| Message ID | 20151211095847.GB54845@msticlxl57.ims.intel.com |
|---|---|
| State | New |
| Headers | show |
On Fri, Dec 11, 2015 at 10:58 AM, Ilya Enkovich <enkovich.gnu@gmail.com> wrote: > Hi, > > Currently when MASK_LOAD and MASK_STORE is vectorized we check > scalar type of a mask but don't check its vector mask. It means > we may vectorize it when mask was just loaded from an array of > booleans. This happens e.g. for following test: > > SUBROUTINE TEST (x, y, z, mask, ims, ime) > IMPLICIT NONE > INTEGER ims, ime, i > LOGICAL mask (ims:ime) > REAL x(ims:ime), y(ims:ime), z(ims:ime) > > DO 812 i=ims,ime > IF (mask(i)) x(i) = y(i) - z(i) > 812 CONTINUE > END > > Produced GIMPLE: > > vect__15.30_76 = MEM[base: _91, offset: 0B]; > vect__17.32_80 = MASK_LOAD (vectp_y.33_78, 4B, vect__15.30_76); > vect__19.35_83 = MASK_LOAD (vectp_z.36_81, 4B, vect__15.30_76); > vect__20.38_84 = vect__17.32_80 - vect__19.35_83; > MASK_STORE (vectp_x.39_85, 4B, vect__15.30_76, vect__20.38_84); > > Loaded values don't have required all 0s and all 1s values and > shouldn't be used as a mask. This patch checks vector type of > a mask used for MASK_LOAD and MASK_STORE. Bootstrapped and > regtested for x86_64-pc-linux-gnu. OK for trunk? Ok. Thanks, Richard. > Thanks, > Ilya > -- > gcc/ > > 2015-12-11 Ilya Enkovich <enkovich.gnu@gmail.com> > > * tree-vect-stmts.c (vectorizable_mask_load_store): Check > mask vectype. > > > diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c > index 5377d15..abcd9a4 100644 > --- a/gcc/tree-vect-stmts.c > +++ b/gcc/tree-vect-stmts.c > @@ -1780,7 +1780,7 @@ vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, > if (!mask_vectype) > mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); > > - if (!mask_vectype) > + if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) > return false; > > if (is_store)
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c index 5377d15..abcd9a4 100644 --- a/gcc/tree-vect-stmts.c +++ b/gcc/tree-vect-stmts.c @@ -1780,7 +1780,7 @@ vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, if (!mask_vectype) mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); - if (!mask_vectype) + if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) return false; if (is_store)
Hi, Currently when MASK_LOAD and MASK_STORE is vectorized we check scalar type of a mask but don't check its vector mask. It means we may vectorize it when mask was just loaded from an array of booleans. This happens e.g. for following test: SUBROUTINE TEST (x, y, z, mask, ims, ime) IMPLICIT NONE INTEGER ims, ime, i LOGICAL mask (ims:ime) REAL x(ims:ime), y(ims:ime), z(ims:ime) DO 812 i=ims,ime IF (mask(i)) x(i) = y(i) - z(i) 812 CONTINUE END Produced GIMPLE: vect__15.30_76 = MEM[base: _91, offset: 0B]; vect__17.32_80 = MASK_LOAD (vectp_y.33_78, 4B, vect__15.30_76); vect__19.35_83 = MASK_LOAD (vectp_z.36_81, 4B, vect__15.30_76); vect__20.38_84 = vect__17.32_80 - vect__19.35_83; MASK_STORE (vectp_x.39_85, 4B, vect__15.30_76, vect__20.38_84); Loaded values don't have required all 0s and all 1s values and shouldn't be used as a mask. This patch checks vector type of a mask used for MASK_LOAD and MASK_STORE. Bootstrapped and regtested for x86_64-pc-linux-gnu. OK for trunk? Thanks, Ilya -- gcc/ 2015-12-11 Ilya Enkovich <enkovich.gnu@gmail.com> * tree-vect-stmts.c (vectorizable_mask_load_store): Check mask vectype.