From: "Kewen.Lin" <linkw@linux.ibm.com>
To: Jiufu Guo <guojiufu@linux.ibm.com>
Cc: segher@kernel.crashing.org, dje.gcc@gmail.com, linkw@gcc.gnu.org,
bergner@linux.ibm.com, gcc-patches@gcc.gnu.org
Subject: Re: [PATCH V2 1/3]rs6000: update num_insns_constant for 2 insns
Date: Wed, 22 Nov 2023 17:12:29 +0800 [thread overview]
Message-ID: <82d4c257-fd86-6879-62f4-814e40ee3f58@linux.ibm.com> (raw)
In-Reply-To: <20231115030237.1188073-1-guojiufu@linux.ibm.com>
Hi,
on 2023/11/15 11:02, Jiufu Guo wrote:
> Hi,
>
> Trunk gcc supports more constants to be built via two instructions: e.g.
> "li/lis; xori/xoris/rldicl/rldicr/rldic".
> And then num_insns_constant should also be updated.
>
> Function "rs6000_emit_set_long_const" is used to build complicate
> constants; and "num_insns_constant_gpr" is used to compute 'how
> many instructions are needed" to build the constant. So, these
> two functions should be aligned.
>
> The idea is: reusing "rs6000_emit_set_long_const" to compute/record
> the instruction number(when computing the insn_num, then do not emit
> instructions).
>
> Compare with previous verions:
> https://gcc.gnu.org/pipermail/gcc-patches/2023-October/634195.html
> This verion adds an argument to "rs6000_emit_set_long_const" to
> indicate computing instruction number instead emit intructions.
>
> Bootstrap & regtest pass ppc64{,le}.
> Is this ok for trunk?
>
> BR,
> Jeff (Jiufu Guo)
>
> gcc/ChangeLog:
>
> * config/rs6000/rs6000.cc (rs6000_emit_set_long_const): Add new
> parameter to record number of instructions to build the constant.
> (num_insns_constant_gpr): Call rs6000_emit_set_long_const to compute
> num_insn.
> (ADJUST_INSN_NUM_AND_RET): New macro.
> (rs6000_emit_set_const): Call rs6000_emit_set_long_const with NULL
> argument.
>
> ---
> gcc/config/rs6000/rs6000.cc | 245 +++++++++++++++++++-----------------
> 1 file changed, 133 insertions(+), 112 deletions(-)
>
> diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
> index cc24dd5301e..ba40dd6eee4 100644
> --- a/gcc/config/rs6000/rs6000.cc
> +++ b/gcc/config/rs6000/rs6000.cc
> @@ -1115,7 +1115,7 @@ static tree rs6000_handle_longcall_attribute (tree *, tree, tree, int, bool *);
> static tree rs6000_handle_altivec_attribute (tree *, tree, tree, int, bool *);
> static tree rs6000_handle_struct_attribute (tree *, tree, tree, int, bool *);
> static tree rs6000_builtin_vectorized_libmass (combined_fn, tree, tree);
> -static void rs6000_emit_set_long_const (rtx, HOST_WIDE_INT);
> +static void rs6000_emit_set_long_const (rtx, HOST_WIDE_INT, int *);
Make the new argument default as nullptr...
> static int rs6000_memory_move_cost (machine_mode, reg_class_t, bool);
> static bool rs6000_debug_rtx_costs (rtx, machine_mode, int, int, int *, bool);
> static int rs6000_debug_address_cost (rtx, machine_mode, addr_space_t,
> @@ -6054,21 +6054,9 @@ num_insns_constant_gpr (HOST_WIDE_INT value)
>
> else if (TARGET_POWERPC64)
> {
> - HOST_WIDE_INT low = sext_hwi (value, 32);
> - HOST_WIDE_INT high = value >> 31;
> -
> - if (high == 0 || high == -1)
> - return 2;
> -
> - high >>= 1;
> -
> - if (low == 0 || low == high)
> - return num_insns_constant_gpr (high) + 1;
> - else if (high == 0)
> - return num_insns_constant_gpr (low) + 1;
> - else
> - return (num_insns_constant_gpr (high)
> - + num_insns_constant_gpr (low) + 1);
> + int num_insns = 0;
> + rs6000_emit_set_long_const (NULL, value, &num_insns);
> + return num_insns;
> }
>
> else
> @@ -10284,7 +10272,7 @@ rs6000_emit_set_const (rtx dest, rtx source)
> emit_move_insn (lo, GEN_INT (c));
> }
> else
> - rs6000_emit_set_long_const (dest, c);
> + rs6000_emit_set_long_const (dest, c, NULL);
... then we don't need to change this line.
