Hi,
Gently ping for this.
This version has removed the target hook and added a new optab for cfcmov.
Thanks,
Lingling
From: Kong, Lingling <lingling.k...@intel.com>
Sent: Tuesday, June 18, 2024 3:41 PM
To: gcc-patches@gcc.gnu.org
Cc: Alexander Monakov <amona...@ispras.ru>; Uros Bizjak <ubiz...@gmail.com>;
lingling.ko...@gmail.com; Hongtao Liu <crazy...@gmail.com>; Jeff Law
<jeffreya...@gmail.com>; Richard Biener <richard.guent...@gmail.com>
Subject: [PATCH v2 1/2] [APX CFCMOV] Support APX CFCMOV in if_convert pass
APX CFCMOV feature implements conditionally faulting which means
that all memory faults are suppressed when the condition code
evaluates to false and load or store a memory operand. Now we
could load or store a memory operand may trap or fault for
conditional move.
In middle-end, now we don't support a conditional move if we knew
that a load from A or B could trap or fault. To enable CFCMOV, we
added a new optab.
Conditional move suppress_fault for condition mem store would not
move any arithmetic calculations. For condition mem load now just
support a conditional move one trap mem and one no trap and no mem
cases.
gcc/ChangeLog:
* ifcvt.cc (noce_try_cmove_load_mem_notrap): Allow convert
to cfcmov for conditional load.
(noce_try_cmove_store_mem_notrap): Convert to conditional store.
(noce_process_if_block): Ditto.
* optabs.def (OPTAB_D): New optab.
---
gcc/ifcvt.cc | 246 ++++++++++++++++++++++++++++++++++++++++++++++++-
gcc/optabs.def | 1 +
2 files changed, 246 insertions(+), 1 deletion(-)
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index 58ed42673e5..65c069b8cc6 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -783,6 +783,8 @@ static rtx noce_emit_cmove (struct noce_if_info *, rtx,
enum rtx_code, rtx,
rtx, rtx, rtx, rtx = NULL, rtx
= NULL);
static bool noce_try_cmove (struct noce_if_info *);
static bool noce_try_cmove_arith (struct noce_if_info *);
+static bool noce_try_cmove_load_mem_notrap (struct noce_if_info *);
+static bool noce_try_cmove_store_mem_notrap (struct noce_if_info *, rtx *,
rtx);
static rtx noce_get_alt_condition (struct noce_if_info *, rtx, rtx_insn **);
static bool noce_try_minmax (struct noce_if_info *);
static bool noce_try_abs (struct noce_if_info *);
@@ -2401,6 +2403,233 @@ noce_try_cmove_arith (struct noce_if_info *if_info)
return false;
}
+/* When target support suppress memory fault, try more complex cases involving
+ conditional_move's source or dest may trap or fault. */
+
+static bool
+noce_try_cmove_load_mem_notrap (struct noce_if_info *if_info)
+{
+ rtx a = if_info->a;
+ rtx b = if_info->b;
+ rtx x = if_info->x;
+
+ if (MEM_P (x))
+ return false;
+ /* Just handle a conditional move from one trap MEM + other non_trap,
+ non mem cases. */
+ if (!(MEM_P (a) ^ MEM_P (b)))
+ return false;
+ bool a_trap = may_trap_or_fault_p (a);
+ bool b_trap = may_trap_or_fault_p (b);
+
+ if (!(a_trap ^ b_trap))
+ return false;
+ if (a_trap && !MEM_P (a))
+ return false;
+ if (b_trap && !MEM_P (b))
+ return false;
+
+ rtx orig_b;
+ rtx_insn *insn_a, *insn_b;
+ bool a_simple = if_info->then_simple;
+ bool b_simple = if_info->else_simple;
+ basic_block then_bb = if_info->then_bb;
+ basic_block else_bb = if_info->else_bb;
+ rtx target;
+ enum rtx_code code;
+ rtx cond = if_info->cond;
+ rtx_insn *ifcvt_seq;
+
+ /* if (test) x = *a; else x = c - d;
