Message ID | SN2PR07MB27028DC0EDBA36D4C0E9671E839B0@SN2PR07MB2702.namprd07.prod.outlook.com |
---|---|
State | New |
Headers | show |
Hi Naveen, On 13/12/16 11:51, Hurugalawadi, Naveen wrote: > Hi Kyrill, > > Thanks for reviewing the patch and your useful comments. > >>> looks good to me if it has gone through the normal required >>> bootstrap and testing, but I can't approve. > Bootstrapped and Regression Tested on aarch64-thunderx-linux. > >>> The rest of the MD file uses the term AdvSIMD. Also, the instrurction >>> is CNT rather than "pop count". > Done. > >>> __builtin_popcount takes an unsigned int, so this should be >>> scanning for absence of popcountsi2 instead? > Done. > > Please find attached the modified patch as per review comments > and let me know if its okay for Stage-1 or current branch. This looks much better, thanks. I still have a minor nit about the testcase. +long +foo1 (int x) +{ + return __builtin_popcountl (x); +} On ILP32 systems this would still use the SImode patterns, so I suggest you use __builtin_popcountll and an unsigned long long return type to ensure you always exercise the 64-bit code. + +/* { dg-final { scan-assembler-not "popcount" } } */ This looks ok to me otherwise, but you'll need an ok from the aarch64 folk. Kyrill > Thanks, > Naveen
On Tue, Dec 13, 2016 at 11:59:36AM +0000, Kyrill Tkachov wrote: > Hi Naveen, > > On 13/12/16 11:51, Hurugalawadi, Naveen wrote: > >Hi Kyrill, > > > >Thanks for reviewing the patch and your useful comments. > > > >>>looks good to me if it has gone through the normal required > >>>bootstrap and testing, but I can't approve. > >Bootstrapped and Regression Tested on aarch64-thunderx-linux. > > > >>>The rest of the MD file uses the term AdvSIMD. Also, the instrurction > >>>is CNT rather than "pop count". > >Done. > > > >>>__builtin_popcount takes an unsigned int, so this should be > >>>scanning for absence of popcountsi2 instead? > >Done. > > > >Please find attached the modified patch as per review comments > >and let me know if its okay for Stage-1 or current branch. > > This looks much better, thanks. > I still have a minor nit about the testcase. > > +long > +foo1 (int x) > +{ > + return __builtin_popcountl (x); > +} > > On ILP32 systems this would still use the SImode patterns, so I suggest you use __builtin_popcountll and > an unsigned long long return type to ensure you always exercise the 64-bit code. > > + > +/* { dg-final { scan-assembler-not "popcount" } } */ > > > This looks ok to me otherwise, but you'll need an ok from the aarch64 folk. I didn't see a follow-up patch posted with Kyrill's comments addressed? Thanks, James
Hi James and Kyrill, Thanks for the review and comments on the patch. >> On ILP32 systems this would still use the SImode patterns, >> so I suggest you use __builtin_popcountll and >> an unsigned long long return type to ensure you always exercise the 64-bit code. Sorry for not commenting on this part. The issue is that code generates "__popcountdi2" for all the codes in testcase for LP64 and ILP32 variants. __builtin_popcount, __builtin_popcountl and __builtin_popcount. Hence, modified the patch to check for "popcount". Bootstrapped and regression tested on AArch64-Thunderx-Linux machine. Please find attached the modified patch and let us know if its okay? Thanks, Naveen
On Thu, Feb 02, 2017 at 04:03:42AM +0000, Hurugalawadi, Naveen wrote: > Hi James and Kyrill, > > Thanks for the review and comments on the patch. > > >> On ILP32 systems this would still use the SImode patterns, > >> so I suggest you use __builtin_popcountll and > >> an unsigned long long return type to ensure you always exercise the 64-bit code. > > Sorry for not commenting on this part. > The issue is that code generates "__popcountdi2" for all the codes in testcase > for LP64 and ILP32 variants. > __builtin_popcount, __builtin_popcountl and __builtin_popcount. > > Hence, modified the patch to check for "popcount". I don't understand this comment and how it relates to your updated patch (which still checks for CNT). > Bootstrapped and regression tested on AArch64-Thunderx-Linux machine. > > Please find attached the modified patch and let us know if its okay? Could you please clarify what you meant above? The patch looks fine (though possibly not for Stage 4), but I'm not sure of your intent. Thanks, James
