This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[RFC PATCH] Masked load/store vectorization
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Richard Henderson <rth at redhat dot com>, Richard Biener <rguenther at suse dot de>
- Cc: gcc-patches at gcc dot gnu dot org, "H.J. Lu" <hjl dot tools at gmail dot com>, "Zolotukhin, Michael V" <michael dot v dot zolotukhin at intel dot com>, "Yukhin, Kirill" <kirill dot yukhin at intel dot com>
- Date: Fri, 2 Nov 2012 18:14:15 +0100
- Subject: [RFC PATCH] Masked load/store vectorization
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
This is a RFC patch which implements masked (conditional) vector
loads/stores roughly as discussed with richi (i.e. internal function
calls, added during if-conversion pass, and if not vectorized
removed after vectorization again).
The if-unconversion pass is just very basic, I guess it could often
do better job by trying to guess which statements in the bb with
MASK_LOAD/MASK_STORE were in which bb, depending on what condition,
because right now vrp2 is able to fix up something, but not undo
COND_EXPRs on PHI nodes, and there is no sink pass after vect.
Testcases it can handle include e.g.
extern void abort (void);
__attribute__((noinline, noclone)) void
foo (float *__restrict x, float *__restrict y, float *__restrict z)
{
float *__restrict p = __builtin_assume_aligned (x, 32);
float *__restrict q = __builtin_assume_aligned (y, 32);
float *__restrict r = __builtin_assume_aligned (z, 32);
int i;
for (i = 0; i < 1024; i++)
{
if (p[i] < 0.0f)
q[i] = p[i] + 2.0f;
else
p[i] = r[i] + 3.0f;
}
}
float a[1024] __attribute__((aligned (32)));
float b[1024] __attribute__((aligned (32)));
float c[1024] __attribute__((aligned (32)));
int
main ()
{
int i;
for (i = 0; i < 1024; i++)
{
a[i] = (i & 1) ? -i : i;
b[i] = 7 * i;
c[i] = a[i] - 3.0f;
}
foo (a, b, c);
for (i = 0; i < 1024; i++)
if (a[i] != ((i & 1) ? -i : i)
|| b[i] != ((i & 1) ? a[i] + 2.0f : 7 * i)
|| c[i] != a[i] - 3.0f)
abort ();
for (i = 0; i < 5000000; i++)
foo (a, b, c);
return 0;
}
or
void
foo (double *x, double *y)
{
double *p = __builtin_assume_aligned (x, 16);
double *q = __builtin_assume_aligned (y, 16);
double z, h;
int i;
for (i = 0; i < 1024; i++)
{
if (p[i] < 0.0)
z = q[i], h = q[i] * 7.0 + 3.0;
else
z = p[i] + 6.0, h = p[1024 + i];
p[i] = z + 2.0 * h;
}
}
or
#define N 1024
float vf1[N+16], vf2[N], vf3[N];
double vd1[N+16], vd2[N];
int k[N];
long l[N];
short n[N];
__attribute__((noinline, noclone)) void
f1 (void)
{
int i;
for (i = 0; i < N; i++)
{
float f;
if (vf3[i] < 0.0f)
f = vf1[k[i]];
else
f = 7.0f;
vf2[i] = f;
}
}
(all with -O3 -mavx resp. -O3 -mavx2). vmaskmov*/vpmaskmov* and
masked gather insns are then used.
2012-11-02 Jakub Jelinek <jakub@redhat.com>
* Makefile.in (tree-if-conv.o): Depend on $(TARGET_H), $(EXPR_H)
and $(OPTABS_H).
* config/i386/sse.md (maskload<mode>, maskstore<mode>): New expanders.
* tree-data-ref.c (struct data_ref_loc_d): Replace pos field with ref.
(get_references_in_stmt): Don't record operand addresses, but
operands themselves. Handle MASK_LOAD and MASK_STORE.
(find_data_references_in_stmt, graphite_find_data_references_in_stmt,
create_rdg_vertices): Adjust users of pos field of data_ref_loc_d.
* internal-fn.def (MASK_LOAD, MASK_STORE): New internal fns.
* tree-if-conv.c: Add target.h, expr.h and optabs.h includes.
(if_convertible_phi_p, insert_gimplified_predicates): Add
any_mask_load_store argument, if true, handle it like
flag_tree_loop_if_convert_stores.
(ifcvt_can_use_mask_load_store): New function.
(if_convertible_gimple_assign_stmt_p): Add any_mask_load_store
argument, check if some conditional loads or stores can't be
converted into MASK_LOAD or MASK_STORE.
(if_convertible_stmt_p): Add any_mask_load_store argument,
pass it down to if_convertible_gimple_assign_stmt_p.
(if_convertible_loop_p_1): Add any_mask_load_store argument,
pass it down to if_convertible_stmt_p and if_convertible_phi_p,
call if_convertible_phi_p only after all if_convertible_stmt_p
calls.
(if_convertible_loop_p): Add any_mask_load_store argument,
pass it down to if_convertible_loop_p_1.
(predicate_mem_writes): Emit MASK_LOAD and/or MASK_STORE calls.
(combine_blocks): Add any_mask_load_store argument, pass
it down to insert_gimplified_predicates and call predicate_mem_writes
if it is set.
(tree_if_conversion): Add any_mask_load_store_p argument,
adjust if_convertible_loop_p, combine_blocks calls and gather
whether any mask loads/stores have been generated.
(need_if_unconversion): New variable.
(main_tree_if_conversion): Adjust tree_if_conversion caller,
if any masked loads/stores have been created, set
need_if_unconversion and return TODO_update_ssa_only_virtuals.
(gate_tree_if_unconversion, main_tree_if_unconversion): New
functions.
(pass_if_unconversion): New pass descriptor.
* tree-vect-data-refs.c (vect_check_gather): Handle
MASK_LOAD/MASK_STORE.
(vect_analyze_data_refs, vect_supportable_dr_alignment): Likewise.
* gimple.h (gimple_expr_type): Handle MASK_STORE.
* internal-fn.c (expand_MASK_LOAD, expand_MASK_STORE): New functions.
* tree-vect-loop.c (vect_determine_vectorization_factor): Handle
MASK_STORE.
* passes.c (init_optimization_passes): Add pass_if_unconversion.
* optabs.def (maskload_optab, maskstore_optab): New optabs.
* tree-pass.h (pass_if_unconversion): New extern decl.
* tree-vect-stmts.c (vect_mark_relevant): Don't crash if lhs
is NULL.
(exist_non_indexing_operands_for_use_p): Handle MASK_LOAD
and MASK_STORE.
(vectorizable_mask_load_store): New function.
(vectorizable_call): Call it for MASK_LOAD or MASK_STORE.
(vect_transform_stmt): Handle MASK_STORE.
