[APX,NF] Add a pass to convert legacy insn to NF insns

Hi,

For APX ccmp, current infrastructure will always generate cstore for
the ccmp flag user, like

	cmpe    %rcx, %r8
	ccmpnel %rax, %rbx
	seta	%dil
	add	%rcx, %r9
	add     %r9, %rdx
	testb   %dil, %dil
	je	.L2

For such case, the legacy add clobbers FLAGS_REG so there should have
extra cstore to avoid the flag be reset before using it. If the
instructions between flag producer and user are NF insns, the setcc/
test sequence is not required.

Add a pass to convert legacy flag clobber insns to their NF counterpart.
The convertion only happens when
1. APX_NF enabled.
2. For a BB, cstore was find, and there are insns between such cstore
and next explicit set insn to FLAGS_REG (test or cmp).
3. All the insns between should have NF counterpart.

The pass was added after rtl-ifcvt which eliminates some branch when
profitable, which could cause some flag-clobbering insn put between
cstore and jcc.

Bootstrapped & regtested on x86_64-pc-linux-gnu and SDE. Also passed
spec2017 simulation run on SDE.

Ok for trunk?

gcc/ChangeLog:

	* config/i386/i386.md (has_nf): New define_attr, add to all
	nf related patterns.
	* config/i386/i386-features.cc (apx_nf_convert): New function
	to convert Non-NF insns to their NF counterparts.
	(class pass_apx_nf_convert): New pass class.
	(make_pass_apx_nf_convert): New.
	* config/i386/i386-passes.def: Add pass_apx_nf_convert after
	rtl_ifcvt.
	* config/i386/i386-protos.h (make_pass_apx_nf_convert): Declare.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/apx-nf-2.c: New test.
---
 gcc/config/i386/i386-features.cc         | 163 +++++++++++++++++++++++
 gcc/config/i386/i386-passes.def          |   1 +
 gcc/config/i386/i386-protos.h            |   1 +
 gcc/config/i386/i386.md                  |  67 +++++++++-
 gcc/testsuite/gcc.target/i386/apx-nf-2.c |  32 +++++
 5 files changed, 259 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/apx-nf-2.c

On Wed, Jul 10, 2024 at 2:46 PM Hongyu Wang <hongyu.wang@intel.com> wrote:
>
> Hi,
>
> For APX ccmp, current infrastructure will always generate cstore for
> the ccmp flag user, like
>
>         cmpe    %rcx, %r8
>         ccmpnel %rax, %rbx
>         seta    %dil
>         add     %rcx, %r9
>         add     %r9, %rdx
>         testb   %dil, %dil
>         je      .L2
>
> For such case, the legacy add clobbers FLAGS_REG so there should have
> extra cstore to avoid the flag be reset before using it. If the
> instructions between flag producer and user are NF insns, the setcc/
> test sequence is not required.
>
> Add a pass to convert legacy flag clobber insns to their NF counterpart.
> The convertion only happens when
> 1. APX_NF enabled.
> 2. For a BB, cstore was find, and there are insns between such cstore
> and next explicit set insn to FLAGS_REG (test or cmp).
> 3. All the insns between should have NF counterpart.
>
> The pass was added after rtl-ifcvt which eliminates some branch when
> profitable, which could cause some flag-clobbering insn put between
> cstore and jcc.
>
> Bootstrapped & regtested on x86_64-pc-linux-gnu and SDE. Also passed
> spec2017 simulation run on SDE.
>
> Ok for trunk?
Ok.
>
> gcc/ChangeLog:
>
>         * config/i386/i386.md (has_nf): New define_attr, add to all
>         nf related patterns.
>         * config/i386/i386-features.cc (apx_nf_convert): New function
>         to convert Non-NF insns to their NF counterparts.
>         (class pass_apx_nf_convert): New pass class.
>         (make_pass_apx_nf_convert): New.
>         * config/i386/i386-passes.def: Add pass_apx_nf_convert after
>         rtl_ifcvt.
>         * config/i386/i386-protos.h (make_pass_apx_nf_convert): Declare.
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.target/i386/apx-nf-2.c: New test.
