From: juzhe.zhong@rivai.ai
To: gcc-patches@gcc.gnu.org
Cc: richard.sandiford@arm.com, rguenther@suse.de,
Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
Subject: [PATCH V2] SSA MATH: Support COND_LEN_FMA for floating-point math optimization
Date: Thu, 13 Jul 2023 16:54:34 +0800 [thread overview]
Message-ID: <20230713085434.3381643-1-juzhe.zhong@rivai.ai> (raw)
From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
Hi, Richard and Richi.
Previous patch we support COND_LEN_* binary operations. However, we didn't
support COND_LEN_* ternary.
Now, this patch support COND_LEN_* ternary. Consider this following case:
#define TEST_TYPE(TYPE) \
__attribute__ ((noipa)) void ternop_##TYPE (TYPE *__restrict dst, \
TYPE *__restrict a, \
TYPE *__restrict b,\
TYPE *__restrict c, int n) \
{ \
for (int i = 0; i < n; i++) \
dst[i] += a[i] * b[i]; \
}
#define TEST_ALL() TEST_TYPE (double)
TEST_ALL ()
Before this patch:
...
COND_LEN_MUL
COND_LEN_ADD
Afther this patch:
...
COND_LEN_FMA
gcc/ChangeLog:
* genmatch.cc (commutative_op): Add COND_LEN_*
* internal-fn.cc (first_commutative_argument): Ditto.
(CASE): Ditto.
(get_unconditional_internal_fn): Ditto.
(can_interpret_as_conditional_op_p): Ditto.
(internal_fn_len_index): Ditto.
* internal-fn.h (can_interpret_as_conditional_op_p): Ditt.
* tree-ssa-math-opts.cc (convert_mult_to_fma_1): Ditto.
(convert_mult_to_fma): Ditto.
(math_opts_dom_walker::after_dom_children): Ditto.
---
gcc/genmatch.cc | 13 ++++++
gcc/internal-fn.cc | 87 ++++++++++++++++++++++++++++++++++-----
gcc/internal-fn.h | 2 +-
gcc/tree-ssa-math-opts.cc | 80 +++++++++++++++++++++++++++++------
4 files changed, 159 insertions(+), 23 deletions(-)
diff --git a/gcc/genmatch.cc b/gcc/genmatch.cc
index 5fceeec9780..2302f2a7ff0 100644
--- a/gcc/genmatch.cc
+++ b/gcc/genmatch.cc
@@ -559,6 +559,19 @@ commutative_op (id_base *id)
case CFN_COND_FMS:
case CFN_COND_FNMA:
case CFN_COND_FNMS:
+ case CFN_COND_LEN_ADD:
+ case CFN_COND_LEN_MUL:
+ case CFN_COND_LEN_MIN:
+ case CFN_COND_LEN_MAX:
+ case CFN_COND_LEN_FMIN:
+ case CFN_COND_LEN_FMAX:
+ case CFN_COND_LEN_AND:
+ case CFN_COND_LEN_IOR:
+ case CFN_COND_LEN_XOR:
+ case CFN_COND_LEN_FMA:
+ case CFN_COND_LEN_FMS:
+ case CFN_COND_LEN_FNMA:
+ case CFN_COND_LEN_FNMS:
return 1;
default:
diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
index c11123a1173..e698f0bffc7 100644
--- a/gcc/internal-fn.cc
+++ b/gcc/internal-fn.cc
@@ -4191,6 +4191,19 @@ first_commutative_argument (internal_fn fn)
case IFN_COND_FMS:
case IFN_COND_FNMA:
case IFN_COND_FNMS:
+ case IFN_COND_LEN_ADD:
+ case IFN_COND_LEN_MUL:
+ case IFN_COND_LEN_MIN:
+ case IFN_COND_LEN_MAX:
+ case IFN_COND_LEN_FMIN:
+ case IFN_COND_LEN_FMAX:
+ case IFN_COND_LEN_AND:
+ case IFN_COND_LEN_IOR:
+ case IFN_COND_LEN_XOR:
+ case IFN_COND_LEN_FMA:
+ case IFN_COND_LEN_FMS:
+ case IFN_COND_LEN_FNMA:
+ case IFN_COND_LEN_FNMS:
return 1;
default:
@@ -4330,11 +4343,14 @@ conditional_internal_fn_code (internal_fn ifn)
{
switch (ifn)
{
-#define CASE(CODE, IFN) case IFN_COND_##IFN: return CODE;
- FOR_EACH_CODE_MAPPING(CASE)
+#define CASE(CODE, IFN) \
+ case IFN_COND_##IFN: \
+ case IFN_COND_LEN_##IFN: \
+ return CODE;
+ FOR_EACH_CODE_MAPPING (CASE)
#undef CASE
- default:
- return ERROR_MARK;
+ default:
+ return ERROR_MARK;
}
}
@@ -4433,6 +4449,18 @@ get_unconditional_internal_fn (internal_fn ifn)
operating elementwise if the operands are vectors. This includes
the case of an all-true COND, so that the operation always happens.