--- gcc/Makefile.in.jj 2012-11-02 09:01:55.000000000 +0100
+++ gcc/Makefile.in 2012-11-02 17:48:10.144549562 +0100
@@ -2378,7 +2378,7 @@ tree-nested.o: tree-nested.c $(CONFIG_H)
tree-if-conv.o: tree-if-conv.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) \
$(TREE_H) $(FLAGS_H) $(BASIC_BLOCK_H) $(TREE_FLOW_H) \
$(CFGLOOP_H) $(TREE_DATA_REF_H) $(TREE_PASS_H) $(DIAGNOSTIC_H) \
- $(DBGCNT_H) $(GIMPLE_PRETTY_PRINT_H)
+ $(DBGCNT_H) $(GIMPLE_PRETTY_PRINT_H) $(TARGET_H) $(EXPR_H) $(OPTABS_H)
tree-iterator.o : tree-iterator.c $(CONFIG_H) $(SYSTEM_H) $(TREE_H) \
coretypes.h $(GGC_H) tree-iterator.h $(GIMPLE_H) gt-tree-iterator.h
tree-dfa.o : tree-dfa.c $(TREE_FLOW_H) $(CONFIG_H) $(SYSTEM_H) \
--- gcc/config/i386/sse.md.jj 2012-11-02 09:01:47.493572793 +0100
+++ gcc/config/i386/sse.md 2012-11-02 10:12:48.873876499 +0100
@@ -11294,6 +11294,23 @@ (define_insn "<avx_avx2>_maskstore<ssemo
(set_attr "prefix" "vex")
(set_attr "mode" "<sseinsnmode>")])
+(define_expand "maskload<mode>"
+ [(set (match_operand:V48_AVX2 0 "register_operand")
+ (unspec:V48_AVX2
+ [(match_operand:<sseintvecmode> 2 "register_operand")
+ (match_operand:V48_AVX2 1 "memory_operand")]
+ UNSPEC_MASKMOV))]
+ "TARGET_AVX")
+
+(define_expand "maskstore<mode>"
+ [(set (match_operand:V48_AVX2 0 "memory_operand")
+ (unspec:V48_AVX2
+ [(match_operand:<sseintvecmode> 2 "register_operand")
+ (match_operand:V48_AVX2 1 "register_operand")
+ (match_dup 0)]
+ UNSPEC_MASKMOV))]
+ "TARGET_AVX")
+
(define_insn_and_split "avx_<castmode><avxsizesuffix>_<castmode>"
[(set (match_operand:AVX256MODE2P 0 "nonimmediate_operand" "=x,m")
(unspec:AVX256MODE2P
--- gcc/tree-data-ref.c.jj 2012-11-02 09:01:47.069575609 +0100
+++ gcc/tree-data-ref.c 2012-11-02 09:13:34.203881900 +0100
@@ -4283,11 +4283,11 @@ compute_all_dependences (VEC (data_refer
typedef struct data_ref_loc_d
{
- /* Position of the memory reference. */
- tree *pos;
+ /* The memory reference. */
+ tree ref;
- /* True if the memory reference is read. */
- bool is_read;
+ /* True if the memory reference is read. */
+ bool is_read;
} data_ref_loc;
DEF_VEC_O (data_ref_loc);
@@ -4301,7 +4301,7 @@ get_references_in_stmt (gimple stmt, VEC
{
bool clobbers_memory = false;
data_ref_loc ref;
- tree *op0, *op1;
+ tree op0, op1;
enum gimple_code stmt_code = gimple_code (stmt);
*references = NULL;
@@ -4310,7 +4310,10 @@ get_references_in_stmt (gimple stmt, VEC
As we cannot model data-references to not spelled out
accesses give up if they may occur. */
if ((stmt_code == GIMPLE_CALL
- && !(gimple_call_flags (stmt) & ECF_CONST))
+ && !(gimple_call_flags (stmt) & ECF_CONST)
+ && (!gimple_call_internal_p (stmt)
+ || (gimple_call_internal_fn (stmt) != IFN_MASK_LOAD
+ && gimple_call_internal_fn (stmt) != IFN_MASK_STORE)))
|| (stmt_code == GIMPLE_ASM
&& (gimple_asm_volatile_p (stmt) || gimple_vuse (stmt))))
clobbers_memory = true;
@@ -4321,15 +4324,15 @@ get_references_in_stmt (gimple stmt, VEC
if (stmt_code == GIMPLE_ASSIGN)
{
tree base;
- op0 = gimple_assign_lhs_ptr (stmt);
- op1 = gimple_assign_rhs1_ptr (stmt);
+ op0 = gimple_assign_lhs (stmt);
+ op1 = gimple_assign_rhs1 (stmt);
- if (DECL_P (*op1)
- || (REFERENCE_CLASS_P (*op1)
- && (base = get_base_address (*op1))
+ if (DECL_P (op1)
+ || (REFERENCE_CLASS_P (op1)
+ && (base = get_base_address (op1))
&& TREE_CODE (base) != SSA_NAME))
{
- ref.pos = op1;
+ ref.ref = op1;
ref.is_read = true;
VEC_safe_push (data_ref_loc, heap, *references, ref);
}
@@ -4338,16 +4341,35 @@ get_references_in_stmt (gimple stmt, VEC
{
unsigned i, n;
- op0 = gimple_call_lhs_ptr (stmt);
+ ref.is_read = false;
+ if (gimple_call_internal_p (stmt))
+ switch (gimple_call_internal_fn (stmt))
+ {
+ case IFN_MASK_LOAD:
+ ref.is_read = true;
+ case IFN_MASK_STORE:
+ ref.ref = build2 (MEM_REF,
+ ref.is_read
+ ? TREE_TYPE (gimple_call_lhs (stmt))
+ : TREE_TYPE (gimple_call_arg (stmt, 3)),
+ gimple_call_arg (stmt, 0),
+ gimple_call_arg (stmt, 1));
+ VEC_safe_push (data_ref_loc, heap, *references, ref);
+ return false;
+ default:
+ break;
+ }
+
+ op0 = gimple_call_lhs (stmt);
n = gimple_call_num_args (stmt);
for (i = 0; i < n; i++)
{
- op1 = gimple_call_arg_ptr (stmt, i);
+ op1 = gimple_call_arg (stmt, i);
- if (DECL_P (*op1)
- || (REFERENCE_CLASS_P (*op1) && get_base_address (*op1)))
+ if (DECL_P (op1)
+ || (REFERENCE_CLASS_P (op1) && get_base_address (op1)))
{
- ref.pos = op1;
+ ref.ref = op1;
ref.is_read = true;
VEC_safe_push (data_ref_loc, heap, *references, ref);
}
@@ -4356,11 +4378,11 @@ get_references_in_stmt (gimple stmt, VEC
else
return clobbers_memory;
- if (*op0
- && (DECL_P (*op0)
- || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0))))
+ if (op0
+ && (DECL_P (op0)
+ || (REFERENCE_CLASS_P (op0) && get_base_address (op0))))
{
- ref.pos = op0;
+ ref.ref = op0;
ref.is_read = false;
VEC_safe_push (data_ref_loc, heap, *references, ref);
}
@@ -4390,7 +4412,7 @@ find_data_references_in_stmt (struct loo
FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
{
dr = create_data_ref (nest, loop_containing_stmt (stmt),
- *ref->pos, stmt, ref->is_read);
+ ref->ref, stmt, ref->is_read);
gcc_assert (dr != NULL);
VEC_safe_push (data_reference_p, heap, *datarefs, dr);
}
@@ -4422,7 +4444,7 @@ graphite_find_data_references_in_stmt (l
FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
{
- dr = create_data_ref (nest, loop, *ref->pos, stmt, ref->is_read);
+ dr = create_data_ref (nest, loop, ref->ref, stmt, ref->is_read);
gcc_assert (dr != NULL);
VEC_safe_push (data_reference_p, heap, *datarefs, dr);
}
@@ -5058,7 +5080,7 @@ create_rdg_vertices (struct graph *rdg,
else
RDGV_HAS_MEM_READS (v) = true;
dr = create_data_ref (loop, loop_containing_stmt (stmt),
- *ref->pos, stmt, ref->is_read);
+ ref->ref, stmt, ref->is_read);
if (dr)
VEC_safe_push (data_reference_p, heap, RDGV_DATAREFS (v), dr);
}
--- gcc/internal-fn.def.jj 2012-11-02 09:01:47.590572147 +0100
+++ gcc/internal-fn.def 2012-11-02 09:13:34.215881819 +0100
@@ -1,5 +1,5 @@
/* Internal functions.
