On Tue, Jun 3, 2025 at 3:07 AM Tamar Christina <tamar.christ...@arm.com> wrote: > > Hi All, > > With the middle-end providing a way to make vectorization more profitable by > scaling vect-scalar-cost-multiplier this makes a more user friendly option > to make it easier to use. > > I propose making it an actual -m option that we document and retain vs using > the parameter name. In the future I would like to extend this option to > modify > additional costing in the AArch64 backend itself. > > This can be used together with --param aarch64-autovec-preference to get the > vectorizer to say, always vectorize with SVE. I did consider making this an > additional enum to --param aarch64-autovec-preference but I also think this is > a useful thing to be able to set with pragmas and attributes, but am open to > suggestions.
I did file a bug about requesting supporting --param in the attributes/pragmas, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116092 . It might be easy to add the support there too. Note this is not to stop adding an -m option which is considered more stable than the --param option but rather to let you know there is a bug about allowing --param. Thanks, Andrew > > Note that as a follow up I plan on extending -fdump-tree-vect to support > -stats > which is then intended to be usable with this flag. > > Bootstrapped Regtested on aarch64-none-linux-gnu, > arm-none-linux-gnueabihf, x86_64-pc-linux-gnu > -m32, -m64 and no issues. > > Ok for master? > > Thanks, > Tamar > > gcc/ChangeLog: > > * config/aarch64/aarch64.opt (max-vectorization): New. > * config/aarch64/aarch64.cc (aarch64_override_options_internal): Save > and restore option. > Implement it through vect-scalar-cost-multiplier. > (aarch64_attributes): Default to off. > * common/config/aarch64/aarch64-common.cc (aarch64_handle_option): > Initialize option. > * doc/extend.texi (max-vectorization): Document attribute. > * doc/invoke.texi (max-vectorization): Document flag. > > gcc/testsuite/ChangeLog: > > * gcc.target/aarch64/sve/cost_model_17.c: New test. > * gcc.target/aarch64/sve/cost_model_18.c: New test. > > --- > diff --git a/gcc/common/config/aarch64/aarch64-common.cc > b/gcc/common/config/aarch64/aarch64-common.cc > index > b9ed83642ade4462f1b030d68cf9744d31d70c23..1488697c6ce43108ae2938e5b8a00ac7ac262da6 > 100644 > --- a/gcc/common/config/aarch64/aarch64-common.cc > +++ b/gcc/common/config/aarch64/aarch64-common.cc > @@ -142,6 +142,10 @@ aarch64_handle_option (struct gcc_options *opts, > opts->x_aarch64_flag_outline_atomics = val; > return true; > > + case OPT_mmax_vectorization: > + opts->x_flag_aarch64_max_vectorization = val; > + return true; > + > default: > return true; > } > diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc > index > 9e3f2885bccb62550c5fcfdf93d72fbc2e63233e..f11f0da28915f49829360cd7a6269e2a3f67a860 > 100644 > --- a/gcc/config/aarch64/aarch64.cc > +++ b/gcc/config/aarch64/aarch64.cc > @@ -18973,6 +18973,12 @@ aarch64_override_options_internal (struct > gcc_options *opts) > if (TARGET_SME && !TARGET_SVE2) > sorry ("no support for %qs without %qs", "sme", "sve2"); > > + /* Set scalar costing to a high value such that we always pick > + vectorization. Increase scalar costing by 10000%. */ > + if (opts->x_flag_aarch64_max_vectorization) > + SET_OPTION_IF_UNSET (opts, &global_options_set, > + param_vect_scalar_cost_multiplier, 10000); > + > aarch64_override_options_after_change_1 (opts); > } > > @@ -19723,6 +19729,8 @@ static const struct aarch64_attribute_info > aarch64_attributes[] = > OPT_msign_return_address_ }, > { "outline-atomics", aarch64_attr_bool, true, NULL, > OPT_moutline_atomics}, > + { "max-vectorization", aarch64_attr_bool, false, NULL, > + OPT_mmax_vectorization}, > { NULL, aarch64_attr_custom, false, NULL, OPT____ } > }; > > diff --git a/gcc/config/aarch64/aarch64.opt b/gcc/config/aarch64/aarch64.opt > index > f32d56d4ffaef7862c1c45a11753be5d480220d0..2725c50da64a2c05489ea6202bdd5eedf1ba7e27 > 100644 > --- a/gcc/config/aarch64/aarch64.opt > +++ b/gcc/config/aarch64/aarch64.opt > @@ -290,6 +290,10 @@ msve-vector-bits= > Target RejectNegative Joined