> ---
>  gcc/config/i386/i386-features.cc         | 163 +++++++++++++++++++++++
>  gcc/config/i386/i386-passes.def          |   1 +
>  gcc/config/i386/i386-protos.h            |   1 +
>  gcc/config/i386/i386.md                  |  67 +++++++++-
>  gcc/testsuite/gcc.target/i386/apx-nf-2.c |  32 +++++
>  5 files changed, 259 insertions(+), 5 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/apx-nf-2.c
>
> diff --git a/gcc/config/i386/i386-features.cc b/gcc/config/i386/i386-features.cc
> index fc224ed06b0..3da56ddbdcc 100644
> --- a/gcc/config/i386/i386-features.cc
> +++ b/gcc/config/i386/i386-features.cc
> @@ -3259,6 +3259,169 @@ make_pass_remove_partial_avx_dependency (gcc::context *ctxt)
>    return new pass_remove_partial_avx_dependency (ctxt);
>  }
>
> +/* Convert legacy instructions that clobbers EFLAGS to APX_NF
> +   instructions when there are no flag set between a flag
> +   producer and user.  */
> +
> +static unsigned int
> +ix86_apx_nf_convert (void)
> +{
> +  timevar_push (TV_MACH_DEP);
> +
> +  basic_block bb;
> +  rtx_insn *insn;
> +  hash_map <rtx_insn *, rtx> converting_map;
> +  auto_vec <rtx_insn *> current_convert_list;
> +
> +  bool converting_seq = false;
> +  rtx cc = gen_rtx_REG (CCmode, FLAGS_REG);
> +
> +  FOR_EACH_BB_FN (bb, cfun)
> +    {
> +      /* Reset conversion for each bb.  */
> +      converting_seq = false;
> +      FOR_BB_INSNS (bb, insn)
> +       {
> +         if (!NONDEBUG_INSN_P (insn))
> +           continue;
> +
> +         if (recog_memoized (insn) < 0)
> +           continue;
> +
> +         /* Convert candidate insns after cstore, which should
> +            satisify the two conditions:
> +            1. Is not flag user or producer, only clobbers
> +            FLAGS_REG.
> +            2. Have corresponding nf pattern.  */
> +
> +         rtx pat = PATTERN (insn);
> +
> +         /* Starting convertion at first cstorecc.  */
> +         rtx set = NULL_RTX;
> +         if (!converting_seq
> +             && (set = single_set (insn))
> +             && ix86_comparison_operator (SET_SRC (set), VOIDmode)
> +             && reg_overlap_mentioned_p (cc, SET_SRC (set))
> +             && !reg_overlap_mentioned_p (cc, SET_DEST (set)))
> +           {
> +             converting_seq = true;
> +             current_convert_list.truncate (0);
> +           }
> +         /* Terminate at the next explicit flag set.  */
> +         else if (reg_set_p (cc, pat)
> +                  && GET_CODE (set_of (cc, pat)) != CLOBBER)
> +           converting_seq = false;
> +
> +         if (!converting_seq)
> +           continue;
> +
> +         if (get_attr_has_nf (insn)
> +             && GET_CODE (pat) == PARALLEL)
> +           {
> +             /* Record the insn to candidate map.  */
> +             current_convert_list.safe_push (insn);
> +             converting_map.put (insn, pat);
> +           }
> +         /* If the insn clobbers flags but has no nf_attr,
> +            revoke all previous candidates.  */
> +         else if (!get_attr_has_nf (insn)
> +                  && reg_set_p (cc, pat)
> +                  && GET_CODE (set_of (cc, pat)) == CLOBBER)
> +           {
> +             for (auto item : current_convert_list)
> +               converting_map.remove (item);
> +             converting_seq = false;
> +           }
> +       }
> +    }
> +
> +  if (!converting_map.is_empty ())
> +    {
> +      for (auto iter = converting_map.begin ();
> +          iter != converting_map.end (); ++iter)
> +       {
> +         rtx_insn *replace = (*iter).first;
> +         rtx pat = (*iter).second;
> +         int i, n = 0, len = XVECLEN (pat, 0);
> +         rtx *new_elems = XALLOCAVEC (rtx, len);
> +         rtx new_pat;
> +         for (i = 0; i < len; i++)
> +           {
> +             rtx temp = XVECEXP (pat, 0, i);
> +             if (! (GET_CODE (temp) == CLOBBER