- Copyright (C) 2011 Free Software Foundation, Inc.
+ Copyright (C) 2011, 2012 Free Software Foundation, Inc.
This file is part of GCC.
@@ -40,3 +40,5 @@ along with GCC; see the file COPYING3.
DEF_INTERNAL_FN (LOAD_LANES, ECF_CONST | ECF_LEAF)
DEF_INTERNAL_FN (STORE_LANES, ECF_CONST | ECF_LEAF)
+DEF_INTERNAL_FN (MASK_LOAD, ECF_PURE | ECF_LEAF)
+DEF_INTERNAL_FN (MASK_STORE, ECF_LEAF)
--- gcc/tree-if-conv.c.jj 2012-11-02 09:01:47.108575349 +0100
+++ gcc/tree-if-conv.c 2012-11-02 16:46:18.306241821 +0100
@@ -96,6 +96,9 @@ along with GCC; see the file COPYING3.
#include "tree-scalar-evolution.h"
#include "tree-pass.h"
#include "dbgcnt.h"
+#include "target.h"
+#include "expr.h"
+#include "optabs.h"
/* List of basic blocks in if-conversion-suitable order. */
static basic_block *ifc_bbs;
@@ -448,7 +451,8 @@ bb_with_exit_edge_p (struct loop *loop,
- there is a virtual PHI in a BB other than the loop->header. */
static bool
-if_convertible_phi_p (struct loop *loop, basic_block bb, gimple phi)
+if_convertible_phi_p (struct loop *loop, basic_block bb, gimple phi,
+ bool any_mask_load_store)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
@@ -463,7 +467,7 @@ if_convertible_phi_p (struct loop *loop,
return false;
}
- if (flag_tree_loop_if_convert_stores)
+ if (flag_tree_loop_if_convert_stores || any_mask_load_store)
return true;
/* When the flag_tree_loop_if_convert_stores is not set, check
@@ -679,6 +683,84 @@ ifcvt_could_trap_p (gimple stmt, VEC (da
return gimple_could_trap_p (stmt);
}
+/* Return true if STMT could be converted into a masked load or store
+ (conditional load or store based on a mask computed from bb predicate). */
+
+static bool
+ifcvt_can_use_mask_load_store (gimple stmt)
+{
+ tree lhs, ref;
+ enum machine_mode mode, vmode;
+ optab op;
+ basic_block bb;
+ unsigned int vector_sizes;
+
+ if (!flag_tree_vectorize
+ || !gimple_assign_single_p (stmt)
+ || gimple_has_volatile_ops (stmt))
+ return false;
+
+ /* Avoid creating mask loads/stores if we'd need to chain
+ conditions, to make it easier to undo them. */
+ bb = gimple_bb (stmt);
+ if (!single_pred_p (bb)
+ || is_predicated (single_pred (bb)))
+ return false;
+
+ /* Check whether this is a load or store. */
+ lhs = gimple_assign_lhs (stmt);
+ if (TREE_CODE (lhs) != SSA_NAME)
+ {
+ if (!is_gimple_val (gimple_assign_rhs1 (stmt)))
+ return false;
+ op = maskstore_optab;
+ ref = lhs;
+ }
+ else if (gimple_assign_load_p (stmt))
+ {
+ op = maskload_optab;
+ ref = gimple_assign_rhs1 (stmt);
+ }
+ else
+ return false;
+
+ /* And whether REF isn't a MEM_REF with non-addressable decl. */
+ if (TREE_CODE (ref) == MEM_REF
+ && TREE_CODE (TREE_OPERAND (ref, 0)) == ADDR_EXPR
+ && DECL_P (TREE_OPERAND (TREE_OPERAND (ref, 0), 0))
+ && !TREE_ADDRESSABLE (TREE_OPERAND (TREE_OPERAND (ref, 0), 0)))
+ return false;
+
+ /* Mask should be integer mode of the same size as the load/store
+ mode. */
+ mode = TYPE_MODE (TREE_TYPE (lhs));
+ if (int_mode_for_mode (mode) == BLKmode)
+ return false;
+
+ /* See if there is any chance the mask load or store might be
+ vectorized. If not, punt. */
+ vmode = targetm.vectorize.preferred_simd_mode (mode);
+ if (!VECTOR_MODE_P (vmode))
+ return false;
+
+ if (optab_handler (op, vmode) != CODE_FOR_nothing)
+ return true;
+
+ vector_sizes = targetm.vectorize.autovectorize_vector_sizes ();
+ while (vector_sizes != 0)
+ {
+ unsigned int cur = 1 << floor_log2 (vector_sizes);
+ vector_sizes &= ~cur;
+ if (cur <= GET_MODE_SIZE (mode))
+ continue;
+ vmode = mode_for_vector (mode, cur / GET_MODE_SIZE (mode));
+ if (VECTOR_MODE_P (vmode)
+ && optab_handler (op, vmode) != CODE_FOR_nothing)
+ return true;
+ }
+ return false;
+}
+
/* Return true when STMT is if-convertible.
GIMPLE_ASSIGN statement is not if-convertible if,
@@ -688,7 +770,8 @@ ifcvt_could_trap_p (gimple stmt, VEC (da
static bool
if_convertible_gimple_assign_stmt_p (gimple stmt,
- VEC (data_reference_p, heap) *refs)
+ VEC (data_reference_p, heap) *refs,
+ bool *any_mask_load_store)
{
tree lhs = gimple_assign_lhs (stmt);
basic_block bb;
@@ -714,10 +797,18 @@ if_convertible_gimple_assign_stmt_p (gim
return false;
}
+ gimple_set_plf (stmt, GF_PLF_1, false);
+
if (flag_tree_loop_if_convert_stores)
{
if (ifcvt_could_trap_p (stmt, refs))
{
+ if (ifcvt_can_use_mask_load_store (stmt))
+ {
+ gimple_set_plf (stmt, GF_PLF_1, true);
+ *any_mask_load_store = true;
+ return true;
+ }
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "tree could trap...\n");
return false;
@@ -727,6 +818,12 @@ if_convertible_gimple_assign_stmt_p (gim
if (gimple_assign_rhs_could_trap_p (stmt))
{
+ if (ifcvt_can_use_mask_load_store (stmt))
+ {
+ gimple_set_plf (stmt, GF_PLF_1, true);
+ *any_mask_load_store = true;
+ return true;
+ }
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "tree could trap...\n");
return false;
@@ -738,6 +835,12 @@ if_convertible_gimple_assign_stmt_p (gim
&& bb != bb->loop_father->header
&& !bb_with_exit_edge_p (bb->loop_father, bb))
{
+ if (ifcvt_can_use_mask_load_store (stmt))
+ {
+ gimple_set_plf (stmt, GF_PLF_1, true);
+ *any_mask_load_store = true;
+ return true;
+ }
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "LHS is not var\n");
@@ -756,7 +859,8 @@ if_convertible_gimple_assign_stmt_p (gim
- it is a GIMPLE_LABEL or a GIMPLE_COND. */
static bool
-if_convertible_stmt_p (gimple stmt, VEC (data_reference_p, heap) *refs)
+if_convertible_stmt_p (gimple stmt, VEC (data_reference_p, heap) *refs,
+ bool *any_mask_load_store)
{
switch (gimple_code (stmt))
{
@@ -766,7 +870,8 @@ if_convertible_stmt_p (gimple stmt, VEC
return true;
case GIMPLE_ASSIGN:
- return if_convertible_gimple_assign_stmt_p (stmt, refs);
+ return if_convertible_gimple_assign_stmt_p (stmt, refs,
+ any_mask_load_store);
case GIMPLE_CALL:
{
@@ -1072,7 +1177,8 @@ static bool
if_convertible_loop_p_1 (struct loop *loop,
VEC (loop_p, heap) **loop_nest,
VEC (data_reference_p, heap) **refs,
