public inbox for gcc-cvs@sourceware.org
help / color / mirror / Atom feed
From: Michael Meissner <meissner@gcc.gnu.org>
To: gcc-cvs@gcc.gnu.org
Subject: [gcc(refs/users/meissner/heads/dmf004)] Limit memcpy inline copy to do just 2 variable moves.
Date: Thu, 17 Nov 2022 21:55:41 +0000 (GMT)	[thread overview]
Message-ID: <20221117215541.3668A384F6D3@sourceware.org> (raw)

https://gcc.gnu.org/g:b3cecdfb43c2134a65833eaa415fef65d4e24b98

commit b3cecdfb43c2134a65833eaa415fef65d4e24b98
Author: Michael Meissner <meissner@linux.ibm.com>
Date:   Wed Nov 16 15:55:46 2022 -0500

    Limit memcpy inline copy to do just 2 variable moves.
    
    2022-11-16   Michael Meissner  <meissner@linux.ibm.com>
    
    gcc/
    
            * config/rs6000/rs6000-string.cc (expand_lxvl_stxvl): New helper
            function.
            (expand_block_move_variable): Only do 2 variable moves per memcpy call.
            * config/rs6000/rs6000.cc (rs6000_option_override_internal): Remove
            bumping up rs6000_memcpy_inline_bytes if -mcpu=future.
            * config/rs6000/rs6000.opt (-param=rs6000-memcpy-inline-bytes=): Set
            default to 32, not 16.

Diff:
---
 gcc/config/rs6000/rs6000-string.cc | 69 +++++++++++++++++++++++++++++++-------
 gcc/config/rs6000/rs6000.cc        |  6 ----
 gcc/config/rs6000/rs6000.opt       |  4 +--
 3 files changed, 58 insertions(+), 21 deletions(-)

diff --git a/gcc/config/rs6000/rs6000-string.cc b/gcc/config/rs6000/rs6000-string.cc
index bc44d27773b..216c0c1cc93 100644
--- a/gcc/config/rs6000/rs6000-string.cc
+++ b/gcc/config/rs6000/rs6000-string.cc
@@ -2735,6 +2735,32 @@ gen_lxvl_stxvl_move (rtx dest, rtx src, int length)
     return gen_lxvl (dest, addr, len);
 }
 
+/* Generate a lxvl/stxvl or lxvprl/stxvprl pair of instructions to move up to
+   16 or 32 bytes at a time.  */
+
+static void
+expand_lxvl_stxvl (rtx dest_addr,
+		   rtx src_addr,
+		   rtx bytes_rtx,
+		   int max_bytes)
+{
+  if (max_bytes > GET_MODE_SIZE (V16QImode))
+    {
+      rtx vreg = gen_reg_rtx (XOmode);
+      emit_insn (gen_lxvprl (vreg, src_addr, bytes_rtx));
+      emit_insn (gen_stxvprl (vreg, dest_addr, bytes_rtx));
+    }
+
+  else
+    {
+      rtx vreg = gen_reg_rtx (V16QImode);
+      emit_insn (gen_lxvl (vreg, src_addr, bytes_rtx));
+      emit_insn (gen_stxvl (vreg, dest_addr, bytes_rtx));
+    }
+
+  return;
+}
+
 /* Expand a variable block move operation, and return 1 if successful.  Return
    0 if we should let the compiler generate normal code.
 
@@ -2766,11 +2792,16 @@ expand_block_move_variable (rtx operands[], bool might_overlap)
 		   ? copy_to_reg (operands[2])
 		   : convert_to_mode (Pmode, operands[2], true));
 
-  int var_size_int
+  /* Maximum size to move at one time.  */
+  int vect_size_int
     = (TARGET_FUTURE ? GET_MODE_SIZE (XOmode) : GET_MODE_SIZE (V16QImode));
 
