Re: [4.9 PATCH, alpha]: Switch alpha to LRA
2013/4/23 Steven Bosscher stevenb@gmail.com: On Mon, Apr 22, 2013 at 7:33 PM, Jeff Law wrote: On 04/22/2013 11:17 AM, Uros Bizjak wrote: On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Right. That would also eliminate the NOTE_INSN_EH_REGION notes bug (PR target/56858). I think it would be a shame to not enable LRA on alpha. It will only be another excuse to never let reload die, and it will hurt stability and reliability for Alpha EV6 in the long term as other targets switch over to LRA. Is it possible to add some EV4/EV5 splitters to work around this Alpha EV4/EV5 oddity? Even if it comes at a code quality cost, it's IMHO still better than tying the fate of apha to reload and vice versa.. Ciao! Steven How about using follow constraints? Index: alpha.c === --- alpha.c (revision 198216) +++ alpha.c (working copy) @@ -9871,6 +9871,9 @@ #undef TARGET_LEGITIMATE_ADDRESS_P #define TARGET_LEGITIMATE_ADDRESS_P alpha_legitimate_address_p +#undef TARGET_LRA_P +#define TARGET_LRA_P hook_bool_void_true + #undef TARGET_CONDITIONAL_REGISTER_USAGE #define TARGET_CONDITIONAL_REGISTER_USAGE alpha_conditional_register_usage Index: alpha.md === --- alpha.md(revision 198216) +++ alpha.md(working copy) @@ -4073,9 +4073,9 @@ (define_insn *movdi [(set (match_operand:DI 0 nonimmediate_operand - =r,r,r,r,r,r,r,r, m, *f,*f, Q, r,*f) + =r,r,r,r,r,r,r,r, m, *f,*f, m, r,*f) (match_operand:DI 1 input_operand - rJ,K,L,T,s,n,s,m,rJ,*fJ, Q,*f,*f, r))] + rJ,K,L,T,s,n,s,m,rJ,*fJ, m,*f,*f, r))] register_operand (operands[0], DImode) || reg_or_0_operand (operands[1], DImode) @ As Uros said, the 'Q' is ignored by LRA. The reason is that the predicate function normal_memory_operand() may change op to a memory location by using resolved_reload_operand(). However, if LRA is enabled, resolve_reload_operand() always returns original reg op because reload_in_progress would never be true, resulting reload loop in this case. So I guess using 'm' constraint instead of 'Q' is able to avoid such abnormal behavior, leaving all the reload jobs to LRA. IMHO that is a simplest solution. At least it passes the case in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 and successfully build libgcc. Best regards, jasonwucj
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
2013/4/23 Vladimir Makarov vmaka...@redhat.com: On 13-04-22 2:19 PM, Steven Bosscher wrote: On Mon, Apr 22, 2013 at 7:33 PM, Jeff Law wrote: On 04/22/2013 11:17 AM, Uros Bizjak wrote: On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Right. That would also eliminate the NOTE_INSN_EH_REGION notes bug (PR target/56858). I think it would be a shame to not enable LRA on alpha. It will only be another excuse to never let reload die, and it will hurt stability and reliability for Alpha EV6 in the long term as other targets switch over to LRA. Is it possible to add some EV4/EV5 splitters to work around this Alpha EV4/EV5 oddity? Even if it comes at a code quality cost, it's IMHO still better than tying the fate of apha to reload and vice versa.. I never tried alpha with LRA. So it is not assumed that LRA should work on alpha. But I am sure LRA can work for alpha if some efforts will be spent. Porting LRA to a new target always involves changes in .md and machine-dependent files. This process was even not started. Actually, Uros showed that Alpha will not require a lot of efforts as code in most cases is already generated successfully. I don't remember any target which I tried to port LRA in such a good shape at the beginning of LRA port process. I would like to share a little bit of my thought. :) We are porting a 32-bit target, 'nds32', for GCC since gcc-4.6, 4.7, and 4.8/trunk. During these years we experienced implementation trade-off between reload and LRA. Solving register allocation is NP-complete (coloring problem). Reload is further to fix up any unsatisfied operands in the instruction and trying to do everything at once. It is powerful, yet complicated. Numbers of target hooks/macros make us confused and hard to implement. LRA, however, its systematic design and clear flow make developers easy to use and trace code. The instruction constraints become primary source of the information. Even some rookie-GCC-developers like me or my team members are also capable of identifying potential allocation issues when developing a new target. (See Shiva Chen's patch http://gcc.gnu.org/ml/gcc/2013-04/msg00225.html http://gcc.gnu.org/ml/gcc-patches/2013-04/msg01350.html) Eventually we decide to enable LRA in our 'nds32' porting. So I believe porting a new target costs less efforts with LRA than that with reload. For existing targets, of course, the efforts depend on how tight it is to reload. :p Best regards, jasonwucj