- VEC (ddr_p, heap) **ddrs)
+ VEC (ddr_p, heap) **ddrs,
+ bool *any_mask_load_store)
{
bool res;
unsigned int i;
@@ -1128,17 +1234,27 @@ if_convertible_loop_p_1 (struct loop *lo
basic_block bb = ifc_bbs[i];
gimple_stmt_iterator itr;
- for (itr = gsi_start_phis (bb); !gsi_end_p (itr); gsi_next (&itr))
- if (!if_convertible_phi_p (loop, bb, gsi_stmt (itr)))
- return false;
-
/* Check the if-convertibility of statements in predicated BBs. */
if (is_predicated (bb))
for (itr = gsi_start_bb (bb); !gsi_end_p (itr); gsi_next (&itr))
- if (!if_convertible_stmt_p (gsi_stmt (itr), *refs))
+ if (!if_convertible_stmt_p (gsi_stmt (itr), *refs,
+ any_mask_load_store))
return false;
}
+ /* Checking PHIs needs to be done after stmts, as the fact whether there
+ are any masked loads or stores affects the tests. */
+ for (i = 0; i < loop->num_nodes; i++)
+ {
+ basic_block bb = ifc_bbs[i];
+ gimple_stmt_iterator itr;
+
+ for (itr = gsi_start_phis (bb); !gsi_end_p (itr); gsi_next (&itr))
+ if (!if_convertible_phi_p (loop, bb, gsi_stmt (itr),
+ *any_mask_load_store))
+ return false;
+ }
+
if (dump_file)
fprintf (dump_file, "Applying if-conversion\n");
@@ -1154,7 +1270,7 @@ if_convertible_loop_p_1 (struct loop *lo
- if its basic blocks and phi nodes are if convertible. */
static bool
-if_convertible_loop_p (struct loop *loop)
+if_convertible_loop_p (struct loop *loop, bool *any_mask_load_store)
{
edge e;
edge_iterator ei;
@@ -1196,7 +1312,8 @@ if_convertible_loop_p (struct loop *loop
refs = VEC_alloc (data_reference_p, heap, 5);
ddrs = VEC_alloc (ddr_p, heap, 25);
loop_nest = VEC_alloc (loop_p, heap, 3);
- res = if_convertible_loop_p_1 (loop, &loop_nest, &refs, &ddrs);
+ res = if_convertible_loop_p_1 (loop, &loop_nest, &refs, &ddrs,
+ any_mask_load_store);
if (flag_tree_loop_if_convert_stores)
{
@@ -1414,7 +1531,7 @@ predicate_all_scalar_phis (struct loop *
gimplification of the predicates. */
static void
-insert_gimplified_predicates (loop_p loop)
+insert_gimplified_predicates (loop_p loop, bool any_mask_load_store)
{
unsigned int i;
@@ -1435,7 +1552,8 @@ insert_gimplified_predicates (loop_p loo
stmts = bb_predicate_gimplified_stmts (bb);
if (stmts)
{
- if (flag_tree_loop_if_convert_stores)
+ if (flag_tree_loop_if_convert_stores
+ || any_mask_load_store)
{
/* Insert the predicate of the BB just after the label,
as the if-conversion of memory writes will use this
@@ -1594,9 +1712,49 @@ predicate_mem_writes (loop_p loop)
}
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
- if ((stmt = gsi_stmt (gsi))
- && gimple_assign_single_p (stmt)
- && gimple_vdef (stmt))
+ if ((stmt = gsi_stmt (gsi)) == NULL
+ || !gimple_assign_single_p (stmt))
+ continue;
+ else if (gimple_plf (stmt, GF_PLF_1))
+ {
+ tree lhs = gimple_assign_lhs (stmt);
+ tree rhs = gimple_assign_rhs1 (stmt);
+ tree ref, addr, ptr, masktype, mask_op0, mask_op1, mask;
+ gimple new_stmt;
+ int bitsize = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (lhs)));
+
+ masktype = build_nonstandard_integer_type (bitsize, 1);
+ mask_op0 = build_int_cst (masktype, swap ? 0 : -1);
+ mask_op1 = build_int_cst (masktype, swap ? -1 : 0);
+ ref = TREE_CODE (lhs) == SSA_NAME ? rhs : lhs;
+ addr = force_gimple_operand_gsi (&gsi, build_fold_addr_expr (ref),
+ true, NULL_TREE, true,
+ GSI_SAME_STMT);
+ cond = force_gimple_operand_gsi_1 (&gsi, unshare_expr (cond),
+ is_gimple_condexpr, NULL_TREE,
+ true, GSI_SAME_STMT);
+ mask = fold_build_cond_expr (masktype, unshare_expr (cond),
+ mask_op0, mask_op1);
+ mask = ifc_temp_var (masktype, mask, &gsi);
+ ptr = build_int_cst (reference_alias_ptr_type (ref), 0);
+ /* Copy points-to info if possible. */
+ if (TREE_CODE (addr) == SSA_NAME && !SSA_NAME_PTR_INFO (addr))
+ copy_ref_info (build2 (MEM_REF, TREE_TYPE (ref), addr, ptr),
+ ref);
+ if (TREE_CODE (lhs) == SSA_NAME)
+ {
+ new_stmt
+ = gimple_build_call_internal (IFN_MASK_LOAD, 3, addr,
+ ptr, mask);
+ gimple_call_set_lhs (new_stmt, lhs);
+ }
+ else
+ new_stmt
+ = gimple_build_call_internal (IFN_MASK_STORE, 4, addr, ptr,
+ mask, rhs);
+ gsi_replace (&gsi, new_stmt, false);
+ }
+ else if (gimple_vdef (stmt))
{
tree lhs = gimple_assign_lhs (stmt);
tree rhs = gimple_assign_rhs1 (stmt);
@@ -1666,7 +1824,7 @@ remove_conditions_and_labels (loop_p loo
blocks. Replace PHI nodes with conditional modify expressions. */
static void
-combine_blocks (struct loop *loop)
+combine_blocks (struct loop *loop, bool any_mask_load_store)
{
basic_block bb, exit_bb, merge_target_bb;
unsigned int orig_loop_num_nodes = loop->num_nodes;
@@ -1675,10 +1833,10 @@ combine_blocks (struct loop *loop)
edge_iterator ei;
remove_conditions_and_labels (loop);
- insert_gimplified_predicates (loop);
+ insert_gimplified_predicates (loop, any_mask_load_store);
predicate_all_scalar_phis (loop);
- if (flag_tree_loop_if_convert_stores)
+ if (flag_tree_loop_if_convert_stores || any_mask_load_store)
predicate_mem_writes (loop);
/* Merge basic blocks: first remove all the edges in the loop,
@@ -1775,23 +1933,25 @@ combine_blocks (struct loop *loop)
profitability analysis. Returns true when something changed. */
static bool
-tree_if_conversion (struct loop *loop)
+tree_if_conversion (struct loop *loop, bool *any_mask_load_store_p)
{
bool changed = false;
ifc_bbs = NULL;
+ bool any_mask_load_store = false;
- if (!if_convertible_loop_p (loop)
+ if (!if_convertible_loop_p (loop, &any_mask_load_store)
|| !dbg_cnt (if_conversion_tree))
goto cleanup;
/* Now all statements are if-convertible. Combine all the basic
blocks into one huge basic block doing the if-conversion
on-the-fly. */
- combine_blocks (loop);
+ combine_blocks (loop, any_mask_load_store);
- if (flag_tree_loop_if_convert_stores)
+ if (flag_tree_loop_if_convert_stores || any_mask_load_store)
mark_virtual_operands_for_renaming (cfun);
+ *any_mask_load_store_p |= any_mask_load_store;
changed = true;
cleanup:
@@ -1809,6 +1969,9 @@ tree_if_conversion (struct loop *loop)