> break;
>
> default:
> @@ -10494,14 +10482,13 @@ can_be_built_by_li_and_rldic (HOST_WIDE_INT c, int *shift, HOST_WIDE_INT *mask)
>
> /* Subroutine of rs6000_emit_set_const, handling PowerPC64 DImode.
> Output insns to set DEST equal to the constant C as a series of
> - lis, ori and shl instructions. */
> + lis, ori and shl instructions. If NUM_INSNS is not NULL, then
> + only increase *NUM_INSNS as the number of insns, and do not output
> + real insns. */
>
> static void
> -rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c)
> +rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c, int *num_insns)
> {
> - rtx temp;
> - int shift;
> - HOST_WIDE_INT mask;
> HOST_WIDE_INT ud1, ud2, ud3, ud4;
>
> ud1 = c & 0xffff;
> @@ -10509,41 +10496,71 @@ rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c)
> ud3 = (c >> 32) & 0xffff;
> ud4 = (c >> 48) & 0xffff;
>
> + /* This macro RETURNs this function after increasing *NUM_INSNS!!! */
> +#define ADJUST_INSN_NUM_AND_RET(N) \
> + if (num_insns) \
> + { \
> + *num_insns += (N); \
> + return; \
> + }
This macro and its uses below can still have the chance to get the inconsistent
counts, as in some arms the counts get pre-computed.
Can we introduce one lambda function named as count_or_emit_insn and use it to
replace all the current uses of emit_move_insn in this function? If so, each place
where we emit insn will do the counting accordingly, it avoids the possible mismatch
pre-computed count (out-of-date in some day).
BR,
Kewen
> +
> if ((ud4 == 0xffff && ud3 == 0xffff && ud2 == 0xffff && (ud1 & 0x8000))
> - || (ud4 == 0 && ud3 == 0 && ud2 == 0 && ! (ud1 & 0x8000)))
> - emit_move_insn (dest, GEN_INT (sext_hwi (ud1, 16)));
> + || (ud4 == 0 && ud3 == 0 && ud2 == 0 && !(ud1 & 0x8000)))
> + {
> + /* li */
> + ADJUST_INSN_NUM_AND_RET (1);
> + emit_move_insn (dest, GEN_INT (sext_hwi (ud1, 16)));
> + return;
> + }
> +
> + rtx temp = num_insns
> + ? NULL
> + : (!can_create_pseudo_p () ? dest : gen_reg_rtx (DImode));
>
> - else if ((ud4 == 0xffff && ud3 == 0xffff && (ud2 & 0x8000))
> - || (ud4 == 0 && ud3 == 0 && ! (ud2 & 0x8000)))
> + if ((ud4 == 0xffff && ud3 == 0xffff && (ud2 & 0x8000))
> + || (ud4 == 0 && ud3 == 0 && !(ud2 & 0x8000)))
> {
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> + /* lis[; ori] */
> + ADJUST_INSN_NUM_AND_RET (ud1 != 0 ? 2 : 1);
>
> emit_move_insn (ud1 != 0 ? temp : dest,
> GEN_INT (sext_hwi (ud2 << 16, 32)));
> if (ud1 != 0)
> emit_move_insn (dest, gen_rtx_IOR (DImode, temp, GEN_INT (ud1)));
> + return;
> }
> - else if (ud4 == 0xffff && ud3 == 0xffff && !(ud2 & 0x8000) && ud1 == 0)
> +
> + if (ud4 == 0xffff && ud3 == 0xffff && !(ud2 & 0x8000) && ud1 == 0)
> {
> /* lis; xoris */
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> + ADJUST_INSN_NUM_AND_RET (2);
> +
> emit_move_insn (temp, GEN_INT (sext_hwi ((ud2 | 0x8000) << 16, 32)));
> emit_move_insn (dest, gen_rtx_XOR (DImode, temp, GEN_INT (0x80000000)));
> + return;
> }
> - else if (ud4 == 0xffff && ud3 == 0xffff && (ud1 & 0x8000))