+ => x = c - d;
+ if (test)
+ x = *a;
+ */
+
+ code = GET_CODE (cond);
+ insn_a = if_info->insn_a;
+ insn_b = if_info->insn_b;
+ machine_mode x_mode = GET_MODE (x);
+
+ /* Because we only handle one trap MEM + other non_trap, non mem cases,
+ just move one trap MEM always in then_bb. */
+ if (noce_reversed_cond_code (if_info) != UNKNOWN)
+ {
+ bool reversep = false;
+ if (b_trap)
+ reversep = true;
+
+ if (reversep)
+ {
+ if (if_info->rev_cond)
+ {
+ cond = if_info->rev_cond;
+ code = GET_CODE (cond);
+ }
+ else
+ code = reversed_comparison_code (cond, if_info->jump);
+ std::swap (a, b);
+ std::swap (insn_a, insn_b);
+ std::swap (a_simple, b_simple);
+ std::swap (then_bb, else_bb);
+ }
+ }
+
+ if (then_bb && else_bb
+ && (!bbs_ok_for_cmove_arith (then_bb, else_bb, if_info->orig_x)
+ || !bbs_ok_for_cmove_arith (else_bb, then_bb,
if_info->orig_x)))
+ return false;
+
+ start_sequence ();
+
+ /* If one of the blocks is empty then the corresponding B or A value
+ came from the test block. The non-empty complex block that we will
+ emit might clobber the register used by B or A, so move it to a pseudo
+ first. */
+
+ rtx tmp_b = NULL_RTX;
+
+ /* Don't move trap mem to a pseudo. */
+ if (!may_trap_or_fault_p (b) && (b_simple || !else_bb))
+ tmp_b = gen_reg_rtx (x_mode);
+
+ orig_b = b;
+
+ rtx emit_a = NULL_RTX;
+ rtx emit_b = NULL_RTX;
+ rtx_insn *tmp_insn = NULL;
+ bool modified_in_a = false;
+ bool modified_in_b = false;
+ /* If either operand is complex, load it into a register first.
+ The best way to do this is to copy the original insn. In this
+ way we preserve any clobbers etc that the insn may have had.
+ This is of course not possible in the IS_MEM case. */
+
+ if (! general_operand (b, GET_MODE (b)) || tmp_b)
+ {
+ if (insn_b)
+ {
+ b = tmp_b ? tmp_b : gen_reg_rtx (GET_MODE (b));
+ rtx_insn *copy_of_b = as_a <rtx_insn *> (copy_rtx (insn_b));
+ rtx set = single_set (copy_of_b);
+
+ SET_DEST (set) = b;
+ emit_b = PATTERN (copy_of_b);
+ }
+ else
+ {
+ rtx tmp_reg = tmp_b ? tmp_b : gen_reg_rtx (GET_MODE (b));
+ emit_b = gen_rtx_SET (tmp_reg, b);
+ b = tmp_reg;
+ }
+ }
+
+ if (tmp_b && then_bb)
+ {
+ FOR_BB_INSNS (then_bb, tmp_insn)
+ /* Don't check inside insn_a. We will have changed it to emit_a
+ with a destination that doesn't conflict. */
+ if (!(insn_a && tmp_insn == insn_a)
+ && modified_in_p (orig_b, tmp_insn))
+ {
+ modified_in_a = true;
+ break;
+ }
+
+ }
+
+ modified_in_b = emit_b != NULL_RTX && modified_in_p (a, emit_b);
+ /* If insn to set up A clobbers any registers B depends on, try to
+ swap insn that sets up A with the one that sets up B. If even
+ that doesn't help, punt. */
+ if (modified_in_a && !modified_in_b)
+ {
+ if (!noce_emit_bb (emit_b, else_bb, b_simple))
+ goto end_seq_and_fail;
+
+ if (!noce_emit_bb (emit_a, then_bb, a_simple))
+ goto end_seq_and_fail;
+ }
+ else if (!modified_in_a)
+ {
+ if (!noce_emit_bb (emit_b, else_bb, b_simple))
+ goto end_seq_and_fail;
+
+ if (!noce_emit_bb (emit_a, then_bb, a_simple))
+ goto end_seq_and_fail;
+ }
+ else
+ goto end_seq_and_fail;
+
+ target = noce_emit_cmove (if_info, x, code, XEXP (cond, 0), XEXP (cond, 1),
+ a, b);
+
+ if (! target)
+ goto end_seq_and_fail;
+
+ if (target != x)
+ noce_emit_move_insn (x, target);
+
+ ifcvt_seq = end_ifcvt_sequence (if_info);
