https://gcc.gnu.org/g:f3d0c95570d59bef636625c8ed4c68a1884bab1b
commit r15-10456-gf3d0c95570d59bef636625c8ed4c68a1884bab1b Author: Tamar Christina <[email protected]> Date: Fri Oct 17 15:43:04 2025 +0100 AArch64: Extend intrinsics framework to account for merging predications without gp [PR121604] In PR121604 the problem was noted that currently the SVE intrinsics infrastructure assumes that for any predicated operation that the GP is at the first argument position which has a svbool_t or for a unary merging operation that it's in the second position. However you have intrinsics like fmov_lane which have an svbool_t but it's not a GP. You also have instructions like BRKB which work only on predicates so it incorrectly determines the first operand to be the GP, while that's the inactive lanes. However during apply_predication we do have the information about where the GP is. This patch re-organizes the code to record this information into the function_instance such that folders have access to this information. For functions that are outliers like pmov_lane we can now override the availability of the intrinsics having a GP. gcc/ChangeLog: PR target/121604 * config/aarch64/aarch64-sve-builtins-shapes.cc (apply_predication): Store gp_index. (struct pmov_to_vector_lane_def): Mark instruction as has no GP. * config/aarch64/aarch64-sve-builtins.h (function_instance::gp_value, function_instance::inactive_values, function_instance::gp_index, function_shape::has_gp_argument_p): New. * config/aarch64/aarch64-sve-builtins.cc (gimple_folder::fold_pfalse): Simplify code and use GP helpers. gcc/testsuite/ChangeLog: PR target/121604 * gcc.target/aarch64/sve/pr121604_brk.c: New test. * gcc.target/aarch64/sve2/pr121604_pmov.c: New test. Co-authored-by: Jennifer Schmitz <[email protected]> (cherry picked from commit d1965b1fd8938f35f78be503e36b98b406751e21) Diff: --- gcc/config/aarch64/aarch64-sve-builtins-shapes.cc | 28 +++++++++++++-- gcc/config/aarch64/aarch64-sve-builtins.cc | 22 ++++++------ gcc/config/aarch64/aarch64-sve-builtins.h | 40 ++++++++++++++++++++++ .../gcc.target/aarch64/sve/pr121604_brk.c | 25 ++++++++++++++ .../gcc.target/aarch64/sve2/pr121604_pmov.c | 16 +++++++++ 5 files changed, 117 insertions(+), 14 deletions(-) diff --git a/gcc/config/aarch64/aarch64-sve-builtins-shapes.cc b/gcc/config/aarch64/aarch64-sve-builtins-shapes.cc index af23a154d0e8..8349221797b3 100644 --- a/gcc/config/aarch64/aarch64-sve-builtins-shapes.cc +++ b/gcc/config/aarch64/aarch64-sve-builtins-shapes.cc @@ -21,7 +21,10 @@ #include "system.h" #include "coretypes.h" #include "tm.h" +#include "basic-block.h" #include "tree.h" +#include "function.h" +#include "gimple.h" #include "rtl.h" #include "tm_p.h" #include "memmodel.h" @@ -68,23 +71,36 @@ za_group_is_pure_overload (const function_group_info &group) types in ARGUMENT_TYPES. RETURN_TYPE is the type returned by the function. */ static void -apply_predication (const function_instance &instance, tree return_type, +apply_predication (function_instance &instance, tree return_type, vec<tree> &argument_types) { + /* Initially mark the function as not being predicated. */ + instance.gp_index = -1; + /* There are currently no SME ZA instructions that have both merging and unpredicated forms, so for simplicity, the predicates are always included in the original format string. */ if (instance.pred != PRED_none && instance.pred != PRED_za_m) { argument_types.quick_insert (0, instance.gp_type ()); + instance.gp_index = 0; /* For unary merge operations, the first argument is a vector with the same type as the result. For unary_convert_narrowt it also provides the "bottom" half of active elements, and is present for all types of predication. */ auto nargs = argument_types.length () - 1; if (instance.shape->has_merge_argument_p (instance, nargs)) - argument_types.quick_insert (0, return_type); + { + argument_types.quick_insert (0, return_type); + instance.gp_index = 1; + } } + + /* In this case the predicate type we added above is a non-governing + predicate operand (and there is no GP), so update the gp_index value + accordingly. */ + if (!instance.shape->has_gp_argument_p (instance)) + instance.gp_index = -1; } /* Parse and move past an element type in FORMAT and return it as a type @@ -3301,6 +3317,14 @@ struct pmov_to_vector_lane_def : public overloaded_base<0> but it doesn't currently have the necessary information. */ return c.require_immediate_range (1, 1, bytes - 1); } + + /* This function has a predicate argument, and is a merging instruction, but + the predicate is not a GP. */ + bool + has_gp_argument_p (const function_instance &) const override + { + return false; + } }; SHAPE (pmov_to_vector_lane) diff --git a/gcc/config/aarch64/aarch64-sve-builtins.cc b/gcc/config/aarch64/aarch64-sve-builtins.cc index 24feb7ce1578..ad0d38934e21 100644 --- a/gcc/config/aarch64/aarch64-sve-builtins.cc +++ b/gcc/config/aarch64/aarch64-sve-builtins.cc @@ -3630,24 +3630,22 @@ gimple_folder::redirect_pred_x () gimple * gimple_folder::fold_pfalse () { - if (pred == PRED_none) + tree gp = gp_value (call); + /* If there isn't a GP then we can't do any folding as the instruction isn't + predicated. */ + if (!gp) return nullptr; - tree arg0 = gimple_call_arg (call, 0); + if (pred == PRED_m) { - /* Unary function shapes with _m predication are folded to the - inactive vector (arg0), while other function shapes are folded - to op1 (arg1). */ - tree arg1 = gimple_call_arg (call, 1); - if (is_pfalse (arg1)) - return fold_call_to (arg0); - if (is_pfalse (arg0)) - return fold_call_to (arg1); + tree val = inactive_values (call); + if (is_pfalse (gp)) + return fold_call_to (val); return nullptr; } - if ((pred == PRED_x || pred == PRED_z) && is_pfalse (arg0)) + if ((pred == PRED_x || pred == PRED_z) && is_pfalse (gp)) return fold_call_to (build_zero_cst (TREE_TYPE (lhs))); - if (pred == PRED_implicit && is_pfalse (arg0)) + if (pred == PRED_implicit && is_pfalse (gp)) { unsigned int flags = call_properties (); /* Folding to lhs = {0, ...} is not appropriate for intrinsics with diff --git a/gcc/config/aarch64/aarch64-sve-builtins.h b/gcc/config/aarch64/aarch64-sve-builtins.h index c145b8065ae3..7e1539a0c722 100644 --- a/gcc/config/aarch64/aarch64-sve-builtins.h +++ b/gcc/config/aarch64/aarch64-sve-builtins.h @@ -402,6 +402,8 @@ public: bool could_trap_p () const; vector_type_index gp_type_index () const; + tree gp_value (gcall *) const; + tree inactive_values (gcall *) const; tree gp_type () const; unsigned int vectors_per_tuple () const; @@ -435,6 +437,7 @@ public: group_suffix_index group_suffix_id; predication_index pred; fpm_mode_index fpm_mode; + int gp_index; }; class registered_function; @@ -800,6 +803,8 @@ public: virtual bool has_merge_argument_p (const function_instance &, unsigned int) const; + virtual bool has_gp_argument_p (const function_instance &) const; + virtual bool explicit_type_suffix_p (unsigned int) const = 0; /* True if the group suffix is present in overloaded names. @@ -948,6 +953,33 @@ function_instance::gp_type () const return acle_vector_types[0][gp_type_index ()]; } +/* Return the tree value that should be used as the governing predicate of + this function. If none then return NULL_TREE. */ +inline tree +function_instance::gp_value (gcall *call) const +{ + if (gp_index < 0) + return NULL_TREE; + + return gimple_call_arg (call, gp_index); +} + +/* Return the tree value that should be used for the inactive lanes should this + function be a predicated function with a gp. Otherwise return NULL_TREE. */ +inline tree +function_instance::inactive_values (gcall *call) const +{ + if (gp_index < 0) + return NULL_TREE; + + /* Function is unary with m predicate. */ + if (gp_index == 1) + return gimple_call_arg (call, 0); + + /* Else the inactive values are the next element. */ + return gimple_call_arg (call, 1); +} + /* If the function operates on tuples of vectors, return the number of vectors in the tuples, otherwise return 1. */ inline unsigned int @@ -1122,6 +1154,14 @@ function_shape::has_merge_argument_p (const function_instance &instance, return nargs == 1 && instance.pred == PRED_m; } +/* Return true if INSTANCE has an predicate argument that can be used as the global + predicate. */ +inline bool +function_shape::has_gp_argument_p (const function_instance &instance) const +{ + return instance.pred != PRED_none; +} + /* Return the mode of the result of a call. */ inline machine_mode function_expander::result_mode () const diff --git a/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c b/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c new file mode 100644 index 000000000000..a474a20554d3 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c @@ -0,0 +1,25 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ +/* { dg-final { check-function-bodies "**" "" "" } } */ + +#include <arm_sve.h> + +/* +** foo: +** ptrue p0\.b, all +** brkb p0\.b, p0/z, p0\.b +** ret +*/ +svbool_t foo () { + return svbrkb_b_m (svpfalse (), svptrue_b8 (), svptrue_b8 ()); +} + +/* +** bar: +** ptrue p0\.b, all +** brka p0\.b, p0/z, p0\.b +** ret +*/ +svbool_t bar () { + return svbrka_b_m (svpfalse (), svptrue_b8 (), svptrue_b8 ()); +} diff --git a/gcc/testsuite/gcc.target/aarch64/sve2/pr121604_pmov.c b/gcc/testsuite/gcc.target/aarch64/sve2/pr121604_pmov.c new file mode 100644 index 000000000000..16844ee4add3 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/sve2/pr121604_pmov.c @@ -0,0 +1,16 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -march=armv8.2-a+sve2p1" } */ +/* { dg-final { check-function-bodies "**" "" "" } } */ + +#include <arm_sve.h> + +/* +** f: +** pfalse p([0-7]+)\.b +** mov z0\.b, #-1 +** pmov z0\[1\], p\1\.d +** ret +*/ +svuint64_t f () { + return svpmov_lane_u64_m (svdup_u64 (~0UL), svpfalse (), 1); +}