> +                    && reg_overlap_mentioned_p (cc,
> +                                                XEXP (temp, 0))))
> +               {
> +                 new_elems[n] = temp;
> +                 n++;
> +               }
> +           }
> +
> +         if (n == 1)
> +           new_pat = new_elems[0];
> +         else
> +           new_pat =
> +             gen_rtx_PARALLEL (VOIDmode,
> +                               gen_rtvec_v (n,
> +                                            new_elems));
> +
> +         PATTERN (replace) = new_pat;
> +         INSN_CODE (replace) = -1;
> +         recog_memoized (replace);
> +         df_insn_rescan (replace);
> +       }
> +    }
> +
> +  timevar_pop (TV_MACH_DEP);
> +  return 0;
> +}
> +
> +
> +namespace {
> +
> +const pass_data pass_data_apx_nf_convert =
> +{
> +  RTL_PASS, /* type */
> +  "apx_nfcvt", /* name */
> +  OPTGROUP_NONE, /* optinfo_flags */
> +  TV_MACH_DEP, /* tv_id */
> +  0, /* properties_required */
> +  0, /* properties_provided */
> +  0, /* properties_destroyed */
> +  0, /* todo_flags_start */
> +  0, /* todo_flags_finish */
> +};
> +
> +class pass_apx_nf_convert : public rtl_opt_pass
> +{
> +public:
> +  pass_apx_nf_convert (gcc::context *ctxt)
> +    : rtl_opt_pass (pass_data_apx_nf_convert, ctxt)
> +  {}
> +
> +  /* opt_pass methods: */
> +  bool gate (function *) final override
> +    {
> +      return (TARGET_APX_NF
> +             && optimize
> +             && optimize_function_for_speed_p (cfun));
> +    }
> +
> +  unsigned int execute (function *) final override
> +    {
> +      return ix86_apx_nf_convert ();
> +    }
> +}; // class pass_rpad
> +
> +} // anon namespace
> +
> +rtl_opt_pass *
> +make_pass_apx_nf_convert (gcc::context *ctxt)
> +{
> +  return new pass_apx_nf_convert (ctxt);
> +}
> +
> +
>  /* This compares the priority of target features in function DECL1
>     and DECL2.  It returns positive value if DECL1 is higher priority,
>     negative value if DECL2 is higher priority and 0 if they are the
> diff --git a/gcc/config/i386/i386-passes.def b/gcc/config/i386/i386-passes.def
> index 2d29f65da88..99fc8805b22 100644
> --- a/gcc/config/i386/i386-passes.def
> +++ b/gcc/config/i386/i386-passes.def
> @@ -33,3 +33,4 @@ along with GCC; see the file COPYING3.  If not see
>    INSERT_PASS_BEFORE (pass_shorten_branches, 1, pass_insert_endbr_and_patchable_area);
>
>    INSERT_PASS_AFTER (pass_late_combine, 1, pass_remove_partial_avx_dependency);
> +  INSERT_PASS_AFTER (pass_rtl_ifcvt, 1, pass_apx_nf_convert);
> diff --git a/gcc/config/i386/i386-protos.h b/gcc/config/i386/i386-protos.h
> index 68f57393c5d..a3629b32a01 100644
> --- a/gcc/config/i386/i386-protos.h
> +++ b/gcc/config/i386/i386-protos.h
> @@ -423,6 +423,7 @@ extern rtl_opt_pass *make_pass_insert_endbr_and_patchable_area
>    (gcc::context *);
>  extern rtl_opt_pass *make_pass_remove_partial_avx_dependency
>    (gcc::context *);
> +extern rtl_opt_pass *make_pass_apx_nf_convert (gcc::context *);
>
>  extern bool ix86_has_no_direct_extern_access;
>  extern bool ix86_rpad_gate ();
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index 214cb2e239a..0d4ee514e40 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -992,6 +992,9 @@ (define_attr "enabled" ""
>  (define_attr "preferred_for_size" "" (const_int 1))
>  (define_attr "preferred_for_speed" "" (const_int 1))
>
> +;; Define attribute to mark the insn has nf variant.
> +(define_attr "has_nf" "0,1" (const_string "0"))
> +
>  ;; Describe a user's asm statement.
>  (define_asm_attributes
>    [(set_attr "length" "128")
> @@ -6565,6 +6568,7 @@ (define_insn "*add<mode>_1<nf_name>"
>         (and (eq_attr "type" "alu") (match_operand 2 "const128_operand"))
>         (const_string "1")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; It may seem that nonimmediate operand is proper one for operand 1.
