Hi Richard,
For Aarch64 target, I was trying to vectorize the expression
"arr[i]=arr[i]*4;" via vector shifts instructions since they don't have
vector mults.
unsigned long int __attribute__ ((aligned (64)))arr[100];
int i;
#if 1
void test_vector_shifts()
{
for(i=0; i<=99;i++)
arr[i]=arr[i]<<2;
}
#endif
void test_vectorshift_via_mul()
{
for(i=0; i<=99;i++)
arr[i]=arr[i]*4;
}
I found a similar PR and your comments
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65952#c6.
Based on that and IRC discussion I had with you, I added vector recog pattern
that transforms mults to shifts. The vectorizer is now able to generate vector
shifts for the above test case.
PR case also gets vectorized
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65952#c10.
This is just an initial patch and tries to optimize integer type power 2
constants. I wanted to get feedback on this . I bootstrapped and reg tested
on aarch64-none-linux-gnu .
Regards,
Venkat.
diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
index f034635..948203d 100644
--- a/gcc/tree-vect-patterns.c
+++ b/gcc/tree-vect-patterns.c
@@ -76,6 +76,10 @@ static gimple vect_recog_vector_vector_shift_pattern
(vec<gimple> *,
tree *, tree *);
static gimple vect_recog_divmod_pattern (vec<gimple> *,
tree *, tree *);
+
+static gimple vect_recog_multconst_pattern (vec<gimple> *,
+ tree *, tree *);
+
static gimple vect_recog_mixed_size_cond_pattern (vec<gimple> *,
tree *, tree *);
static gimple vect_recog_bool_pattern (vec<gimple> *, tree *, tree *);
@@ -90,6 +94,7 @@ static vect_recog_func_ptr
vect_vect_recog_func_ptrs[NUM_PATTERNS] = {
vect_recog_rotate_pattern,
vect_recog_vector_vector_shift_pattern,
vect_recog_divmod_pattern,
+ vect_recog_multconst_pattern,
vect_recog_mixed_size_cond_pattern,
vect_recog_bool_pattern};
@@ -2147,6 +2152,87 @@ vect_recog_vector_vector_shift_pattern (vec<gimple>
*stmts,
return pattern_stmt;
}
+static gimple
+vect_recog_multconst_pattern (vec<gimple> *stmts,
+ tree *type_in, tree *type_out)
+{
+ gimple last_stmt = stmts->pop ();
+ tree oprnd0, oprnd1, vectype, itype;
+ gimple pattern_stmt;
+ enum tree_code rhs_code;
+ optab optab;
+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
+
+ if (!is_gimple_assign (last_stmt))
+ return NULL;
+
+ rhs_code = gimple_assign_rhs_code (last_stmt);
+ switch (rhs_code)
+ {
+ case MULT_EXPR:
+ break;
+ default:
+ return NULL;
+ }
+
+ if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
+ return NULL;
+
+ oprnd0 = gimple_assign_rhs1 (last_stmt);
+ oprnd1 = gimple_assign_rhs2 (last_stmt);
+ itype = TREE_TYPE (oprnd0);
+
+ if (TREE_CODE (oprnd0) != SSA_NAME
+ || TREE_CODE (oprnd1) != INTEGER_CST
+ || TREE_CODE (itype) != INTEGER_TYPE
+ || TYPE_PRECISION (itype) != GET_MODE_PRECISION (TYPE_MODE (itype)))
+ return NULL;
+
+ vectype = get_vectype_for_scalar_type (itype);
+ if (vectype == NULL_TREE)
+ return NULL;
+
+ /* If the target can handle vectorized multiplication natively,
+ don't attempt to optimize this. */
+ optab = optab_for_tree_code (rhs_code, vectype, optab_default);
+ if (optab != unknown_optab)
+ {
+ machine_mode vec_mode = TYPE_MODE (vectype);
+ int icode = (int) optab_handler (optab, vec_mode);
+ if (icode != CODE_FOR_nothing)
+ return NULL;
+ }
+
+ /* If target cannot handle vector left shift then we cannot
+ optimize and bail out. */
+ optab = optab_for_tree_code (LSHIFT_EXPR, vectype, optab_vector);
+ if (!optab
+ || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing)
+ return NULL;
+
+ if (integer_pow2p (oprnd1))
+ {
+ /* Pattern detected. */
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "vect_recog_multconst_pattern: detected:\n");
+
+ tree shift;
+ shift = build_int_cst (itype, tree_log2 (oprnd1));
+ pattern_stmt = gimple_build_assign (vect_recog_temp_ssa_var (itype,
NULL),
+ LSHIFT_EXPR, oprnd0, shift);
+ if (dump_enabled_p ())
+ dump_gimple_stmt_loc (MSG_NOTE, vect_location, TDF_SLIM, pattern_stmt,
+ 0);
+ stmts->safe_push (last_stmt);
+ *type_in = vectype;
+ *type_out = vectype;
+ return pattern_stmt;
+ }
+
+ return NULL;
+}
+
/* Detect a signed division by a constant that wouldn't be
otherwise vectorized:
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 48c1f8d..833fe4b 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -1131,7 +1131,7 @@ extern void vect_slp_transform_bb (basic_block);
Additional pattern recognition functions can (and will) be added
in the future. */
typedef gimple (* vect_recog_func_ptr) (vec<gimple> *, tree *, tree *);
-#define NUM_PATTERNS 12
+#define NUM_PATTERNS 13
void vect_pattern_recog (loop_vec_info, bb_vec_info);
/* In tree-vectorizer.c. */