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On Thu, Apr 25, 2013 at 10:04 AM, Chung-Ju Wu jasonw...@gmail.com wrote: * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Right. That would also eliminate the NOTE_INSN_EH_REGION notes bug (PR target/56858). I think it would be a shame to not enable LRA on alpha. It will only be another excuse to never let reload die, and it will hurt stability and reliability for Alpha EV6 in the long term as other targets switch over to LRA. Is it possible to add some EV4/EV5 splitters to work around this Alpha EV4/EV5 oddity? Even if it comes at a code quality cost, it's IMHO still better than tying the fate of apha to reload and vice versa.. Ciao! Steven How about using follow constraints? Index: alpha.c === --- alpha.c (revision 198216) +++ alpha.c (working copy) @@ -9871,6 +9871,9 @@ #undef TARGET_LEGITIMATE_ADDRESS_P #define TARGET_LEGITIMATE_ADDRESS_P alpha_legitimate_address_p +#undef TARGET_LRA_P +#define TARGET_LRA_P hook_bool_void_true + #undef TARGET_CONDITIONAL_REGISTER_USAGE #define TARGET_CONDITIONAL_REGISTER_USAGE alpha_conditional_register_usage Index: alpha.md === --- alpha.md(revision 198216) +++ alpha.md(working copy) @@ -4073,9 +4073,9 @@ (define_insn *movdi [(set (match_operand:DI 0 nonimmediate_operand - =r,r,r,r,r,r,r,r, m, *f,*f, Q, r,*f) + =r,r,r,r,r,r,r,r, m, *f,*f, m, r,*f) (match_operand:DI 1 input_operand - rJ,K,L,T,s,n,s,m,rJ,*fJ, Q,*f,*f, r))] + rJ,K,L,T,s,n,s,m,rJ,*fJ, m,*f,*f, r))] register_operand (operands[0], DImode) || reg_or_0_operand (operands[1], DImode) @ As Uros said, the 'Q' is ignored by LRA. The reason is that the predicate function normal_memory_operand() may change op to a memory location by using resolved_reload_operand(). However, if LRA is enabled, resolve_reload_operand() always returns original reg op because reload_in_progress would never be true, resulting reload loop in this case. So I guess using 'm' constraint instead of 'Q' is able to avoid such abnormal behavior, leaving all the reload jobs to LRA. IMHO that is a simplest solution. At least it passes the case in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 and successfully build libgcc. Yes, this patch will work up to building libstdc++, where it will fail with the same reason on movqi pattern (on non-BWX target). I suspect that QImode access is generated during LRA (where reload_in_progress is false!) and it directly generates movqi, which can't use memory operand. Maybe simply adding lra_in_progress to reload_in_progress would fix this problem. Uros.
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 2013-04-25 09:18, Uros Bizjak wrote: Yes, this patch will work up to building libstdc++, where it will fail with the same reason on movqi pattern (on non-BWX target). I suspect that QImode access is generated during LRA (where reload_in_progress is false!) and it directly generates movqi, which can't use memory operand. Maybe simply adding lra_in_progress to reload_in_progress would fix this problem. Possibly. I've wondered in the past if it wouldn't be better to spill QI/HImode values in SImode for pre-BWX targets. Back in the day that would have involved hacking reload, but now it would appear that SECONDARY_MEMORY_NEEDED_RTX plus a reload_[qh]i pattern would do the job. Of course, that macro itself is in reload.c... Is there some other way we're supposed to get the same effect from LRA? r~
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Uros.
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 04/22/2013 11:17 AM, Uros Bizjak wrote: On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? jeff
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On Mon, Apr 22, 2013 at 7:33 PM, Jeff Law l...@redhat.com wrote: Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Looking at [1], I would propose a cutoff point at ev6. IIUC, the core of the problem is with *movdi, *movhi and *movqi patterns, where moves to/from memory are blocked for FP (there remains Q constraint in the *movdi case, ignored by LRA) and integer regs. This situation gets eventually fixed during reload (please see various instances of reload_in_progress), which is somehow incompatible with LRA. [1] http://en.wikipedia.org/wiki/DEC_Alpha Uros.