> @@ -6682,6 +6686,7 @@ (define_insn "*addhi_1<nf_name>"
>         (and (eq_attr "type" "alu") (match_operand 2 "const128_operand"))
>         (const_string "1")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "HI,HI,HI,SI,HI,HI")])
>
>  (define_insn "*addqi_1<nf_name>"
> @@ -6750,6 +6755,7 @@ (define_insn "*addqi_1<nf_name>"
>         (and (eq_attr "type" "alu") (match_operand 2 "const128_operand"))
>         (const_string "1")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "QI,QI,QI,SI,SI,SI,QI,QI")
>     ;; Potential partial reg stall on alternatives 3 and 4.
>     (set (attr "preferred_for_speed")
> @@ -7963,6 +7969,7 @@ (define_insn "*sub<mode>_1<nf_name>"
>    <nf_prefix>sub{<imodesuffix>}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "isa" "*,*,*,apx_ndd,apx_ndd,apx_ndd")
>     (set_attr "type" "alu")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*subsi_1_zext"
> @@ -10078,6 +10085,7 @@ (define_insn "*mul<mode>3_1<nf_name>"
>           (match_test "<MODE>mode == HImode")
>             (const_string "double")
>             (const_string "direct")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*imulhi<mode>zu<nf_name>"
> @@ -10145,6 +10153,7 @@ (define_insn "*mulqi3_1<nf_name>"
>          (const_string "direct")))
>     (set_attr "amdfam10_decode" "direct")
>     (set_attr "bdver1_decode" "direct")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "QI")])
>
>  ;; Multiply with jump on overflow.
> @@ -11309,6 +11318,7 @@ (define_insn "*<u>divmod<mode>4_noext"
>    ""
>    "<sgnprefix>div{<imodesuffix>}\t%3"
>    [(set_attr "type" "idiv")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*<u>divmodsi4_noext_zext_1"
> @@ -11461,6 +11471,7 @@ (define_insn "<u>divmodhiqi3<nf_name>"
>     && <nf_condition>"
>    "<nf_prefix><sgnprefix>div{b}\t%2"
>    [(set_attr "type" "idiv")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "QI")])
>
>  ;; We cannot use div/idiv for double division, because it causes
> @@ -12006,6 +12017,7 @@ (define_insn "*anddi_1<nf_name>"
>                  (match_operand 1 "ext_QIreg_operand")))
>         (const_string "1")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI,SI,DI,DI,DI,DI,DI,DI,SI,DI")])
>
>  (define_insn_and_split "*anddi_1_btr"
> @@ -12112,6 +12124,7 @@ (define_insn "*and<mode>_1<nf_name>"
>                  (match_operand 1 "ext_QIreg_operand")))
>         (const_string "1")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>,<MODE>,<MODE>,<MODE>,<MODE>,<MODE>,SI,<MODE>")])
>
>  (define_insn "*andqi_1<nf_name>"
> @@ -12129,6 +12142,7 @@ (define_insn "*andqi_1<nf_name>"
>     #"
>    [(set_attr "type" "alu,alu,alu,alu,alu,msklog")
>     (set_attr "isa" "*,*,*,apx_ndd,apx_ndd,*")
> +   (set_attr "has_nf" "1")
>     (set (attr "mode")
>         (cond [(eq_attr "alternative" "2")
>                  (const_string "SI")
> @@ -13005,6 +13019,7 @@ (define_insn "*<code><mode>_1<nf_name>"
>     #"
>    [(set_attr "isa" "*,*,*,apx_ndd,apx_ndd,apx_ndd,<kmov_isa>")
>     (set_attr "type" "alu,alu, alu, alu, alu, alu, msklog")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn_and_split "*notxor<mode>_1"
> @@ -13165,6 +13180,7 @@ (define_insn "*<code>qi_1<nf_name>"
>     #"
>    [(set_attr "isa" "*,*,*,apx_ndd,apx_ndd,avx512f")
>     (set_attr "type" "alu,alu,alu,alu,alu,msklog")
> +   (set_attr "has_nf" "1")
>     (set (attr "mode")
