Message ID | 20161214100147.GA3427@linux.vnet.ibm.com |
---|---|
State | New |
Headers | show |
On Wed, Dec 14, 2016 at 11:01:47AM +0100, Dominik Vogt wrote: > This is another micro-optimisation in change_zero_ext. If an > > (and (lshiftrt ... (N)) (M)) > > generated by change_zero_ext is equivalent to just > > (lshiftrt ... (N)) > > (because the AND constant selects the N rightmost bits of the > result), strip off the AND. > > _But_ I'm still not completely convinced whether this is a good > idea. It may become necessary to add md patterns to deal with > just the LSHIFTRT. On the other hand it saves the need for > another special case in change_zero_ext, and a less obvious, very > specific risbg pattern on s390 For PowerPC we should already have all such patterns with a "bare" shift (they can be created in other ways, too). > Bootstrapped and regression tested on s390x and s390. (Targets > with risbg-like instructions (Power, others?) may need some > tuning.) But, it is also possible I missed some. So please wait until I have tested it. > diff --git a/gcc/combine.c b/gcc/combine.c > index 19851a2..5ebf31c 100644 > --- a/gcc/combine.c > +++ b/gcc/combine.c > @@ -11280,8 +11280,13 @@ change_zero_ext (rtx pat) > else > continue; > > - wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > - x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > + if (!(GET_CODE (x) == LSHIFTRT > + && CONST_INT_P (XEXP (x, 1)) > + && size + INTVAL (XEXP (x, 1)) == GET_MODE_PRECISION (mode))) > + { > + wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > + x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > + } One could argue that this should have been an lshiftrt in the first place then, not a zero_ext*. Hrm. Segher
On Wed, Dec 14, 2016 at 01:32:48PM -0600, Segher Boessenkool wrote: > On Wed, Dec 14, 2016 at 11:01:47AM +0100, Dominik Vogt wrote: > > This is another micro-optimisation in change_zero_ext. If an > > > > (and (lshiftrt ... (N)) (M)) > > > > generated by change_zero_ext is equivalent to just > > > > (lshiftrt ... (N)) > > > > (because the AND constant selects the N rightmost bits of the > > result), strip off the AND. > > > > _But_ I'm still not completely convinced whether this is a good > > idea. It may become necessary to add md patterns to deal with > > just the LSHIFTRT. On the other hand it saves the need for > > another special case in change_zero_ext, and a less obvious, very > > specific risbg pattern on s390 > > For PowerPC we should already have all such patterns with a "bare" shift > (they can be created in other ways, too). > > > Bootstrapped and regression tested on s390x and s390. (Targets > > with risbg-like instructions (Power, others?) may need some > > tuning.) > > But, it is also possible I missed some. So please wait until I have > tested it. > > > > diff --git a/gcc/combine.c b/gcc/combine.c > > index 19851a2..5ebf31c 100644 > > --- a/gcc/combine.c > > +++ b/gcc/combine.c > > @@ -11280,8 +11280,13 @@ change_zero_ext (rtx pat) > > else > > continue; > > > > - wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > > - x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > > + if (!(GET_CODE (x) == LSHIFTRT > > + && CONST_INT_P (XEXP (x, 1)) > > + && size + INTVAL (XEXP (x, 1)) == GET_MODE_PRECISION (mode))) > > + { > > + wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > > + x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > > + } > > One could argue that this should have been an lshiftrt in the first place > then, not a zero_ext*. Hrm. This one void g2(ui64 *pl, i32 seed) { seed = 69607 * seed; pl[0] = (seed >> 8) & 0xffffff; } generates (zero_extract:DI (reg:SI 75 [ seed ]) (const_int 24 [0x18]) (const_int 0 [0])) on s390x. Ciao Dominik ^_^ ^_^
On Thu, Dec 15, 2016 at 09:55:52AM +0100, Dominik Vogt wrote: > > > Bootstrapped and regression tested on s390x and s390. (Targets > > > with risbg-like instructions (Power, others?) may need some > > > tuning.) > > > > But, it is also possible I missed some. So please wait until I have > > tested it. > > > > > diff --git a/gcc/combine.c b/gcc/combine.c > > > index 19851a2..5ebf31c 100644 > > > --- a/gcc/combine.c > > > +++ b/gcc/combine.c > > > @@ -11280,8 +11280,13 @@ change_zero_ext (rtx pat) > > > else > > > continue; > > > > > > - wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > > > - x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > > > + if (!(GET_CODE (x) == LSHIFTRT > > > + && CONST_INT_P (XEXP (x, 1)) > > > + && size + INTVAL (XEXP (x, 1)) == GET_MODE_PRECISION (mode))) > > > + { > > > + wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); > > > + x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); > > > + } > > > > One could argue that this should have been an lshiftrt in the first place > > then, not a zero_ext*. Hrm. > > This one > > void g2(ui64 *pl, i32 seed) > { > seed = 69607 * seed; > pl[0] = (seed >> 8) & 0xffffff; > } > > generates > > (zero_extract:DI (reg:SI 75 [ seed ]) > (const_int 24 [0x18]) > (const_int 0 [0])) > > on s390x. Ah, right, it changes mode as well. I see. Tested on powerpc64-linux {-m32,-m64}, no new failures. The patch is okay for trunk. Thanks! Segher
On Wed, Dec 14, 2016 at 11:01:47AM +0100, Dominik Vogt wrote: > gcc/ChangeLog-change_zero_ext-2 > > * combine.c (change_zero_ext): Skip generation of redundant AND. Applied. Thanks! -Andreas-
diff --git a/gcc/combine.c b/gcc/combine.c index 19851a2..5ebf31c 100644 --- a/gcc/combine.c +++ b/gcc/combine.c @@ -11280,8 +11280,13 @@ change_zero_ext (rtx pat) else continue; - wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); - x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); + if (!(GET_CODE (x) == LSHIFTRT + && CONST_INT_P (XEXP (x, 1)) + && size + INTVAL (XEXP (x, 1)) == GET_MODE_PRECISION (mode))) + { + wide_int mask = wi::mask (size, false, GET_MODE_PRECISION (mode)); + x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); + } SUBST (**iter, x); changed = true;