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On Mon, Apr 22, 2013 at 7:33 PM, Jeff Law wrote: On 04/22/2013 11:17 AM, Uros Bizjak wrote: On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Right. That would also eliminate the NOTE_INSN_EH_REGION notes bug (PR target/56858). I think it would be a shame to not enable LRA on alpha. It will only be another excuse to never let reload die, and it will hurt stability and reliability for Alpha EV6 in the long term as other targets switch over to LRA. Is it possible to add some EV4/EV5 splitters to work around this Alpha EV4/EV5 oddity? Even if it comes at a code quality cost, it's IMHO still better than tying the fate of apha to reload and vice versa.. Ciao! Steven
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 13-04-22 2:19 PM, Steven Bosscher wrote: On Mon, Apr 22, 2013 at 7:33 PM, Jeff Law wrote: On 04/22/2013 11:17 AM, Uros Bizjak wrote: On Tue, Jan 29, 2013 at 12:34 AM, Richard Henderson r...@redhat.com wrote: On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. Unfortunately, alphas are much more tied to reload than it was hoped. While latest alphas (with FIX and BWX ISAs) survived transition to LRA without problems, further testing on ev4 and ev5 triggered various problems, one of them is PR57032 [1] that exposed rather unique way of handling aligned/nonaligned memory operands. The patch was reverted. I suspect that fixing older alphas to live with LRA would be quite involved task, and I guess nobody (including me) wants to spend considerable amount of time on a dying architecture. Consequently, this also means that alphas will die together with reload as far as gcc is concerned. [1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57032 Would it make sense to deprecate the older Alpha implementations without killing the modern ones? Right. That would also eliminate the NOTE_INSN_EH_REGION notes bug (PR target/56858). I think it would be a shame to not enable LRA on alpha. It will only be another excuse to never let reload die, and it will hurt stability and reliability for Alpha EV6 in the long term as other targets switch over to LRA. Is it possible to add some EV4/EV5 splitters to work around this Alpha EV4/EV5 oddity? Even if it comes at a code quality cost, it's IMHO still better than tying the fate of apha to reload and vice versa.. I never tried alpha with LRA. So it is not assumed that LRA should work on alpha. But I am sure LRA can work for alpha if some efforts will be spent. Porting LRA to a new target always involves changes in .md and machine-dependent files. This process was even not started. Actually, Uros showed that Alpha will not require a lot of efforts as code in most cases is already generated successfully. I don't remember any target which I tried to port LRA in such a good shape at the beginning of LRA port process.
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On Mon, Apr 22, 2013 at 9:17 PM, Vladimir Makarov vmaka...@redhat.com wrote: I never tried alpha with LRA. So it is not assumed that LRA should work on alpha. But I am sure LRA can work for alpha if some efforts will be spent. Porting LRA to a new target always involves changes in .md and machine-dependent files. This process was even not started. Actually, Uros showed that Alpha will not require a lot of efforts as code in most cases is already generated successfully. I don't remember any target which I tried to port LRA in such a good shape at the beginning of LRA port process. Vladimir, thanks for encouraging words, it looks that all hope is not lost yet. However, I would like to point out that I have tested only ev68 architecture, and apparently all the interesting stuff is in the way ev4 and ev5 are handled. I can certainly spend a reasonable time to bring the port up to life, but for now, I propose to revert the patch, until ev4 bootstrap is fixed. Uros.
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 13-04-22 3:35 PM, Uros Bizjak wrote: On Mon, Apr 22, 2013 at 9:17 PM, Vladimir Makarov vmaka...@redhat.com wrote: I never tried alpha with LRA. So it is not assumed that LRA should work on alpha. But I am sure LRA can work for alpha if some efforts will be spent. Porting LRA to a new target always involves changes in .md and machine-dependent files. This process was even not started. Actually, Uros showed that Alpha will not require a lot of efforts as code in most cases is already generated successfully. I don't remember any target which I tried to port LRA in such a good shape at the beginning of LRA port process. Vladimir, thanks for encouraging words, it looks that all hope is not lost yet. However, I would like to point out that I have tested only ev68 architecture, and apparently all the interesting stuff is in the way ev4 and ev5 are handled. I can certainly spend a reasonable time to bring the port up to life, but for now, I propose to revert the patch, until ev4 bootstrap is fixed. Sure, Uros. I agree. Thanks for trying alpha.
[4.9 PATCH, alpha]: Switch alpha to LRA
Hello! 2013-01-28 Uros Bizjak ubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? [1] http://gcc.gnu.org/ml/gcc-testresults/2013-01/msg02998.html Uros. Index: config/alpha/alpha.c === --- config/alpha/alpha.c(revision 195502) +++ config/alpha/alpha.c(working copy) @@ -9872,6 +9872,9 @@ #undef TARGET_LEGITIMATE_ADDRESS_P #define TARGET_LEGITIMATE_ADDRESS_P alpha_legitimate_address_p +#undef TARGET_LRA_P +#define TARGET_LRA_P hook_bool_void_true + #undef TARGET_CONDITIONAL_REGISTER_USAGE #define TARGET_CONDITIONAL_REGISTER_USAGE alpha_conditional_register_usage
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 01/28/2013 03:14 PM, Uros Bizjak wrote: 2013-01-28 Uros Bizjakubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? Yep. r~
Re: [4.9 PATCH, alpha]: Switch alpha to LRA
On 01/28/2013 04:14 PM, Uros Bizjak wrote: Hello! 2013-01-28 Uros Bizjak ubiz...@gmail.com * config/alpha/alpha.c (TARGET_LRA_P): New define. Bootstrapped and regression tested [1] on alphaev68-unknown-linux-gnu. OK for 4.9? [1] http://gcc.gnu.org/ml/gcc-testresults/2013-01/msg02998.html Can you attach this to PR 55996, the 4.9 pending patches metabug. Thanks, jeff