| Message ID | 20140204085918.GC12671@tucnak.redhat.com |
|---|---|
| State | New |
| Headers | show |
On Tue, 4 Feb 2014, Jakub Jelinek wrote: > Hi! > > Apparently ia64 has mulv8qi3 and addv8qi3 insns, but not ashlv8qi3 > nor vashlv8qi3, thus the vectorizer can create or vector generic lowering > keep a V8QImode multiplication by constant, but expand_mult if it is > multiplied by a scalar (CONST_VECTOR with the same CONST_INT everywhere) > attempts to expand it as left shift (if power of 8) or tests various costs > and attempts to emit it as combination of left shift and add etc. > I think there are no targets that have vector multiply but not vector add > for the same mode, and similarly it doesn't make sense to test for scalar > left shifts, those are pretty much assumed everywhere in the compiler, > so this patch just ensures we emit normal vector multiply if there is no > vector left shift (be it with scalar or vector shift count). > > Bootstrapped/regtested on i686-linux and Andreas has bootstrapped/regtested > it on ia64-linux. Ok for trunk? Ok. Thanks, Richard. > 2014-02-04 Jakub Jelinek <jakub@redhat.com> > > PR middle-end/59261 > * expmed.c (expand_mult): For MODE_VECTOR_INT multiplication > if there is no vashl<mode>3 or ashl<mode>3 insn, skip_synth. > > * gcc.dg/pr59261.c: New test. > > --- gcc/expmed.c.jj 2014-01-03 11:40:57.000000000 +0100 > +++ gcc/expmed.c 2014-02-03 19:06:30.459304401 +0100 > @@ -3136,6 +3136,14 @@ expand_mult (enum machine_mode mode, rtx > if (do_trapv) > goto skip_synth; > > + /* If mode is integer vector mode, check if the backend supports > + vector lshift (by scalar or vector) at all. If not, we can't use > + synthetized multiply. */ > + if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT > + && optab_handler (vashl_optab, mode) == CODE_FOR_nothing > + && optab_handler (ashl_optab, mode) == CODE_FOR_nothing) > + goto skip_synth; > + > /* These are the operations that are potentially turned into > a sequence of shifts and additions. */ > mode_bitsize = GET_MODE_UNIT_BITSIZE (mode); > --- gcc/testsuite/gcc.dg/pr59261.c.jj 2014-02-03 19:14:39.457797016 +0100 > +++ gcc/testsuite/gcc.dg/pr59261.c 2014-02-03 19:14:20.000000000 +0100 > @@ -0,0 +1,17 @@ > +/* PR middle-end/59261 */ > +/* { dg-do compile } */ > +/* { dg-options "-O2" } */ > + > +typedef signed char V __attribute__((vector_size (8))); > + > +void > +foo (V *a, V *b) > +{ > + *a = *b * 3; > +} > + > +void > +bar (V *a, V *b) > +{ > + *a = *b * 4; > +} > > Jakub > >
--- gcc/expmed.c.jj 2014-01-03 11:40:57.000000000 +0100 +++ gcc/expmed.c 2014-02-03 19:06:30.459304401 +0100 @@ -3136,6 +3136,14 @@ expand_mult (enum machine_mode mode, rtx if (do_trapv) goto skip_synth; + /* If mode is integer vector mode, check if the backend supports + vector lshift (by scalar or vector) at all. If not, we can't use + synthetized multiply. */ + if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT + && optab_handler (vashl_optab, mode) == CODE_FOR_nothing + && optab_handler (ashl_optab, mode) == CODE_FOR_nothing) + goto skip_synth; + /* These are the operations that are potentially turned into a sequence of shifts and additions. */ mode_bitsize = GET_MODE_UNIT_BITSIZE (mode); --- gcc/testsuite/gcc.dg/pr59261.c.jj 2014-02-03 19:14:39.457797016 +0100 +++ gcc/testsuite/gcc.dg/pr59261.c 2014-02-03 19:14:20.000000000 +0100 @@ -0,0 +1,17 @@ +/* PR middle-end/59261 */ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +typedef signed char V __attribute__((vector_size (8))); + +void +foo (V *a, V *b) +{ + *a = *b * 3; +} + +void +bar (V *a, V *b) +{ + *a = *b * 4; +}