Enum(sve_vector_bits) > Var(aarch64_sve_vector_bits) Init(SVE_SCALABLE) > -msve-vector-bits=<number> Set the number of bits in an SVE vector > register. > > +mmax-vectorization > +Target Undocumented Var(flag_aarch64_max_vectorization) Save > +Override the scalar cost model such that vectorization is always profitable. > + > mverbose-cost-dump > Target Undocumented Var(flag_aarch64_verbose_cost) > Enables verbose cost model dumping in the debug dump files. > diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi > index > 40ccf22b29f4316928f905ec2c978fdaf30a55ec..429cbeb8a3c8186af60b1441acb52bc052b60721 > 100644 > --- a/gcc/doc/extend.texi > +++ b/gcc/doc/extend.texi > @@ -3882,6 +3882,15 @@ Enable or disable calls to out-of-line helpers to > implement atomic operations. > This corresponds to the behavior of the command-line options > @option{-moutline-atomics} and @option{-mno-outline-atomics}. > > +@cindex @code{max-vectorization} function attribute, AArch64 > +@item max-vectorization > +@itemx no-max-vectorization > +@code{max-vectorization} tells GCC's vectorizer to treat all vector > +loops as being more profitable than the original scalar loops when > +optimizing the current function. @code{no-max-vectorization} disables > +this behavior. > +@option{-mmax-vectorization} and @option{-mno-max-vectorization}. > + > @cindex @code{indirect_return} function attribute, AArch64 > @item indirect_return > The @code{indirect_return} attribute can be applied to a function type > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi > index > 95a25c0f63b77f26db05a7b48bfad8f9c58bcc5f..54d4cf88cccd4d0e4ede3442ad9907faac325d52 > 100644 > --- a/gcc/doc/invoke.texi > +++ b/gcc/doc/invoke.texi > @@ -21984,6 +21984,14 @@ used directly. The same applies when using > @option{-mcpu=} when the > selected cpu supports the @samp{lse} feature. > This option is on by default. > > +@item -mmax-vectorization > +@itemx -mno-max-vectorization > +Enable or disable override to vectorizer cost model making vectorization > always > +profitable. This option can be combined with -mautovec-preference=... > allowing > +precise control over which ISA will be used for auto-vectorization. Unlike > +-fno-vect-cost-model or -fvect-cost-model=unlimited this option does not turn > +off cost comparison between different vector modes. > + > @opindex march > @item -march=@var{name} > Specify the name of the target architecture and, optionally, one or > diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_17.c > b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_17.c > new file mode 100644 > index > 0000000000000000000000000000000000000000..c405591a101d50b4734bc6d65a6d6c01888bea48 > --- /dev/null > +++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_17.c > @@ -0,0 +1,21 @@ > +/* { dg-do compile } */ > +/* { dg-options "-Ofast -march=armv8-a+sve -mmax-vectorization > -fdump-tree-vect-details" } */ > + > +void > +foo (char *restrict a, int *restrict b, int *restrict c, > + int *restrict d, int stride) > +{ > + if (stride <= 1) > + return; > + > + for (int i = 0; i < 3; i++) > + { > + int res = c[i]; > + int t = b[i * stride]; > + if (a[i] != 0) > + res = t * d[i]; > + c[i] = res; > + } > +} > + > +/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" } } */ > diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_18.c > b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_18.c > new file mode 100644 > index > 0000000000000000000000000000000000000000..8e91f9e9c29971a4bb7033be6be4d2fc4c71d05a > --- /dev/null > +++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_18.c > @@ -0,0 +1,21 @@ > +/* { dg-do compile } */ > +/* { dg-options "-Ofast -march=armv8-a+sve -fdump-tree-vect-details" } */ > + > +void __attribute__ (( target ("max-vectorization"))) > +foo (char *restrict a, int *restrict b, int *restrict c, > + int *restrict d, int stride) > +{ > + if (stride <= 1) > + return; > + > + for (int i = 0; i < 3; i++) > + { > + int res = c[i]; > + int t = b[i * stride]; > + if (a[i] != 0) > + res = t * d[i]; > + c[i] = res; > + } > +} > + > +/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" } } */ > > > --