>         (cond [(eq_attr "alternative" "2")
>                  (const_string "SI")
> @@ -13731,6 +13747,7 @@ (define_insn "*neg<mode>_1<nf_name>"
>    <nf_prefix>neg{<imodesuffix>}\t{%1, %0|%0, %1}"
>    [(set_attr "type" "negnot")
>     (set_attr "isa" "*,apx_ndd")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*negsi_1_zext"
> @@ -14744,6 +14761,7 @@ (define_insn "x86_64_shld<nf_name>"
>    "<nf_prefix>shld{q}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")
>     (set_attr "athlon_decode" "vector")
>     (set_attr "amdfam10_decode" "vector")
> @@ -14763,6 +14781,7 @@ (define_insn "x86_64_shld_ndd<nf_name>"
>    "TARGET_APX_NDD && <nf_condition>"
>    "<nf_prefix>shld{q}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")])
>
>  (define_insn "x86_64_shld_1<nf_name>"
> @@ -14780,6 +14799,7 @@ (define_insn "x86_64_shld_1<nf_name>"
>    "<nf_prefix>shld{q}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")
>     (set_attr "length_immediate" "1")
>     (set_attr "athlon_decode" "vector")
> @@ -14800,6 +14820,7 @@ (define_insn "x86_64_shld_ndd_1<nf_name>"
>     && <nf_condition>"
>    "<nf_prefix>shld{q}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")
>     (set_attr "length_immediate" "1")])
>
> @@ -14909,7 +14930,8 @@ (define_insn_and_split "*x86_64_shld_shrd_1_nozext"
>       emit_move_insn (operands[0], tmp);
>     }
>     DONE;
> -})
> +}
> +  [(set_attr "has_nf" "1")])
>
>  (define_insn_and_split "*x86_64_shld_2"
>    [(set (match_operand:DI 0 "nonimmediate_operand")
> @@ -14974,6 +14996,7 @@ (define_insn "x86_shld<nf_name>"
>    "<nf_prefix>shld{l}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")
>     (set_attr "pent_pair" "np")
>     (set_attr "athlon_decode" "vector")
> @@ -14994,6 +15017,7 @@ (define_insn "x86_shld_ndd<nf_name>"
>    "TARGET_APX_NDD && <nf_condition>"
>    "<nf_prefix>shld{l}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")])
>
>
> @@ -15012,6 +15036,7 @@ (define_insn "x86_shld_1<nf_name>"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")
>     (set_attr "pent_pair" "np")
>     (set_attr "athlon_decode" "vector")
> @@ -15033,6 +15058,7 @@ (define_insn "x86_shld_ndd_1<nf_name>"
>    "<nf_prefix>shld{l}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")])
>
>  (define_insn_and_split "*x86_shld_shrd_1_nozext_nf"
> @@ -15140,7 +15166,8 @@ (define_insn_and_split "*x86_shld_shrd_1_nozext"
>       emit_move_insn (operands[0], tmp);
>     }
>     DONE;
> -})
> +}
> +  [(set_attr "has_nf" "1")])
>
>  (define_insn_and_split "*x86_shld_2"
>    [(set (match_operand:SI 0 "nonimmediate_operand")
> @@ -15356,6 +15383,7 @@ (define_insn "*ashl<mode>3_1<nf_name>"
>                            (match_test "optimize_function_for_size_p (cfun)")))))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; Convert shift to the shiftx pattern to avoid flags dependency.
> @@ -15512,6 +15540,7 @@ (define_insn "*ashlhi3_1<nf_name>"
>                            (match_test "optimize_function_for_size_p (cfun)")))))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "HI,SI,HI,HI")])
>
>  (define_insn "*ashlqi3_1<nf_name>"
> @@ -15583,6 +15612,7 @@ (define_insn "*ashlqi3_1<nf_name>"
>                            (match_test "optimize_function_for_size_p (cfun)")))))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "QI,SI,SI,QI,QI")
>     ;; Potential partial reg stall on alternative 1.