On Thu, Feb 2, 2017 at 3:55 AM, James Greenhalgh <james.greenhalgh@arm.com> wrote: > On Thu, Feb 02, 2017 at 04:03:42AM +0000, Hurugalawadi, Naveen wrote: >> Hi James and Kyrill, >> >> Thanks for the review and comments on the patch. >> >> >> On ILP32 systems this would still use the SImode patterns, >> >> so I suggest you use __builtin_popcountll and >> >> an unsigned long long return type to ensure you always exercise the 64-bit code. >> >> Sorry for not commenting on this part. >> The issue is that code generates "__popcountdi2" for all the codes in testcase >> for LP64 and ILP32 variants. >> __builtin_popcount, __builtin_popcountl and __builtin_popcount. >> >> Hence, modified the patch to check for "popcount". > > I don't understand this comment and how it relates to your updated > patch (which still checks for CNT). Now I understand Kyrill's comments as Naveen and I did not understand before. What Naveen was trying to test was that there was no longer any call to __popcountdi2 or __popcountsi2 any more for all three: __builtin_popcount, __builtin_popcountl and __builtin_popcountll. Kyrill was asking him to change the test for __builtin_popcountl to __builtin_popcountll even though there was already one test in that file already. Now of course what should change still is the argument types to foo1/foo2 so they take long and long long respectively. He will post a new version of the patch soon. > >> Bootstrapped and regression tested on AArch64-Thunderx-Linux machine. >> >> Please find attached the modified patch and let us know if its okay? > > Could you please clarify what you meant above? The patch looks fine (though > possibly not for Stage 4), but I'm not sure of your intent. What he meant was he tested on aarch64-linux-gnu with no regressions. Is it ok for when stage 1 opens with the change of the testcase? Thanks, Andrew > > Thanks, > James
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md index 65eb326..0acb3f0 100644 --- a/gcc/config/aarch64/aarch64.md +++ b/gcc/config/aarch64/aarch64.md @@ -3785,6 +3785,39 @@ } ) +;; Pop count be done via the "CNT" instruction in AdvSIMD. +;; +;; MOV v.1d, x0 +;; CNT v1.8b, v.8b +;; ADDV b2, v1.8b +;; MOV w0, v2.b[0] + +(define_expand "popcount<mode>2" + [(match_operand:GPI 0 "register_operand") + (match_operand:GPI 1 "register_operand")] + "TARGET_SIMD" +{ + rtx v = gen_reg_rtx (V8QImode); + rtx v1 = gen_reg_rtx (V8QImode); + rtx r = gen_reg_rtx (QImode); + rtx in = operands[1]; + rtx out = operands[0]; + if(<MODE>mode == SImode) + { + rtx tmp; + tmp = gen_reg_rtx (DImode); + /* If we have SImode, zero extend to DImode, pop count does + not change if we have extra zeros. */ + emit_insn (gen_zero_extendsidi2 (tmp, in)); + in = tmp; + } + emit_move_insn (v, gen_lowpart (V8QImode, in)); + emit_insn (gen_popcountv8qi2 (v1, v)); + emit_insn (gen_reduc_plus_scal_v8qi (r, v1)); + emit_insn (gen_zero_extendqi<mode>2 (out, r)); + DONE; +}) + (define_insn "clrsb<mode>2" [(set (match_operand:GPI 0 "register_operand" "=r") (clrsb:GPI (match_operand:GPI 1 "register_operand" "r")))] diff --git a/gcc/testsuite/gcc.target/aarch64/popcnt.c b/gcc/testsuite/gcc.target/aarch64/popcnt.c new file mode 100644 index 0000000..37cf4b9 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/popcnt.c @@ -0,0 +1,17 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +int +foo (int x) +{ + return __builtin_popcount (x); +} + +long +foo1 (int x) +{ + return __builtin_popcountl (x); +} + +/* { dg-final { scan-assembler-not "popcount" } } */ +/* { dg-final { scan-assembler-times "cnt\t" 2 } } */