Message ID | 20151112160843.GG51435@msticlxl57.ims.intel.com |
---|---|
State | New |
Headers | show |
On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: > Hi, > > When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. Confuses aka fails to vectorize? > This patch transforms such conversions into comparison. > > I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: > > real*8 a(18) > logical b(18) > integer i > > do i=1,18 > if(a(i).gt.0.d0) then > b(i)=.true. > else > b(i)=.false. > endif > enddo This looks the the "error" comes from if-conversion - can't we do better there then? Richard. > Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? > > Thanks, > Ilya > -- > gcc/ > > 2015-11-12 Ilya Enkovich <enkovich.gnu@gmail.com> > > * tree-vect-patterns.c (vect_recog_mask_conversion_pattern): > Transform useless boolean conversion into assignment. > > > diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c > index b9d900c..62070da 100644 > --- a/gcc/tree-vect-patterns.c > +++ b/gcc/tree-vect-patterns.c > @@ -3674,6 +3674,38 @@ vect_recog_mask_conversion_pattern (vec<gimple *> *stmts, tree *type_in, > if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE) > return NULL; > > + /* Check conversion between boolean types of different sizes. > + If no vectype is specified, then we have a regular mask > + assignment with no actual conversion. */ > + if (rhs_code == CONVERT_EXPR > + && !STMT_VINFO_DATA_REF (stmt_vinfo) > + && !STMT_VINFO_VECTYPE (stmt_vinfo)) > + { > + if (TREE_CODE (rhs1) != SSA_NAME) > + return NULL; > + > + rhs1_type = search_type_for_mask (rhs1, vinfo); > + if (!rhs1_type) > + return NULL; > + > + vectype1 = get_mask_type_for_scalar_type (rhs1_type); > + > + if (!vectype1) > + return NULL; > + > + lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL); > + pattern_stmt = gimple_build_assign (lhs, rhs1); > + > + *type_out = vectype1; > + *type_in = vectype1; > + stmts->safe_push (last_stmt); > + if (dump_enabled_p ()) > + dump_printf_loc (MSG_NOTE, vect_location, > + "vect_recog_mask_conversion_pattern: detected:\n"); > + > + return pattern_stmt; > + } > + > if (rhs_code != BIT_IOR_EXPR > && rhs_code != BIT_XOR_EXPR > && rhs_code != BIT_AND_EXPR)
2015-11-13 13:03 GMT+03:00 Richard Biener <richard.guenther@gmail.com>: > On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: >> Hi, >> >> When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. > > Confuses aka fails to vectorize? Right. > >> This patch transforms such conversions into comparison. >> >> I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: >> >> real*8 a(18) >> logical b(18) >> integer i >> >> do i=1,18 >> if(a(i).gt.0.d0) then >> b(i)=.true. >> else >> b(i)=.false. >> endif >> enddo > > This looks the the "error" comes from if-conversion - can't we do > better there then? No, this loop is transformed into a single BB before if-conversion by cselim + phiopt. Ilya > > Richard. > >> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? >> >> Thanks, >> Ilya
Ping 2015-11-13 16:17 GMT+03:00 Ilya Enkovich <enkovich.gnu@gmail.com>: > 2015-11-13 13:03 GMT+03:00 Richard Biener <richard.guenther@gmail.com>: >> On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: >>> Hi, >>> >>> When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. >> >> Confuses aka fails to vectorize? > > Right. > >> >>> This patch transforms such conversions into comparison. >>> >>> I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: >>> >>> real*8 a(18) >>> logical b(18) >>> integer i >>> >>> do i=1,18 >>> if(a(i).gt.0.d0) then >>> b(i)=.true. >>> else >>> b(i)=.false. >>> endif >>> enddo >> >> This looks the the "error" comes from if-conversion - can't we do >> better there then? > > No, this loop is transformed into a single BB before if-conversion by > cselim + phiopt. > > Ilya > >> >> Richard. >> >>> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? >>> >>> Thanks, >>> Ilya
Ping 2015-11-23 16:05 GMT+03:00 Ilya Enkovich <enkovich.gnu@gmail.com>: > Ping > > 2015-11-13 16:17 GMT+03:00 Ilya Enkovich <enkovich.gnu@gmail.com>: >> 2015-11-13 13:03 GMT+03:00 Richard Biener <richard.guenther@gmail.com>: >>> On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: >>>> Hi, >>>> >>>> When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. >>> >>> Confuses aka fails to vectorize? >> >> Right. >> >>> >>>> This patch transforms such conversions into comparison. >>>> >>>> I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: >>>> >>>> real*8 a(18) >>>> logical b(18) >>>> integer i >>>> >>>> do i=1,18 >>>> if(a(i).gt.0.d0) then >>>> b(i)=.true. >>>> else >>>> b(i)=.false. >>>> endif >>>> enddo >>> >>> This looks the the "error" comes from if-conversion - can't we do >>> better there then? >> >> No, this loop is transformed into a single BB before if-conversion by >> cselim + phiopt. >> >> Ilya >> >>> >>> Richard. >>> >>>> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? >>>> >>>> Thanks, >>>> Ilya