return changed;
}
+/* Flag whether if-unconversion pass will be needed afterwards. */
+static bool need_if_unconversion;
+
/* Tree if-conversion pass management. */
static unsigned int
@@ -1818,17 +1981,20 @@ main_tree_if_conversion (void)
struct loop *loop;
bool changed = false;
unsigned todo = 0;
+ bool any_mask_load_store = false;
if (number_of_loops () <= 1)
return 0;
FOR_EACH_LOOP (li, loop, 0)
- changed |= tree_if_conversion (loop);
+ changed |= tree_if_conversion (loop, &any_mask_load_store);
+
+ need_if_unconversion = any_mask_load_store;
if (changed)
todo |= TODO_cleanup_cfg;
- if (changed && flag_tree_loop_if_convert_stores)
+ if (changed && (flag_tree_loop_if_convert_stores || any_mask_load_store))
todo |= TODO_update_ssa_only_virtuals;
free_dominance_info (CDI_POST_DOMINATORS);
@@ -1865,6 +2031,139 @@ struct gimple_opt_pass pass_if_conversio
NULL, /* sub */
NULL, /* next */
0, /* static_pass_number */
+ TV_NONE, /* tv_id */
+ PROP_cfg | PROP_ssa, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_verify_stmts | TODO_verify_flow
+ /* todo_flags_finish */
+ }
+};
+
+/* Undo creation of MASK_LOAD or MASK_STORE, if it hasn't
+ been successfully vectorized. */
+
+static bool
+gate_tree_if_unconversion (void)
+{
+ return need_if_unconversion;
+}
+
+static unsigned int
+main_tree_if_unconversion (void)
+{
+ basic_block bb;
+ gimple_stmt_iterator gsi;
+
+ need_if_unconversion = false;
+ FOR_EACH_BB (bb)
+ for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple stmt = gsi_stmt (gsi);
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+ || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)
+ && INTEGRAL_TYPE_P (TREE_TYPE (gimple_call_arg (stmt, 2))))
+ {
+ tree cond = gimple_call_arg (stmt, 2), mem, type;
+ edge e1, e2, e3;
+ bool swapped_p = false;
+ gimple cond_stmt, new_stmt;
+
+ if (TREE_CODE (cond) == SSA_NAME
+ && !SSA_NAME_IS_DEFAULT_DEF (cond))
+ {
+ gimple def_stmt = SSA_NAME_DEF_STMT (cond);
+ if (is_gimple_assign (def_stmt)
+ && gimple_bb (def_stmt) == bb
+ && gimple_assign_rhs_code (def_stmt) == COND_EXPR)
+ {
+ tree rhs2 = gimple_assign_rhs2 (def_stmt);
+ tree rhs3 = gimple_assign_rhs3 (def_stmt);
+ if (integer_all_onesp (rhs2) && integer_zerop (rhs3))
+ cond = gimple_assign_rhs1 (def_stmt);
+ else if (integer_zerop (rhs2) && integer_all_onesp (rhs3))
+ {
+ cond = gimple_assign_rhs1 (def_stmt);
+ swapped_p = true;
+ }
+ }
+ }
+ gsi_prev (&gsi);
+ e1 = split_block (bb, gsi_stmt (gsi));
+ e2 = split_block (e1->dest, stmt);
+ e3 = make_edge (e1->src, e2->dest,
+ swapped_p ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE);
+ e1->flags = (e1->flags & ~EDGE_FALLTHRU)
+ | (swapped_p ? EDGE_FALSE_VALUE : EDGE_TRUE_VALUE);
+ set_immediate_dominator (CDI_DOMINATORS, e2->dest, e1->src);
+ if (cond == gimple_call_arg (stmt, 2))
+ cond_stmt
+ = gimple_build_cond (NE_EXPR, cond,
+ build_int_cst (TREE_TYPE (cond), 0),
+ NULL_TREE, NULL_TREE);
+ else
+ cond_stmt
+ = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE);
+ gsi = gsi_last_bb (e1->src);
+ gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT);
+ if (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD)
+ type = TREE_TYPE (gimple_call_lhs (stmt));
+ else
+ type = TREE_TYPE (gimple_call_arg (stmt, 3));
+ mem = build2 (MEM_REF, type, gimple_call_arg (stmt, 0),
+ gimple_call_arg (stmt, 1));
+ if (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD)
+ new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
+ mem);
+ else
+ new_stmt = gimple_build_assign (mem, gimple_call_arg (stmt, 3));
+ gimple_set_vuse (new_stmt, gimple_vuse (stmt));
+ if (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD)
+ {
+ gimple phi;
+ tree res = gimple_assign_lhs (new_stmt);
+ tree tem = make_ssa_name (TREE_TYPE (res), NULL);
+ tree zero = build_zero_cst (TREE_TYPE (res));
+ gimple_assign_set_lhs (new_stmt, tem);
+ gimple_call_set_lhs (stmt, NULL_TREE);
+ phi = create_phi_node (res, e2->dest);
+ add_phi_arg (phi, tem, e2, gimple_location (stmt));
+ add_phi_arg (phi, zero, e3, gimple_location (stmt));
+ SSA_NAME_DEF_STMT (res) = phi;
+ }
+ else
+ {
+ gimple phi;
+ tree new_vdef = copy_ssa_name (gimple_vuse (stmt), new_stmt);
+ gimple_set_vdef (new_stmt, new_vdef);
+ phi = create_phi_node (gimple_vdef (stmt), e2->dest);
+ add_phi_arg (phi, new_vdef, e2, UNKNOWN_LOCATION);
+ add_phi_arg (phi, gimple_vuse (stmt), e3, UNKNOWN_LOCATION);
+ SSA_NAME_DEF_STMT (gimple_vdef (stmt)) = phi;
+ }
+ gsi = gsi_for_stmt (stmt);
+ gsi_replace (&gsi, new_stmt, false);
+ gsi = gsi_for_stmt (cond_stmt);
+ }
+ }
+
+ return 0;
+}
+
+struct gimple_opt_pass pass_if_unconversion =
+{
+ {
+ GIMPLE_PASS,
+ "ifuncvt", /* name */
+ OPTGROUP_NONE, /* optinfo_flags */
+ gate_tree_if_unconversion, /* gate */
+ main_tree_if_unconversion, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
TV_NONE, /* tv_id */
PROP_cfg | PROP_ssa, /* properties_required */
0, /* properties_provided */
--- gcc/tree-vect-data-refs.c.jj 2012-11-02 09:01:47.097575423 +0100
+++ gcc/tree-vect-data-refs.c 2012-11-02 09:29:49.647621187 +0100
@@ -2707,6 +2707,24 @@ vect_check_gather (gimple stmt, loop_vec
enum machine_mode pmode;
int punsignedp, pvolatilep;
+ base = DR_REF (dr);
+ /* For masked loads/stores, DR_REF (dr) is an artificial MEM_REF,
+ see if we can use the def stmt of the address. */
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+ || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)
+ && TREE_CODE (base) == MEM_REF
+ && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME
+ && integer_zerop (TREE_OPERAND (base, 1))
+ && !expr_invariant_in_loop_p (loop, TREE_OPERAND (base, 0)))
+ {
+ gimple def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (base, 0));
+ if (is_gimple_assign (def_stmt)
+ && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR)
+ base = TREE_OPERAND (gimple_assign_rhs1 (def_stmt), 0);
+ }
+
/* The gather builtins need address of the form