+ There is an alternative approach to interpret the STMT when the operands
+ are vectors which is the operation predicated by both conditional mask
+ and loop control length, the equivalent C code:
+
+ for (int i = 0; i < NUNTIS; i++)
+ {
+ if (i < LEN + BIAS && COND[i])
+ LHS[i] = A[i] CODE B[i];
+ else
+ LHS[i] = ELSE[i];
+ }
+
When returning true, set:
- *COND_OUT to the condition COND, or to NULL_TREE if the condition
@@ -4440,13 +4468,18 @@ get_unconditional_internal_fn (internal_fn ifn)
- *CODE_OUT to the tree code
- OPS[I] to operand I of *CODE_OUT
- *ELSE_OUT to the fallback value ELSE, or to NULL_TREE if the
- condition is known to be all true. */
+ condition is known to be all true.
+ - *LEN to the len argument if it COND_LEN_* operations or to NULL_TREE.
+ - *BIAS to the bias argument if it COND_LEN_* operations or to NULL_TREE. */
bool
can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out,
tree_code *code_out,
- tree (&ops)[3], tree *else_out)
+ tree (&ops)[3], tree *else_out,
+ tree *len, tree *bias)
{
+ *len = NULL_TREE;
+ *bias = NULL_TREE;
if (gassign *assign = dyn_cast <gassign *> (stmt))
{
*cond_out = NULL_TREE;
@@ -4462,18 +4495,28 @@ can_interpret_as_conditional_op_p (gimple *stmt, tree *cond_out,
{
internal_fn ifn = gimple_call_internal_fn (call);
tree_code code = conditional_internal_fn_code (ifn);
+ int len_index = internal_fn_len_index (ifn);
+ int cond_nargs = len_index >= 0 ? 4 : 2;
if (code != ERROR_MARK)
{
*cond_out = gimple_call_arg (call, 0);
*code_out = code;
- unsigned int nops = gimple_call_num_args (call) - 2;
+ unsigned int nops = gimple_call_num_args (call) - cond_nargs;
for (unsigned int i = 0; i < 3; ++i)
ops[i] = i < nops ? gimple_call_arg (call, i + 1) : NULL_TREE;
*else_out = gimple_call_arg (call, nops + 1);
- if (integer_truep (*cond_out))
+ if (len_index < 0)
+ {
+ if (integer_truep (*cond_out))
+ {
+ *cond_out = NULL_TREE;
+ *else_out = NULL_TREE;
+ }
+ }
+ else
{
- *cond_out = NULL_TREE;
- *else_out = NULL_TREE;
+ *len = gimple_call_arg (call, len_index);
+ *bias = gimple_call_arg (call, len_index + 1);
}
return true;
}
@@ -4561,8 +4604,32 @@ internal_fn_len_index (internal_fn fn)
case IFN_LEN_MASK_GATHER_LOAD:
case IFN_LEN_MASK_SCATTER_STORE:
+ case IFN_COND_LEN_FMA:
+ case IFN_COND_LEN_FMS:
+ case IFN_COND_LEN_FNMA:
+ case IFN_COND_LEN_FNMS:
return 5;
+ case IFN_COND_LEN_ADD:
+ case IFN_COND_LEN_SUB:
+ case IFN_COND_LEN_MUL:
+ case IFN_COND_LEN_DIV:
+ case IFN_COND_LEN_MOD:
+ case IFN_COND_LEN_RDIV:
+ case IFN_COND_LEN_MIN:
+ case IFN_COND_LEN_MAX:
+ case IFN_COND_LEN_FMIN:
+ case IFN_COND_LEN_FMAX:
+ case IFN_COND_LEN_AND:
+ case IFN_COND_LEN_IOR:
+ case IFN_COND_LEN_XOR:
+ case IFN_COND_LEN_SHL:
+ case IFN_COND_LEN_SHR:
+ return 4;
+
+ case IFN_COND_LEN_NEG:
+ return 3;
+
default:
return -1;
}
diff --git a/gcc/internal-fn.h b/gcc/internal-fn.h
index dd1bab0bddf..a5c3f4765ff 100644
--- a/gcc/internal-fn.h
+++ b/gcc/internal-fn.h
@@ -229,7 +229,7 @@ extern tree_code conditional_internal_fn_code (internal_fn);
extern internal_fn get_unconditional_internal_fn (internal_fn);
extern bool can_interpret_as_conditional_op_p (gimple *, tree *,
tree_code *, tree (&)[3],
- tree *);
+ tree *, tree *, tree *);
extern bool internal_load_fn_p (internal_fn);
extern bool internal_store_fn_p (internal_fn);
diff --git a/gcc/tree-ssa-math-opts.cc b/gcc/tree-ssa-math-opts.cc
index 68fc518b1ab..712097ac5be 100644
--- a/gcc/tree-ssa-math-opts.cc
+++ b/gcc/tree-ssa-math-opts.cc
@@ -3099,10 +3099,11 @@ convert_mult_to_fma_1 (tree mul_result, tree op1, tree op2)
negate_p = true;
}
- tree cond, else_value, ops[3];
+ tree cond, else_value, ops[3], len, bias;