On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: > Hi, > > When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. This patch transforms such conversions into comparison. > > I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: > > real*8 a(18) > logical b(18) > integer i > > do i=1,18 > if(a(i).gt.0.d0) then > b(i)=.true. > else > b(i)=.false. > endif > enddo > > Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? > > Thanks, > Ilya > -- > gcc/ > > 2015-11-12 Ilya Enkovich <enkovich.gnu@gmail.com> > > * tree-vect-patterns.c (vect_recog_mask_conversion_pattern): > Transform useless boolean conversion into assignment. > > > diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c > index b9d900c..62070da 100644 > --- a/gcc/tree-vect-patterns.c > +++ b/gcc/tree-vect-patterns.c > @@ -3674,6 +3674,38 @@ vect_recog_mask_conversion_pattern (vec<gimple *> *stmts, tree *type_in, > if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE) > return NULL; > > + /* Check conversion between boolean types of different sizes. > + If no vectype is specified, then we have a regular mask > + assignment with no actual conversion. */ > + if (rhs_code == CONVERT_EXPR CONVERT_EXPR_CODE_P (rhs_code) > + && !STMT_VINFO_DATA_REF (stmt_vinfo) > + && !STMT_VINFO_VECTYPE (stmt_vinfo)) > + { > + if (TREE_CODE (rhs1) != SSA_NAME) > + return NULL; > + > + rhs1_type = search_type_for_mask (rhs1, vinfo); > + if (!rhs1_type) > + return NULL; > + > + vectype1 = get_mask_type_for_scalar_type (rhs1_type); > + > + if (!vectype1) > + return NULL; > + > + lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL); > + pattern_stmt = gimple_build_assign (lhs, rhs1); So what's the actual issue here? That the conversion is spurious? Why can't you accept this simply in vectorizable_assignment then? Richard. > + *type_out = vectype1; > + *type_in = vectype1; > + stmts->safe_push (last_stmt); > + if (dump_enabled_p ()) > + dump_printf_loc (MSG_NOTE, vect_location, > + "vect_recog_mask_conversion_pattern: detected:\n"); > + > + return pattern_stmt; > + } > + > if (rhs_code != BIT_IOR_EXPR > && rhs_code != BIT_XOR_EXPR > && rhs_code != BIT_AND_EXPR)
2015-12-02 17:52 GMT+03:00 Richard Biener <richard.guenther@gmail.com>: > On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: >> Hi, >> >> When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. This patch transforms such conversions into comparison. >> >> I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: >> >> real*8 a(18) >> logical b(18) >> integer i >> >> do i=1,18 >> if(a(i).gt.0.d0) then >> b(i)=.true. >> else >> b(i)=.false. >> endif >> enddo >> >> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? >> >> Thanks, >> Ilya >> -- >> gcc/ >> >> 2015-11-12 Ilya Enkovich <enkovich.gnu@gmail.com> >> >> * tree-vect-patterns.c (vect_recog_mask_conversion_pattern): >> Transform useless boolean conversion into assignment. >> >> >> diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c >> index b9d900c..62070da 100644 >> --- a/gcc/tree-vect-patterns.c >> +++ b/gcc/tree-vect-patterns.c >> @@ -3674,6 +3674,38 @@ vect_recog_mask_conversion_pattern (vec<gimple *> *stmts, tree *type_in, >> if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE) >> return NULL; >> >> + /* Check conversion between boolean types of different sizes. >> + If no vectype is specified, then we have a regular mask >> + assignment with no actual conversion. */ >> + if (rhs_code == CONVERT_EXPR > > CONVERT_EXPR_CODE_P (rhs_code) > >> + && !STMT_VINFO_DATA_REF (stmt_vinfo) >> + && !STMT_VINFO_VECTYPE (stmt_vinfo)) >> + { >> + if (TREE_CODE (rhs1) != SSA_NAME) >> + return NULL; >> + >> + rhs1_type = search_type_for_mask (rhs1, vinfo); >> + if (!rhs1_type) >> + return NULL; >> + >> + vectype1 = get_mask_type_for_scalar_type (rhs1_type); >> + >> + if (!vectype1) >> + return NULL; >> + >> + lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL); >> + pattern_stmt = gimple_build_assign (lhs, rhs1); > > So what's the actual issue here? That the conversion is spurious? > Why can't you accept this simply in vectorizable_assignment then? The problem is that conversion is supposed to be handled by vectorizable_conversion, but it fails to because it is not actually a conversion. I suppose it may be handled in vectorizable_assignment but I chose this pattern because it's meant to handle mask conversion issues. Thanks, Ilya > > Richard. > >> + *type_out = vectype1; >> + *type_in = vectype1; >> + stmts->safe_push (last_stmt); >> + if (dump_enabled_p ()) >> + dump_printf_loc (MSG_NOTE, vect_location, >> + "vect_recog_mask_conversion_pattern: detected:\n"); >> + >> + return pattern_stmt; >> + } >> + >> if (rhs_code != BIT_IOR_EXPR >> && rhs_code != BIT_XOR_EXPR >> && rhs_code != BIT_AND_EXPR)