loop_invariant + vector * {1, 2, 4, 8}
or
@@ -2719,7 +2737,7 @@ vect_check_gather (gimple stmt, loop_vec
vectorized. The following code attempts to find such a preexistng
SSA_NAME OFF and put the loop invariants into a tree BASE
that can be gimplified before the loop. */
- base = get_inner_reference (DR_REF (dr), &pbitsize, &pbitpos, &off,
+ base = get_inner_reference (base, &pbitsize, &pbitpos, &off,
&pmode, &punsignedp, &pvolatilep, false);
gcc_assert (base != NULL_TREE && (pbitpos % BITS_PER_UNIT) == 0);
@@ -3186,7 +3204,10 @@ vect_analyze_data_refs (loop_vec_info lo
offset = unshare_expr (DR_OFFSET (dr));
init = unshare_expr (DR_INIT (dr));
- if (is_gimple_call (stmt))
+ if (is_gimple_call (stmt)
+ && (!gimple_call_internal_p (stmt)
+ || (gimple_call_internal_fn (stmt) != IFN_MASK_LOAD
+ && gimple_call_internal_fn (stmt) != IFN_MASK_STORE)))
{
if (dump_enabled_p ())
{
@@ -4893,6 +4914,14 @@ vect_supportable_dr_alignment (struct da
if (aligned_access_p (dr) && !check_aligned_accesses)
return dr_aligned;
+ /* For now assume all conditional loads/stores support unaligned
+ access without any special code. */
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+ || gimple_call_internal_fn (stmt) == IFN_MASK_STORE))
+ return dr_unaligned_supported;
+
if (loop_vinfo)
{
vect_loop = LOOP_VINFO_LOOP (loop_vinfo);
--- gcc/gimple.h.jj 2012-11-02 09:01:47.206574700 +0100
+++ gcc/gimple.h 2012-11-02 09:13:34.211881846 +0100
@@ -4938,7 +4938,13 @@ gimple_expr_type (const_gimple stmt)
useless conversion involved. That means returning the
original RHS type as far as we can reconstruct it. */
if (code == GIMPLE_CALL)
- type = gimple_call_return_type (stmt);
+ {
+ if (gimple_call_internal_p (stmt)
+ && gimple_call_internal_fn (stmt) == IFN_MASK_STORE)
+ type = TREE_TYPE (gimple_call_arg (stmt, 3));
+ else
+ type = gimple_call_return_type (stmt);
+ }
else
switch (gimple_assign_rhs_code (stmt))
{
--- gcc/internal-fn.c.jj 2012-11-02 09:01:47.582572197 +0100
+++ gcc/internal-fn.c 2012-11-02 10:06:36.167504880 +0100
@@ -1,5 +1,5 @@
/* Internal functions.
- Copyright (C) 2011 Free Software Foundation, Inc.
+ Copyright (C) 2011, 2012 Free Software Foundation, Inc.
This file is part of GCC.
@@ -109,6 +109,52 @@ expand_STORE_LANES (gimple stmt)
expand_insn (get_multi_vector_move (type, vec_store_lanes_optab), 2, ops);
}
+static void
+expand_MASK_LOAD (gimple stmt)
+{
+ struct expand_operand ops[3];
+ tree type, lhs, rhs, maskt;
+ rtx mem, target, mask;
+
+ maskt = gimple_call_arg (stmt, 2);
+ lhs = gimple_call_lhs (stmt);
+ type = TREE_TYPE (lhs);
+ rhs = build2 (MEM_REF, type, gimple_call_arg (stmt, 0),
+ gimple_call_arg (stmt, 1));
+
+ mem = expand_expr (rhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
+ gcc_assert (MEM_P (mem));
+ mask = expand_normal (maskt);
+ target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
+ create_output_operand (&ops[0], target, TYPE_MODE (type));
+ create_fixed_operand (&ops[1], mem);
+ create_input_operand (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt)));
+ expand_insn (optab_handler (maskload_optab, TYPE_MODE (type)), 3, ops);
+}
+
+static void
+expand_MASK_STORE (gimple stmt)
+{
+ struct expand_operand ops[3];
+ tree type, lhs, rhs, maskt;
+ rtx mem, reg, mask;
+
+ maskt = gimple_call_arg (stmt, 2);
+ rhs = gimple_call_arg (stmt, 3);
+ type = TREE_TYPE (rhs);
+ lhs = build2 (MEM_REF, type, gimple_call_arg (stmt, 0),
+ gimple_call_arg (stmt, 1));
+
+ mem = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
+ gcc_assert (MEM_P (mem));
+ mask = expand_normal (maskt);
+ reg = expand_normal (rhs);
+ create_fixed_operand (&ops[0], mem);
+ create_input_operand (&ops[1], reg, TYPE_MODE (type));
+ create_input_operand (&ops[2], mask, TYPE_MODE (TREE_TYPE (maskt)));
+ expand_insn (optab_handler (maskstore_optab, TYPE_MODE (type)), 3, ops);
+}
+
/* Routines to expand each internal function, indexed by function number.
Each routine has the prototype:
--- gcc/tree-vect-loop.c.jj 2012-11-02 09:01:47.543572462 +0100
+++ gcc/tree-vect-loop.c 2012-11-02 09:13:34.209881860 +0100
@@ -351,7 +351,11 @@ vect_determine_vectorization_factor (loo
analyze_pattern_stmt = false;
}
- if (gimple_get_lhs (stmt) == NULL_TREE)
+ if (gimple_get_lhs (stmt) == NULL_TREE
+ /* MASK_STORE has no lhs, but is ok. */
+ && (!is_gimple_call (stmt)
+ || !gimple_call_internal_p (stmt)
+ || gimple_call_internal_fn (stmt) != IFN_MASK_STORE))
{
if (dump_enabled_p ())
{
@@ -388,7 +392,12 @@ vect_determine_vectorization_factor (loo
else
{
gcc_assert (!STMT_VINFO_DATA_REF (stmt_info));
- scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && gimple_call_internal_fn (stmt) == IFN_MASK_STORE)
+ scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3));
+ else
+ scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
--- gcc/passes.c.jj 2012-11-01 09:33:29.000000000 +0100
+++ gcc/passes.c 2012-11-02 16:26:09.452391414 +0100
@@ -1484,6 +1484,7 @@ init_optimization_passes (void)
struct opt_pass **p = &pass_vectorize.pass.sub;
NEXT_PASS (pass_dce_loop);
}
+ NEXT_PASS (pass_if_unconversion);
NEXT_PASS (pass_predcom);
NEXT_PASS (pass_complete_unroll);
NEXT_PASS (pass_slp_vectorize);
--- gcc/optabs.def.jj 2012-11-02 09:01:47.148575082 +0100
+++ gcc/optabs.def 2012-11-02 09:13:34.205881885 +0100
@@ -242,6 +242,8 @@ OPTAB_D (sdot_prod_optab, "sdot_prod$I$a
OPTAB_D (ssum_widen_optab, "widen_ssum$I$a3")
OPTAB_D (udot_prod_optab, "udot_prod$I$a")
OPTAB_D (usum_widen_optab, "widen_usum$I$a3")
+OPTAB_D (maskload_optab, "maskload$a")
+OPTAB_D (maskstore_optab, "maskstore$a")
OPTAB_D (vec_extract_optab, "vec_extract$a")
OPTAB_D (vec_init_optab, "vec_init$a")
OPTAB_D (vec_pack_sfix_trunc_optab, "vec_pack_sfix_trunc_$a")
--- gcc/tree-pass.h.jj 2012-11-01 09:33:25.000000000 +0100
+++ gcc/tree-pass.h 2012-11-02 11:52:38.598409785 +0100
@@ -1,5 +1,5 @@
/* Definitions for describing one tree-ssa optimization pass.
- Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011
+ Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
Free Software Foundation, Inc.
Contributed by Richard Henderson <rth@redhat.com>
@@ -284,6 +284,7 @@ extern struct gimple_opt_pass pass_recor
extern struct gimple_opt_pass pass_graphite;
extern struct gimple_opt_pass pass_graphite_transforms;
extern struct gimple_opt_pass pass_if_conversion;
+extern struct gimple_opt_pass pass_if_unconversion;
extern struct gimple_opt_pass pass_loop_distribution;
extern struct gimple_opt_pass pass_vectorize;
extern struct gimple_opt_pass pass_slp_vectorize;
--- gcc/tree-vect-stmts.c.jj 2012-11-02 09:01:47.555572381 +0100
+++ gcc/tree-vect-stmts.c 2012-11-02 10:32:47.782340547 +0100
@@ -218,7 +218,7 @@ vect_mark_relevant (VEC(gimple,heap) **w
/* This use is out of pattern use, if LHS has other uses that are
pattern uses, we should mark the stmt itself, and not the pattern
stmt. */
- if (TREE_CODE (lhs) == SSA_NAME)
+ if (lhs && TREE_CODE (lhs) == SSA_NAME)
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs)
{
if (is_gimple_debug (USE_STMT (use_p)))
@@ -376,7 +376,27 @@ exist_non_indexing_operands_for_use_p (t
first case, and whether var corresponds to USE. */
if (!gimple_assign_copy_p (stmt))
- return false;
+ {
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt))
+ switch (gimple_call_internal_fn (stmt))
+ {
+ case IFN_MASK_STORE:
+ operand = gimple_call_arg (stmt, 3);
+ if (operand == use)
+ return true;
+ /* FALLTHRU */
+ case IFN_MASK_LOAD:
+ operand = gimple_call_arg (stmt, 2);
+ if (operand == use)
+ return true;
+ break;
+ default:
+ break;
+ }
+ return false;
+ }
+
if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
return false;
operand = gimple_assign_rhs1 (stmt);
@@ -1693,6 +1713,401 @@ vectorizable_function (gimple call, tree
vectype_in);
}
+
+static tree permute_vec_elements (tree, tree, tree, gimple,
+ gimple_stmt_iterator *);
+
+
+static bool
+vectorizable_mask_load_store (gimple stmt, gimple_stmt_iterator *gsi,
+ gimple *vec_stmt, slp_tree slp_node)
+{
+ tree vec_dest = NULL;
+ stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+ stmt_vec_info prev_stmt_info;
+ loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
+ struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
+ struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
+ tree vectype = STMT_VINFO_VECTYPE (stmt_info);
+ tree elem_type;
+ gimple new_stmt;
+ tree dummy;
+ tree dataref_ptr = NULL_TREE;
+ gimple ptr_incr;
+ int nunits = TYPE_VECTOR_SUBPARTS (vectype);
+ int ncopies;
+ int i, j;
+ bool inv_p;
+ tree gather_base = NULL_TREE, gather_off = NULL_TREE;
+ tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE;
+ int gather_scale = 1;
+ enum vect_def_type gather_dt = vect_unknown_def_type;
+ bool is_store;
+ tree mask;
+ gimple def_stmt;
+ tree def;
+ enum vect_def_type dt;
+
+ if (slp_node != NULL)
+ return false;
+
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
+ gcc_assert (ncopies >= 1);
+
+ is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE;
+ mask = gimple_call_arg (stmt, 2);
+ if (TYPE_PRECISION (TREE_TYPE (mask))
+ != GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))))
+ return false;
+
+ /* FORNOW. This restriction should be relaxed. */
+ if (nested_in_vect_loop && ncopies > 1)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+ "multiple types in nested loop.");
+ return false;
+ }
+
+ if (!STMT_VINFO_RELEVANT_P (stmt_info))
+ return false;
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def)
+ return false;
+
+ if (!STMT_VINFO_DATA_REF (stmt_info))
+ return false;
+
+ elem_type = TREE_TYPE (vectype);
+
+ if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
+ return false;
+
+ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
+ return false;
+
+ if (STMT_VINFO_GATHER_P (stmt_info))
+ {
+ gimple def_stmt;
+ tree def;
+ gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base,
+ &gather_off, &gather_scale);
+ gcc_assert (gather_decl);
+ if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, NULL,
+ &def_stmt, &def, &gather_dt,
+ &gather_off_vectype))
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+ "gather index use not simple.");
+ return false;
+ }
+ }
+ else if (tree_int_cst_compare (nested_in_vect_loop
+ ? STMT_VINFO_DR_STEP (stmt_info)
+ : DR_STEP (dr), size_zero_node) < 0)
+ return false;
+ else if (optab_handler (is_store ? maskstore_optab : maskload_optab,
+ TYPE_MODE (vectype)) == CODE_FOR_nothing)
+ return false;
+
+ if (TREE_CODE (mask) != SSA_NAME)
+ return false;
+
+ if (!vect_is_simple_use (mask, stmt, loop_vinfo, NULL,
+ &def_stmt, &def, &dt))
+ return false;
+
+ if (is_store)
+ {
+ tree rhs = gimple_call_arg (stmt, 3);
+ if (!vect_is_simple_use (rhs, stmt, loop_vinfo, NULL,
+ &def_stmt, &def, &dt))
+ return false;
+ }
+
+ if (!vec_stmt) /* transformation not required. */
+ {
+ STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
+ return true;
+ }
+
+ /** Transform. **/
+
+ if (STMT_VINFO_GATHER_P (stmt_info))
+ {
+ tree vec_oprnd0 = NULL_TREE, op;
+ tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl));
+ tree rettype, srctype, ptrtype, idxtype, masktype, scaletype;
+ tree ptr, vec_mask = NULL_TREE, mask_op, var, scale;
+ tree perm_mask = NULL_TREE, prev_res = NULL_TREE;
+ edge pe = loop_preheader_edge (loop);
+ gimple_seq seq;
+ basic_block new_bb;
+ enum { NARROW, NONE, WIDEN } modifier;
+ int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype);
+
+ if (nunits == gather_off_nunits)
+ modifier = NONE;
+ else if (nunits == gather_off_nunits / 2)
+ {
+ unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits);
+ modifier = WIDEN;
+
+ for (i = 0; i < gather_off_nunits; ++i)
+ sel[i] = i | nunits;
+
+ perm_mask = vect_gen_perm_mask (gather_off_vectype, sel);
+ gcc_assert (perm_mask != NULL_TREE);
+ }
+ else if (nunits == gather_off_nunits * 2)
+ {
+ unsigned char *sel = XALLOCAVEC (unsigned char, nunits);
+ modifier = NARROW;
+
+ for (i = 0; i < nunits; ++i)
+ sel[i] = i < gather_off_nunits
+ ? i : i + nunits - gather_off_nunits;
+
+ perm_mask = vect_gen_perm_mask (vectype, sel);
+ gcc_assert (perm_mask != NULL_TREE);
+ ncopies *= 2;
+ }
+ else
+ gcc_unreachable ();
+
+ rettype = TREE_TYPE (TREE_TYPE (gather_decl));
+ srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
+ ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
+ idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
+ masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
+ scaletype = TREE_VALUE (arglist);
+ gcc_checking_assert (types_compatible_p (srctype, rettype)
+ && types_compatible_p (srctype, masktype));
+
+ vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype);
+
+ ptr = fold_convert (ptrtype, gather_base);
+ if (!is_gimple_min_invariant (ptr))
+ {
+ ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE);
+ new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
+ gcc_assert (!new_bb);
+ }
+
+ scale = build_int_cst (scaletype, gather_scale);
+
+ prev_stmt_info = NULL;
+ for (j = 0; j < ncopies; ++j)
+ {
+ if (modifier == WIDEN && (j & 1))