> +
> + if (ud4 == 0xffff && ud3 == 0xffff && (ud1 & 0x8000))
> {
> /* li; xoris */
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> + ADJUST_INSN_NUM_AND_RET (2);
> +
> emit_move_insn (temp, GEN_INT (sext_hwi (ud1, 16)));
> emit_move_insn (dest, gen_rtx_XOR (DImode, temp,
> GEN_INT ((ud2 ^ 0xffff) << 16)));
> + return;
> }
> - else if (can_be_built_by_li_lis_and_rotldi (c, &shift, &mask)
> - || can_be_built_by_li_lis_and_rldicl (c, &shift, &mask)
> - || can_be_built_by_li_lis_and_rldicr (c, &shift, &mask)
> - || can_be_built_by_li_and_rldic (c, &shift, &mask))
> +
> + int shift;
> + HOST_WIDE_INT mask;
> + if (can_be_built_by_li_lis_and_rotldi (c, &shift, &mask)
> + || can_be_built_by_li_lis_and_rldicl (c, &shift, &mask)
> + || can_be_built_by_li_lis_and_rldicr (c, &shift, &mask)
> + || can_be_built_by_li_and_rldic (c, &shift, &mask))
> {
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> + /* li/lis; rldicX */
> + ADJUST_INSN_NUM_AND_RET (2);
> +
> unsigned HOST_WIDE_INT imm = (c | ~mask);
> imm = (imm >> shift) | (imm << (HOST_BITS_PER_WIDE_INT - shift));
>
> @@ -10553,49 +10570,48 @@ rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c)
> if (mask != HOST_WIDE_INT_M1)
> temp = gen_rtx_AND (DImode, temp, GEN_INT (mask));
> emit_move_insn (dest, temp);
> +
> + return;
> }
> - else if (ud3 == 0 && ud4 == 0)
> - {
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
>
> - gcc_assert (ud2 & 0x8000);
> + if (ud3 == 0 && ud4 == 0)
> + {
> + gcc_assert ((ud2 & 0x8000) && ud1 != 0);
> + ADJUST_INSN_NUM_AND_RET (!(ud1 & 0x8000) ? 2 : 3);
>
> - if (ud1 == 0)
> - {
> - /* lis; rldicl */
> - emit_move_insn (temp, GEN_INT (sext_hwi (ud2 << 16, 32)));
> - emit_move_insn (dest,
> - gen_rtx_AND (DImode, temp, GEN_INT (0xffffffff)));
> - }
> - else if (!(ud1 & 0x8000))
> + if (!(ud1 & 0x8000))
> {
> /* li; oris */
> emit_move_insn (temp, GEN_INT (ud1));
> emit_move_insn (dest,
> gen_rtx_IOR (DImode, temp, GEN_INT (ud2 << 16)));
> + return;
> }
> - else
> - {
> - /* lis; ori; rldicl */
> - emit_move_insn (temp, GEN_INT (sext_hwi (ud2 << 16, 32)));
> - emit_move_insn (temp, gen_rtx_IOR (DImode, temp, GEN_INT (ud1)));
> - emit_move_insn (dest,
> - gen_rtx_AND (DImode, temp, GEN_INT (0xffffffff)));
> - }
> +
> + /* lis; ori; rldicl */
> + emit_move_insn (temp, GEN_INT (sext_hwi (ud2 << 16, 32)));
> + emit_move_insn (temp, gen_rtx_IOR (DImode, temp, GEN_INT (ud1)));
> + emit_move_insn (dest, gen_rtx_AND (DImode, temp, GEN_INT (0xffffffff)));
> + return;
> }
> - else if (ud1 == ud3 && ud2 == ud4)
> +
> + if (ud1 == ud3 && ud2 == ud4)
> {
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> HOST_WIDE_INT num = (ud2 << 16) | ud1;
> - rs6000_emit_set_long_const (temp, sext_hwi (num, 32));
> + rs6000_emit_set_long_const (temp, sext_hwi (num, 32), num_insns);
> +
> + /* Adjust(+1) insn number after half part is adjusted. */
> + ADJUST_INSN_NUM_AND_RET (1);
> +
> rtx one = gen_rtx_AND (DImode, temp, GEN_INT (0xffffffff));