>     (set (attr "preferred_for_speed")
> @@ -16184,7 +16214,8 @@ (define_insn_and_split "<insn><dwi>3_doubleword_lowpart"
>    operands[4] = GEN_INT ((<MODE_SIZE> * BITS_PER_UNIT) - INTVAL (operands[2]));
>    if (!rtx_equal_p (operands[0], operands[1]))
>      emit_move_insn (operands[0], operands[1]);
> -})
> +}
> +  [(set_attr "has_nf" "1")])
>
>  (define_insn "x86_64_shrd<nf_name>"
>    [(set (match_operand:DI 0 "nonimmediate_operand" "+r*m")
> @@ -16201,6 +16232,7 @@ (define_insn "x86_64_shrd<nf_name>"
>    "<nf_prefix>shrd{q}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")
>     (set_attr "athlon_decode" "vector")
>     (set_attr "amdfam10_decode" "vector")
> @@ -16220,6 +16252,7 @@ (define_insn "x86_64_shrd_ndd<nf_name>"
>    "TARGET_APX_NDD && <nf_condition>"
>    "<nf_prefix>shrd{q}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")])
>
>  (define_insn "x86_64_shrd_1<nf_name>"
> @@ -16238,6 +16271,7 @@ (define_insn "x86_64_shrd_1<nf_name>"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")
>     (set_attr "athlon_decode" "vector")
>     (set_attr "amdfam10_decode" "vector")
> @@ -16258,6 +16292,7 @@ (define_insn "x86_64_shrd_ndd_1<nf_name>"
>    "<nf_prefix>shrd{q}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "DI")])
>
>  (define_insn_and_split "*x86_64_shrd_shld_1_nozext_nf"
> @@ -16366,7 +16401,8 @@ (define_insn_and_split "*x86_64_shrd_shld_1_nozext"
>       emit_move_insn (operands[0], tmp);
>     }
>     DONE;
> -})
> +}
> +  [(set_attr "has_nf" "1")])
>
>  (define_insn_and_split "*x86_64_shrd_2"
>    [(set (match_operand:DI 0 "nonimmediate_operand")
> @@ -16431,6 +16467,7 @@ (define_insn "x86_shrd<nf_name>"
>    "<nf_prefix>shrd{l}\t{%2, %1, %0|%0, %1, %2}"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")
>     (set_attr "pent_pair" "np")
>     (set_attr "athlon_decode" "vector")
> @@ -16451,6 +16488,7 @@ (define_insn "x86_shrd_ndd<nf_name>"
>    "TARGET_APX_NDD && <nf_condition>"
>    "<nf_prefix>shrd{l}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")])
>
>  (define_insn "x86_shrd_1<nf_name>"
> @@ -16468,6 +16506,7 @@ (define_insn "x86_shrd_1<nf_name>"
>    [(set_attr "type" "ishift")
>     (set_attr "prefix_0f" "1")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")
>     (set_attr "pent_pair" "np")
>     (set_attr "athlon_decode" "vector")
> @@ -16489,6 +16528,7 @@ (define_insn "x86_shrd_ndd_1<nf_name>"
>    "<nf_prefix>shrd{l}\t{%3, %2, %1, %0|%0, %1, %2, %3}"
>    [(set_attr "type" "ishift")
>     (set_attr "length_immediate" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "SI")])
>
>  (define_insn_and_split "*x86_shrd_shld_1_nozext_nf"
> @@ -16596,7 +16636,8 @@ (define_insn_and_split "*x86_shrd_shld_1_nozext"
>       emit_move_insn (operands[0], tmp);
>     }
>     DONE;
> -})
> +}
> +  [(set_attr "has_nf" "1")])
>
>  (define_insn_and_split "*x86_shrd_2"
>    [(set (match_operand:SI 0 "nonimmediate_operand")
> @@ -16668,6 +16709,7 @@ (define_insn "ashr<mode>3_cvt<nf_name>"
>     (set_attr "prefix_0f" "0,*,*")
>     (set_attr "length_immediate" "0,*,*")
>     (set_attr "modrm" "0,1,1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*ashrsi3_cvt_zext"
> @@ -16761,6 +16803,7 @@ (define_insn "*ashr<mode>3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; Specialization of *lshr<mode>3_1 below, extracting the SImode
> @@ -16824,6 +16867,7 @@ (define_insn "*lshr<mode>3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; Convert shift to the shiftx pattern to avoid flags dependency.
> @@ -16939,6 +16983,7 @@ (define_insn "*ashr<mode>3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "*lshrqi3_1<nf_name>"
> @@ -16976,6 +17021,7 @@ (define_insn "*lshrqi3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "QI")])
>
>  (define_insn "*lshrhi3_1<nf_name>"
> @@ -17013,6 +17059,7 @@ (define_insn "*lshrhi3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "HI")])
>
>  ;; Alternative 1 is needed to work around LRA limitation, see PR82524.