tree_code code;
if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code,
- ops, &else_value))
+ ops, &else_value,
+ &len, &bias))
gcc_unreachable ();
addop = ops[0] == result ? ops[1] : ops[0];
@@ -3122,7 +3123,11 @@ convert_mult_to_fma_1 (tree mul_result, tree op1, tree op2)
if (seq)
gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
- if (cond)
+ if (len)
+ fma_stmt
+ = gimple_build_call_internal (IFN_COND_LEN_FMA, 7, cond, mulop1, op2,
+ addop, else_value, len, bias);
+ else if (cond)
fma_stmt = gimple_build_call_internal (IFN_COND_FMA, 5, cond, mulop1,
op2, addop, else_value);
else
@@ -3307,7 +3312,8 @@ last_fma_candidate_feeds_initial_phi (fma_deferring_state *state,
static bool
convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2,
- fma_deferring_state *state, tree mul_cond = NULL_TREE)
+ fma_deferring_state *state, tree mul_cond = NULL_TREE,
+ tree mul_len = NULL_TREE, tree mul_bias = NULL_TREE)
{
tree mul_result = gimple_get_lhs (mul_stmt);
/* If there isn't a LHS then this can't be an FMA. There can be no LHS
@@ -3420,10 +3426,10 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2,
negate_p = seen_negate_p = true;
}
- tree cond, else_value, ops[3];
+ tree cond, else_value, ops[3], len, bias;
tree_code code;
if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code, ops,
- &else_value))
+ &else_value, &len, &bias))
return false;
switch (code)
@@ -3439,15 +3445,49 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2,
return false;
}
- if (mul_cond && cond != mul_cond)
- return false;
-
- if (cond)
+ if (len)
{
- if (cond == result || else_value == result)
+ /* For COND_LEN_* operations, we may have dummpy mask which is
+ the all true mask. Such TREE type may be mul_cond != cond
+ but we still consider they are equal. */
+ if (mul_cond && cond != mul_cond
+ && !(integer_truep (mul_cond) && integer_truep (cond)))
return false;
- if (!direct_internal_fn_supported_p (IFN_COND_FMA, type, opt_type))
+
+ if (else_value == result)
+ return false;
+
+ if (!direct_internal_fn_supported_p (IFN_COND_LEN_FMA, type,
+ opt_type))
return false;
+
+ if (mul_len)
+ {
+ poly_int64 mul_value, value;
+ if (poly_int_tree_p (mul_len, &mul_value)
+ && poly_int_tree_p (len, &value)
+ && maybe_ne (mul_value, value))
+ return false;
+ else if (mul_len != len)
+ return false;
+
+ if (wi::to_widest (mul_bias) != wi::to_widest (bias))
+ return false;
+ }
+ }
+ else
+ {
+ if (mul_cond && cond != mul_cond)
+ return false;
+
+ if (cond)
+ {
+ if (cond == result || else_value == result)
+ return false;
+ if (!direct_internal_fn_supported_p (IFN_COND_FMA, type,
+ opt_type))
+ return false;
+ }
}
/* If the subtrahend (OPS[1]) is computed by a MULT_EXPR that
@@ -5632,6 +5672,22 @@ math_opts_dom_walker::after_dom_children (basic_block bb)
}
break;
+ case CFN_COND_LEN_MUL:
+ if (convert_mult_to_fma (stmt,
+ gimple_call_arg (stmt, 1),
+ gimple_call_arg (stmt, 2),
+ &fma_state,
+ gimple_call_arg (stmt, 0),
+ gimple_call_arg (stmt, 4),
+ gimple_call_arg (stmt, 5)))
+
+ {
+ gsi_remove (&gsi, true);
+ release_defs (stmt);
+ continue;
+ }
+ break;
+
case CFN_LAST:
cancel_fma_deferring (&fma_state);
break;
--
2.36.3
next reply other threads:[~2023-07-13 8:54 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-07-13 8:54 juzhe.zhong [this message]
2023-07-13 10:50 ` Richard Biener
2023-07-13 13:48 ` Li, Pan2
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20230713085434.3381643-1-juzhe.zhong@rivai.ai \
--to=juzhe.zhong@rivai.ai \
--cc=gcc-patches@gcc.gnu.org \
--cc=rguenther@suse.de \
--cc=richard.sandiford@arm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).