> rtx two = gen_rtx_ASHIFT (DImode, temp, GEN_INT (32));
> emit_move_insn (dest, gen_rtx_IOR (DImode, one, two));
> + return;
> }
> - else if ((ud4 == 0xffff && (ud3 & 0x8000))
> - || (ud4 == 0 && ! (ud3 & 0x8000)))
> +
> + if ((ud4 == 0xffff && (ud3 & 0x8000)) || (ud4 == 0 && !(ud3 & 0x8000)))
> {
> - temp = !can_create_pseudo_p () ? dest : gen_reg_rtx (DImode);
> + ADJUST_INSN_NUM_AND_RET (2 + (ud2 != 0 ? 1 : 0) + (ud1 != 0 ? 1 : 0));
>
> emit_move_insn (temp, GEN_INT (sext_hwi (ud3 << 16, 32)));
> if (ud2 != 0)
> @@ -10604,73 +10620,78 @@ rs6000_emit_set_long_const (rtx dest, HOST_WIDE_INT c)
> gen_rtx_ASHIFT (DImode, temp, GEN_INT (16)));
> if (ud1 != 0)
> emit_move_insn (dest, gen_rtx_IOR (DImode, temp, GEN_INT (ud1)));
> + return;
> }
> - else if (TARGET_PREFIXED)
> +
> + if (TARGET_PREFIXED)
> {
> if (can_create_pseudo_p ())
> {
> - /* pli A,L + pli B,H + rldimi A,B,32,0. */
> - temp = gen_reg_rtx (DImode);
> + /* pli A,L; pli B,H; rldimi A,B,32,0. */
> + ADJUST_INSN_NUM_AND_RET (3);
> +
> rtx temp1 = gen_reg_rtx (DImode);
> emit_move_insn (temp, GEN_INT ((ud4 << 16) | ud3));
> emit_move_insn (temp1, GEN_INT ((ud2 << 16) | ud1));
> -
> emit_insn (gen_rotldi3_insert_3 (dest, temp, GEN_INT (32), temp1,
> GEN_INT (0xffffffff)));
> + return;
> }
> - else
> - {
> - /* pli A,H + sldi A,32 + paddi A,A,L. */
> - emit_move_insn (dest, GEN_INT ((ud4 << 16) | ud3));
>
> - emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (32)));
> + /* There may be 1 insn inaccurate because of no info about dest. */
> + ADJUST_INSN_NUM_AND_RET (2 + (ud2 != 0 && ud1 != 0 ? 2 : 1));
>
> - bool can_use_paddi = REGNO (dest) != FIRST_GPR_REGNO;
> + bool can_use_paddi = REGNO (dest) != FIRST_GPR_REGNO;
>
> - /* Use paddi for the low 32 bits. */
> - if (ud2 != 0 && ud1 != 0 && can_use_paddi)
> - emit_move_insn (dest, gen_rtx_PLUS (DImode, dest,
> - GEN_INT ((ud2 << 16) | ud1)));
> + /* pli A,H; sldi A,32; paddi A,A,L. */
> + emit_move_insn (dest, GEN_INT ((ud4 << 16) | ud3));
> + emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (32)));
>
> - /* Use oris, ori for low 32 bits. */
> - if (ud2 != 0 && (ud1 == 0 || !can_use_paddi))
> - emit_move_insn (dest,
> - gen_rtx_IOR (DImode, dest, GEN_INT (ud2 << 16)));
> - if (ud1 != 0 && (ud2 == 0 || !can_use_paddi))
> - emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
> - }
> + /* Use paddi for the low 32 bits. */
> + if (ud2 != 0 && ud1 != 0 && can_use_paddi)
> + emit_move_insn (dest, gen_rtx_PLUS (DImode, dest,
> + GEN_INT ((ud2 << 16) | ud1)));
> + /* Use oris, ori for low 32 bits. */
> + if (ud2 != 0 && (ud1 == 0 || !can_use_paddi))
> + emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud2 << 16)));
> + if (ud1 != 0 && (ud2 == 0 || !can_use_paddi))
> + emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
> + return;
> }
> - else
> - {
> - if (can_create_pseudo_p ())
> - {
> - /* lis HIGH,UD4 ; ori HIGH,UD3 ;
> - lis LOW,UD2 ; ori LOW,UD1 ; rldimi LOW,HIGH,32,0. */
> - rtx high = gen_reg_rtx (DImode);
> - rtx low = gen_reg_rtx (DImode);