> @@ -17562,6 +17609,7 @@ (define_insn "*<insn><mode>3_1<nf_name>"
>                       (match_test "optimize_function_for_size_p (cfun)"))))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; Convert rotate to the rotatex pattern to avoid flags dependency.
> @@ -17706,6 +17754,7 @@ (define_insn "*<insn><mode>3_1<nf_name>"
>                  (match_test "optimize_function_for_size_p (cfun)")))
>         (const_string "0")
>         (const_string "*")))
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ;; Alternative 1 is needed to work around LRA limitation, see PR82524.
> @@ -20497,6 +20546,7 @@ (define_insn_and_split "clz<mode>2_lzcnt"
>    "ix86_expand_clear (operands[0]);"
>    [(set_attr "prefix_rep" "1")
>     (set_attr "type" "bitmanip")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ; False dependency happens when destination is only updated by tzcnt,
> @@ -20525,6 +20575,7 @@ (define_insn "*clz<mode>2_lzcnt_falsedep"
>    "lzcnt{<imodesuffix>}\t{%1, %0|%0, %1}"
>    [(set_attr "prefix_rep" "1")
>     (set_attr "type" "bitmanip")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn_and_split "*clzsi2_lzcnt_zext"
> @@ -20658,6 +20709,7 @@ (define_insn_and_split "<lt_zcnt>_<mode>"
>    [(set_attr "type" "<lt_zcnt_type>")
>     (set_attr "prefix_0f" "1")
>     (set_attr "prefix_rep" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ; False dependency happens when destination is only updated by tzcnt,
> @@ -20688,6 +20740,7 @@ (define_insn "*<lt_zcnt>_<mode>_falsedep"
>    [(set_attr "type" "<lt_zcnt_type>")
>     (set_attr "prefix_0f" "1")
>     (set_attr "prefix_rep" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn "<lt_zcnt>_hi<nf_name>"
> @@ -20699,6 +20752,7 @@ (define_insn "<lt_zcnt>_hi<nf_name>"
>    [(set_attr "type" "<lt_zcnt_type>")
>     (set_attr "prefix_0f" "1")
>     (set_attr "prefix_rep" "1")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "HI")])
>
>  ;; BMI instructions.
> @@ -21161,6 +21215,7 @@ (define_insn_and_split "popcount<mode>2"
>    "ix86_expand_clear (operands[0]);"
>    [(set_attr "prefix_rep" "1")
>     (set_attr "type" "bitmanip")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  ; False dependency happens when destination is only updated by tzcnt,
> @@ -21201,6 +21256,7 @@ (define_insn "*popcount<mode>2_falsedep"
>  }
>    [(set_attr "prefix_rep" "1")
>     (set_attr "type" "bitmanip")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "<MODE>")])
>
>  (define_insn_and_split "*popcountsi2_zext"
> @@ -21355,6 +21411,7 @@ (define_insn "popcounthi2<nf_name>"
>  }
>    [(set_attr "prefix_rep" "1")
>     (set_attr "type" "bitmanip")
> +   (set_attr "has_nf" "1")
>     (set_attr "mode" "HI")])
>
>  (define_expand "bswapdi2"
> diff --git a/gcc/testsuite/gcc.target/i386/apx-nf-2.c b/gcc/testsuite/gcc.target/i386/apx-nf-2.c
> new file mode 100644
> index 00000000000..8f5b8f6b2b7
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/apx-nf-2.c
> @@ -0,0 +1,32 @@
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-O3 -mapxf" } */
> +
> +int foo(int a, int b, int c, int d)
> +{
> +  int sum = a;
> +
> +  if (a != c)
> +    {
> +      c += d;
> +      a += b;
> +      sum += a + c;
> +      if (b != d && sum < c || sum > d)
> +       {
> +         b -= d;
> +         sum -= b;
> +       }
> +    }
> +
> +  return sum;
> +}
> +
> +int foo2 (unsigned a, unsigned b, unsigned d, unsigned e, int *p)
> +{
> +  unsigned r;
> +  int c = __builtin_mul_overflow (a, b, &r);
> +  *p += a;
> +  return c ? d : e;
> +}
> +
> +/* { dg-final { scan-assembler-not "set" } } */
> +
> --
> 2.31.1
>