On Wed, Dec 2, 2015 at 4:24 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: > 2015-12-02 17:52 GMT+03:00 Richard Biener <richard.guenther@gmail.com>: >> On Thu, Nov 12, 2015 at 5:08 PM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: >>> Hi, >>> >>> When we use LTO for fortran we may have a mix 32bit and 1bit scalar booleans. It means we may have conversion of one scalar type to another which confuses vectorizer because values with different scalar boolean type may get the same vectype. This patch transforms such conversions into comparison. >>> >>> I managed to make a small fortran test which gets vectorized with this patch but I didn't find how I can run fortran test with LTO and then scan tree dump to check it is vectorized. BTW here is a loop from the test: >>> >>> real*8 a(18) >>> logical b(18) >>> integer i >>> >>> do i=1,18 >>> if(a(i).gt.0.d0) then >>> b(i)=.true. >>> else >>> b(i)=.false. >>> endif >>> enddo >>> >>> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk? >>> >>> Thanks, >>> Ilya >>> -- >>> gcc/ >>> >>> 2015-11-12 Ilya Enkovich <enkovich.gnu@gmail.com> >>> >>> * tree-vect-patterns.c (vect_recog_mask_conversion_pattern): >>> Transform useless boolean conversion into assignment. >>> >>> >>> diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c >>> index b9d900c..62070da 100644 >>> --- a/gcc/tree-vect-patterns.c >>> +++ b/gcc/tree-vect-patterns.c >>> @@ -3674,6 +3674,38 @@ vect_recog_mask_conversion_pattern (vec<gimple *> *stmts, tree *type_in, >>> if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE) >>> return NULL; >>> >>> + /* Check conversion between boolean types of different sizes. >>> + If no vectype is specified, then we have a regular mask >>> + assignment with no actual conversion. */ >>> + if (rhs_code == CONVERT_EXPR >> >> CONVERT_EXPR_CODE_P (rhs_code) >> >>> + && !STMT_VINFO_DATA_REF (stmt_vinfo) >>> + && !STMT_VINFO_VECTYPE (stmt_vinfo)) >>> + { >>> + if (TREE_CODE (rhs1) != SSA_NAME) >>> + return NULL; >>> + >>> + rhs1_type = search_type_for_mask (rhs1, vinfo); >>> + if (!rhs1_type) >>> + return NULL; >>> + >>> + vectype1 = get_mask_type_for_scalar_type (rhs1_type); >>> + >>> + if (!vectype1) >>> + return NULL; >>> + >>> + lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL); >>> + pattern_stmt = gimple_build_assign (lhs, rhs1); >> >> So what's the actual issue here? That the conversion is spurious? >> Why can't you accept this simply in vectorizable_assignment then? > > The problem is that conversion is supposed to be handled by > vectorizable_conversion, > but it fails to because it is not actually a conversion. I suppose it > may be handled > in vectorizable_assignment but I chose this pattern because it's meant > to handle mask > conversion issues. I think it's always better to avoid patterns if you can. Richard. > Thanks, > Ilya > >> >> Richard. >> >>> + *type_out = vectype1; >>> + *type_in = vectype1; >>> + stmts->safe_push (last_stmt); >>> + if (dump_enabled_p ()) >>> + dump_printf_loc (MSG_NOTE, vect_location, >>> + "vect_recog_mask_conversion_pattern: detected:\n"); >>> + >>> + return pattern_stmt; >>> + } >>> + >>> if (rhs_code != BIT_IOR_EXPR >>> && rhs_code != BIT_XOR_EXPR >>> && rhs_code != BIT_AND_EXPR)
diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c index b9d900c..62070da 100644 --- a/gcc/tree-vect-patterns.c +++ b/gcc/tree-vect-patterns.c @@ -3674,6 +3674,38 @@ vect_recog_mask_conversion_pattern (vec<gimple *> *stmts, tree *type_in, if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE) return NULL; + /* Check conversion between boolean types of different sizes. + If no vectype is specified, then we have a regular mask + assignment with no actual conversion. */ + if (rhs_code == CONVERT_EXPR + && !STMT_VINFO_DATA_REF (stmt_vinfo) + && !STMT_VINFO_VECTYPE (stmt_vinfo)) + { + if (TREE_CODE (rhs1) != SSA_NAME) + return NULL; + + rhs1_type = search_type_for_mask (rhs1, vinfo); + if (!rhs1_type) + return NULL; + + vectype1 = get_mask_type_for_scalar_type (rhs1_type); + + if (!vectype1) + return NULL; + + lhs = vect_recog_temp_ssa_var (TREE_TYPE (lhs), NULL); + pattern_stmt = gimple_build_assign (lhs, rhs1); + + *type_out = vectype1; + *type_in = vectype1; + stmts->safe_push (last_stmt); + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "vect_recog_mask_conversion_pattern: detected:\n"); + + return pattern_stmt; + } + if (rhs_code != BIT_IOR_EXPR && rhs_code != BIT_XOR_EXPR && rhs_code != BIT_AND_EXPR)