+ if (!ifcvt_seq || !targetm.noce_conversion_profitable_p (ifcvt_seq, if_info))
+ return false;
+
+ emit_insn_before_setloc (ifcvt_seq, if_info->jump,
+ INSN_LOCATION (if_info->insn_a));
+ if_info->transform_name = "noce_try_cmove_load_mem_notrap";
+ return true;
+
+ end_seq_and_fail:
+ end_sequence ();
+ return false;
+}
+
+static bool
+noce_try_cmove_store_mem_notrap (struct noce_if_info *if_info, rtx *x_ptr, rtx
orig_x)
+{
+ rtx a = if_info->a;
+ rtx b = if_info->b;
+ rtx x = orig_x;
+ machine_mode x_mode = GET_MODE (x);
+
+ if (!MEM_P (x) || !rtx_equal_p (x, b))
+ return false;
+ if (!may_trap_or_fault_p (x))
+ return false;
+ if (!if_info->then_simple || !register_operand (a, x_mode))
+ return false;
+
+ rtx cond = if_info->cond;
+ enum rtx_code code = GET_CODE (cond);
+ rtx_insn *ifcvt_seq;
+
+ start_sequence ();
+
+ rtx target = noce_emit_cmove (if_info, x, code, XEXP (cond, 0), XEXP (cond,
1),
+ a, b);
+
+ if (! target)
+ goto end_seq_and_fail;
+
+ if (target != x)
+ noce_emit_move_insn (x, target);
+
+ ifcvt_seq = end_ifcvt_sequence (if_info);
+ if (!ifcvt_seq || !targetm.noce_conversion_profitable_p (ifcvt_seq, if_info))
+ return false;
+
+ emit_insn_before_setloc (ifcvt_seq, if_info->jump,
+ INSN_LOCATION (if_info->insn_a));
+ if_info->transform_name = "noce_try_cmove_load_mem_notrap";
+ if_info->x = orig_x;
+ *x_ptr = orig_x;
+ return true;
+
+ end_seq_and_fail:
+ end_sequence ();
+ return false;
+}
+
/* For most cases, the simplified condition we found is the best
choice, but this is not the case for the min/max/abs transforms.
For these we wish to know that it is A or B in the condition. */
@@ -4055,6 +4284,8 @@ noce_process_if_block (struct noce_if_info *if_info)
the lifetime of hard registers on small register class machines. */
orig_x = x;
if_info->orig_x = orig_x;
+ bool have_cfmovcc = (optab_handler (cfmovcc_optab, GET_MODE (orig_x))
+ != CODE_FOR_nothing);
if (!REG_P (x)
|| (HARD_REGISTER_P (x)
&& targetm.small_register_classes_for_mode_p (GET_MODE (x))))
@@ -4121,12 +4352,21 @@ noce_process_if_block (struct noce_if_info *if_info)
}
if (!set_b && MEM_P (orig_x))
+ {
+ /* Conditional_move_suppress_fault for condition mem store would not
+ move any arithmetic calculations. */
+ if (have_cfmovcc
+ && HAVE_conditional_move
+ && noce_try_cmove_store_mem_notrap (if_info, &x, orig_x))
+ goto success;
+ else
/* We want to avoid store speculation to avoid cases like
if (pthread_mutex_trylock(mutex))
++global_variable;
Rather than go to much effort here, we rely on the SSA optimizers,
which do a good enough job these days. */
- return false;
+ return false;
+ }
if (noce_try_move (if_info))
goto success;
@@ -4160,6 +4400,10 @@ noce_process_if_block (struct noce_if_info *if_info)
if (HAVE_conditional_move
&& noce_try_cmove_arith (if_info))
goto success;
+ if (HAVE_conditional_move
+ && have_cfmovcc
+ && noce_try_cmove_load_mem_notrap (if_info))
+ goto success;
if (noce_try_sign_mask (if_info))
goto success;
}
diff --git a/gcc/optabs.def b/gcc/optabs.def
index bc2611abdc2..49335ec3212 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -540,3 +540,4 @@ OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
OPTAB_D (len_load_optab, "len_load_$a")
OPTAB_D (len_store_optab, "len_store_$a")
OPTAB_D (select_vl_optab, "select_vl$a")
+OPTAB_D (cfmovcc_optab, "cfmov$acc")
--
2.31.1