+ op = permute_vec_elements (vec_oprnd0, vec_oprnd0,
+ perm_mask, stmt, gsi);
+ else if (j == 0)
+ op = vec_oprnd0
+ = vect_get_vec_def_for_operand (gather_off, stmt, NULL);
+ else
+ op = vec_oprnd0
+ = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0);
+
+ if (!useless_type_conversion_p (idxtype, TREE_TYPE (op)))
+ {
+ gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op))
+ == TYPE_VECTOR_SUBPARTS (idxtype));
+ var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL);
+ var = make_ssa_name (var, NULL);
+ op = build1 (VIEW_CONVERT_EXPR, idxtype, op);
+ new_stmt
+ = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var,
+ op, NULL_TREE);
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+ op = var;
+ }
+
+ if (j == 0)
+ vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL);
+ else
+ {
+ vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt,
+ &def, &dt);
+ vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask);
+ }
+
+ mask_op = vec_mask;
+ if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask)))
+ {
+ gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op))
+ == TYPE_VECTOR_SUBPARTS (masktype));
+ var = vect_get_new_vect_var (masktype, vect_simple_var, NULL);
+ var = make_ssa_name (var, NULL);
+ mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op);
+ new_stmt
+ = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var,
+ mask_op, NULL_TREE);
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+ mask_op = var;
+ }
+
+ new_stmt
+ = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op,
+ scale);
+
+ if (!useless_type_conversion_p (vectype, rettype))
+ {
+ gcc_assert (TYPE_VECTOR_SUBPARTS (vectype)
+ == TYPE_VECTOR_SUBPARTS (rettype));
+ var = vect_get_new_vect_var (rettype, vect_simple_var, NULL);
+ op = make_ssa_name (var, new_stmt);
+ gimple_call_set_lhs (new_stmt, op);
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+ var = make_ssa_name (vec_dest, NULL);
+ op = build1 (VIEW_CONVERT_EXPR, vectype, op);
+ new_stmt
+ = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op,
+ NULL_TREE);
+ }
+ else
+ {
+ var = make_ssa_name (vec_dest, new_stmt);
+ gimple_call_set_lhs (new_stmt, var);
+ }
+
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+
+ if (modifier == NARROW)
+ {
+ if ((j & 1) == 0)
+ {
+ prev_res = var;
+ continue;
+ }
+ var = permute_vec_elements (prev_res, var,
+ perm_mask, stmt, gsi);
+ new_stmt = SSA_NAME_DEF_STMT (var);
+ }
+
+ if (prev_stmt_info == NULL)
+ STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
+ else
+ STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
+ prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
+ return true;
+ }
+ else if (is_store)
+ {
+ tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE;
+ prev_stmt_info = NULL;
+ for (i = 0; i < ncopies; i++)
+ {
+ unsigned align, misalign;
+
+ if (i == 0)
+ {
+ tree rhs = gimple_call_arg (stmt, 3);
+ vec_rhs = vect_get_vec_def_for_operand (rhs, stmt, NULL);
+ vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL);
+ /* We should have catched mismatched types earlier. */
+ gcc_assert (useless_type_conversion_p (vectype,
+ TREE_TYPE (vec_rhs)));
+ dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL,
+ NULL_TREE, &dummy, gsi,
+ &ptr_incr, false, &inv_p);
+ gcc_assert (!inv_p);
+ }
+ else
+ {
+ vect_is_simple_use (vec_rhs, NULL, loop_vinfo, NULL, &def_stmt,
+ &def, &dt);
+ vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs);
+ vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt,
+ &def, &dt);
+ vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask);
+ dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
+ TYPE_SIZE_UNIT (vectype));
+ }
+
+ align = TYPE_ALIGN_UNIT (vectype);
+ if (aligned_access_p (dr))
+ misalign = 0;
+ else if (DR_MISALIGNMENT (dr) == -1)
+ {
+ align = TYPE_ALIGN_UNIT (elem_type);
+ misalign = 0;
+ }
+ else
+ misalign = DR_MISALIGNMENT (dr);
+ set_ptr_info_alignment (get_ptr_info (dataref_ptr), align,
+ misalign);
+ new_stmt
+ = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr,
+ gimple_call_arg (stmt, 1),
+ vec_mask, vec_rhs);
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+ if (i == 0)
+ STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
+ else
+ STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
+ prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
+ }
+ else
+ {
+ tree vec_mask = NULL_TREE;
+ prev_stmt_info = NULL;
+ vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype);
+ for (i = 0; i < ncopies; i++)
+ {
+ unsigned align, misalign;
+
+ if (i == 0)
+ {
+ vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL);
+ dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL,
+ NULL_TREE, &dummy, gsi,
+ &ptr_incr, false, &inv_p);
+ gcc_assert (!inv_p);
+ }
+ else
+ {
+ vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt,
+ &def, &dt);
+ vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask);
+ dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt,
+ TYPE_SIZE_UNIT (vectype));
+ }
+
+ align = TYPE_ALIGN_UNIT (vectype);
+ if (aligned_access_p (dr))
+ misalign = 0;
+ else if (DR_MISALIGNMENT (dr) == -1)
+ {
+ align = TYPE_ALIGN_UNIT (elem_type);
+ misalign = 0;
+ }
+ else
+ misalign = DR_MISALIGNMENT (dr);
+ set_ptr_info_alignment (get_ptr_info (dataref_ptr), align,
+ misalign);
+ new_stmt
+ = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr,
+ gimple_call_arg (stmt, 1),
+ vec_mask);
+ gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest, NULL));
+ vect_finish_stmt_generation (stmt, new_stmt, gsi);
+ if (i == 0)
+ STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
+ else
+ STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
+ prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
+ }
+
+ return true;
+}
+
+
/* Function vectorizable_call.
Check if STMT performs a function call that can be vectorized.
@@ -1735,10 +2150,16 @@ vectorizable_call (gimple stmt, gimple_s
if (!is_gimple_call (stmt))
return false;
- if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
+ if (stmt_can_throw_internal (stmt))
return false;
- if (stmt_can_throw_internal (stmt))
+ if (gimple_call_internal_p (stmt)
+ && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+ || gimple_call_internal_fn (stmt) == IFN_MASK_STORE))
+ return vectorizable_mask_load_store (stmt, gsi, vec_stmt,
+ slp_node);
+
+ if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
return false;
vectype_out = STMT_VINFO_VECTYPE (stmt_info);
@@ -3442,10 +3863,6 @@ vectorizable_shift (gimple stmt, gimple_
}
-static tree permute_vec_elements (tree, tree, tree, gimple,
- gimple_stmt_iterator *);
-
-
/* Function vectorizable_operation.
Check if STMT performs a binary, unary or ternary operation that can
@@ -5836,6 +6253,10 @@ vect_transform_stmt (gimple stmt, gimple
case call_vec_info_type:
done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node);
stmt = gsi_stmt (*gsi);
+ if (is_gimple_call (stmt)
+ && gimple_call_internal_p (stmt)
+ && gimple_call_internal_fn (stmt) == IFN_MASK_STORE)
+ is_store = true;
break;
case reduc_vec_info_type:
Jakub