> - HOST_WIDE_INT num = (ud2 << 16) | ud1;
> - rs6000_emit_set_long_const (low, sext_hwi (num, 32));
> - num = (ud4 << 16) | ud3;
> - rs6000_emit_set_long_const (high, sext_hwi (num, 32));
> - emit_insn (gen_rotldi3_insert_3 (dest, high, GEN_INT (32), low,
> - GEN_INT (0xffffffff)));
> - }
> - else
> - {
> - /* lis DEST,UD4 ; ori DEST,UD3 ; rotl DEST,32 ;
> - oris DEST,UD2 ; ori DEST,UD1. */
> - emit_move_insn (dest, GEN_INT (sext_hwi (ud4 << 16, 32)));
> - if (ud3 != 0)
> - emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud3)));
>
> - emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (32)));
> - if (ud2 != 0)
> - emit_move_insn (dest,
> - gen_rtx_IOR (DImode, dest, GEN_INT (ud2 << 16)));
> - if (ud1 != 0)
> - emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
> - }
> + if (can_create_pseudo_p ())
> + {
> + /* lis HIGH,UD4 ; ori HIGH,UD3 ;
> + lis LOW,UD2 ; ori LOW,UD1 ; rldimi LOW,HIGH,32,0. */
> + rtx high = num_insns ? nullptr : gen_reg_rtx (DImode);
> + rtx low = num_insns ? nullptr : gen_reg_rtx (DImode);
> + HOST_WIDE_INT num = (ud2 << 16) | ud1;
> + rs6000_emit_set_long_const (low, sext_hwi (num, 32), num_insns);
> + num = (ud4 << 16) | ud3;
> + rs6000_emit_set_long_const (high, sext_hwi (num, 32), num_insns);
> +
> + /* Adjust(+1) insn number after 'high' and 'low' parts are adjusted. */
> + ADJUST_INSN_NUM_AND_RET (1);
> + emit_insn (gen_rotldi3_insert_3 (dest, high, GEN_INT (32), low,
> + GEN_INT (0xffffffff)));
> + return;
> }
> +
> + /* lis DEST,UD4 ; ori DEST,UD3 ; rotl DEST,32 ;
> + oris DEST,UD2 ; ori DEST,UD1. */
> + ADJUST_INSN_NUM_AND_RET (2 + (ud3 != 0 ? 1 : 0) + (ud2 != 0 ? 1 : 0)
> + + (ud1 != 0 ? 1 : 0));
> + emit_move_insn (dest, GEN_INT (sext_hwi (ud4 << 16, 32)));
> + if (ud3 != 0)
> + emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud3)));
> +
> + emit_move_insn (dest, gen_rtx_ASHIFT (DImode, dest, GEN_INT (32)));
> + if (ud2 != 0)
> + emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud2 << 16)));
> + if (ud1 != 0)
> + emit_move_insn (dest, gen_rtx_IOR (DImode, dest, GEN_INT (ud1)));
> +
> + return;
> }
>
> /* Helper for the following. Get rid of [r+r] memory refs
next prev parent reply other threads:[~2023-11-22 9:15 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-15 3:02 Jiufu Guo
2023-11-15 3:02 ` [PATCH V2 2/3] Using pli to split 34bits constant Jiufu Guo
2023-11-22 9:18 ` Kewen.Lin
2023-11-27 2:49 ` Jiufu Guo
2023-11-15 3:02 ` [PATCH V2 3/3] split complicate constant to memory Jiufu Guo
2023-11-22 9:12 ` Kewen.Lin [this message]
2023-11-27 2:59 ` [PATCH V2 1/3]rs6000: update num_insns_constant for 2 insns Jiufu Guo
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=82d4c257-fd86-6879-62f4-814e40ee3f58@linux.ibm.com \
--to=linkw@linux.ibm.com \
--cc=bergner@linux.ibm.com \
--cc=dje.gcc@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=guojiufu@linux.ibm.com \
--cc=linkw@gcc.gnu.org \
--cc=segher@kernel.crashing.org \
/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).