe to use REG_DEAD info without re-computing it. Richard.
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
> still can anyone confirm that it is safe to re-use REG_DEAD info there > without calling df_note_add_problem and df_analysis first? or I am > using those info passed down from the previous pass which calculated > these info and maybe broken? It is generally _not_ safe to consume REG_UNUSED and REG_DEAD notes without recomputing them first. fwprop.c recomputes them on entry through the call to build_single_def_use_links (not clear why since it doesn't seem to use them) but it can presumably invalidate them before calling fwprop_done. I think it would be better to avoid using them in fwprop.c if possible. -- Eric Botcazou
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On Fri, Dec 19, 2014 at 11:51:06AM +0800, Bin.Cheng wrote: > >> yes, we want to restrict the transformation on single-use pseudo only, > >> and it's better the transformation could re-use existed info and helper > >> function to avoid increase compile time. but I haven't found anything I > >> can reuse at the stage the transformation happen. > >> > >> the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study > >> the code about build LOG_LINKS, and try to see if we can do some factor > >> out. > > > > LOG_LINKs in combine are just historical. combine should be converted > > to use DF fully. > > > > LOG_LINKs have nothing to do with single use; they point from the _first_ > > use to its corresponding def. > > > > You might want to look at what fwprop does instead. > Pass rtl fwprop uses df information in single-definition way, it > doesn't really take into consideration if register is a single use. Sure, because that's not what fwprop is meant to do. It has all the information though, and combine (in the LOG_LINKs) does not: it does not encode single uses, it does not encode all uses, it does not encode all defs, it doesn't even look outside of a single basic block. Segher
Re: [PATCH] PR 62173, re-shuffle insns for RTL loop invariant hoisting
On 12/19/2014 06:26 AM, Richard Biener wrote: On Fri, Dec 19, 2014 at 11:28 AM, Jiong Wang wrote: 2014-12-19 3:51 GMT+00:00 Bin.Cheng : On Fri, Dec 19, 2014 at 6:09 AM, Segher Boessenkool wrote: On Thu, Dec 18, 2014 at 05:00:01PM +, Jiong Wang wrote: On 17/12/14 15:54, Richard Biener wrote: ick. I realize we don't have SSA form on RTL but doesn't DF provide at least some help in looking up definition statements for pseudos? In fact we want to restrict the transform to single-use pseudos, thus hopefully it can at least tell us that... (maybe not and this is what LOG_LINKS are for in combine...?) At least loop-invariant alreadly computes df_chain with DF_UD_CHAIN which seems exactly what is needed (apart from maybe getting at single-use info). thanks very much for these inspiring questions. yes, we want to restrict the transformation on single-use pseudo only, and it's better the transformation could re-use existed info and helper function to avoid increase compile time. but I haven't found anything I can reuse at the stage the transformation happen. the info similar as LOG_LINKS is what I want, but maybe simpler. I'd study the code about build LOG_LINKS, and try to see if we can do some factor out. LOG_LINKs in combine are just historical. combine should be converted to use DF fully. As noted loop-invariant already computes DF use-def chains. The question is whether it does compute single-use info when doing that (should be trivially possible). Kenny? In the US, (where i live), there is something called the "statute of limitations" which says that after 7 years you cannot be found accountable for anything except the most heinous of crimes. It has been more than 7 years since i did this. however, since i am a nice person loop-invariant solves the DF_UD_CHAIN which builds a data structure that connects each use with all of the defs that reach it. I believe that this is the opposite of what you want here. if you really need this, you need to also turn on the DF_DU_CHAIN which builds the opposite structure.Both structures can be space hogs, but they are both turned on in other places in the compiler so it is unlikely to be an issue. LOG_LINKs have nothing to do with single use; they point from the _first_ use to its corresponding def. You might want to look at what fwprop does instead. Pass rtl fwprop uses df information in single-definition way, it doesn't really take into consideration if register is a single use. This often corrupts other optimizations like post-increment and load/store pair. For example: add r2, r1, r0 ldr rx, [r2] add r2, r2, #4 is transformed into below form: add r2, r1, r0 ldr rx, [r1, r0] add r2, r2, #4 As a result, post-increment opportunity is corrupted, also definition of r2 can't be deleted because it's not single use. Thanks, bin thanks for all these suggestion. Have look at the LOG_LINK build function, a simple reverse scan, while needs to allocate big map array for all pseudo regs. Haven't looked at similar code in fwprop, actually, when found the first matched insn pattern, I just want to scan several insns next, then abort quickly if nothing meet requirement. there is no need to build full single-use information. still can anyone confirm that it is safe to re-use REG_DEAD info there without calling df_note_add_problem and df_analysis first? or I am using those info passed down from the previous pass which calculated these info and maybe broken? It's not safe to use REG_DEAD info without re-computing it. not sure that reg_dead is the right answer even if you did re-compute it. I believe you can have two parallel uses (on both sides of an if-then-else) for a single def (above the if then else) and have two REG_DEAD notes. Richard.