-  if (var_size_int > rs6000_memcpy_inline_bytes)
-    var_size_int = rs6000_memcpy_inline_bytes;
+  /* Total size to move.  Limit the number of bytes that we do in this
+     optimization to just 2 variable moves.  Anything larger let the memcpy
+     glibc function handle it and do the extra optimizations it provides.  */
+  int var_size_int = (rs6000_memcpy_inline_bytes > (2 * vect_size_int)
+		      ? 2 * vect_size_int
+		      : rs6000_memcpy_inline_bytes);
 
   rtx var_size = GEN_INT (var_size_int);
   rtx var_cr = gen_reg_rtx (CCUNSmode);
@@ -2797,17 +2828,29 @@ expand_block_move_variable (rtx operands[], bool might_overlap)
   emit_label (var_label);
 
   /* We want to move bytes inline.  Move 0..16 or 0..32 bytes now.  */
-  if (var_size_int > GET_MODE_SIZE (V16QImode))
-    {
-      rtx vreg = gen_reg_rtx (XOmode);
-      emit_insn (gen_lxvprl (vreg, src_addr, bytes_rtx));
-      emit_insn (gen_stxvprl (vreg, dest_addr, bytes_rtx));
-    }
-  else
+  if (vect_size_int > var_size_int)
+    vect_size_int = var_size_int;
+
+  expand_lxvl_stxvl (dest_addr, src_addr, bytes_rtx, vect_size_int);
+
+  /* If we have more than 16/32 bytes, adjust the pointers/length and generate
+     a second move.  */
+  if (var_size_int > vect_size_int)
     {
-      rtx vreg = gen_reg_rtx (V16QImode);
-      emit_insn (gen_lxvl (vreg, src_addr, bytes_rtx));
-      emit_insn (gen_stxvl (vreg, dest_addr, bytes_rtx));
+      emit_insn (gen_add2_insn (bytes_rtx, GEN_INT (- vect_size_int)));
+
+      rtx vect_cr = gen_reg_rtx (CCmode);
+      emit_insn (gen_rtx_SET (vect_cr,
+			      gen_rtx_COMPARE (CCmode, bytes_rtx,
+					       const0_rtx)));
+
+      do_ifelse (CCmode, GT, NULL_RTX, NULL_RTX, vect_cr, join_label,
+		 profile_probability::likely ());
+
+      rtx ptr_adjust = GEN_INT (vect_size_int);
+      emit_insn (gen_add2_insn (dest_addr, ptr_adjust));
+      emit_insn (gen_add2_insn (src_addr, ptr_adjust));
+      expand_lxvl_stxvl (dest_addr, src_addr, bytes_rtx, vect_size_int);
     }
 
   emit_label (join_label);
diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
index e401aa4c02b..dba37df3c61 100644
--- a/gcc/config/rs6000/rs6000.cc
+++ b/gcc/config/rs6000/rs6000.cc
@@ -4878,12 +4878,6 @@ rs6000_option_override_internal (bool global_init_p)
 	SET_OPTION_IF_UNSET (&global_options, &global_options_set,
 			     param_vect_partial_vector_usage, 0);
 
-      /* If we have the lxvprl/stxvprl instructions, bump up the default size
-	 for doing inlining memcpy moves.  */
-      if (TARGET_FUTURE)
-	SET_OPTION_IF_UNSET (&global_options, &global_options_set,
-			     rs6000_memcpy_inline_bytes, 32);
-
       /* Use the 'model' -fsched-pressure algorithm by default.  */
       SET_OPTION_IF_UNSET (&global_options, &global_options_set,
 			   param_sched_pressure_algorithm,
diff --git a/gcc/config/rs6000/rs6000.opt b/gcc/config/rs6000/rs6000.opt
index c594877ebc6..602930063cd 100644
--- a/gcc/config/rs6000/rs6000.opt
+++ b/gcc/config/rs6000/rs6000.opt
@@ -689,6 +689,6 @@ When reduction factor computed for a loop exceeds the threshold specified by
 this parameter, prefer to unroll this loop.  The default value is 1.
 
 -param=rs6000-memcpy-inline-bytes=
-Target Undocumented Joined UInteger Var(rs6000_memcpy_inline_bytes) Init(16) Param
+Target Undocumented Joined UInteger Var(rs6000_memcpy_inline_bytes) Init(32) Param
 Maximum number of bytes to move with inline code before calling the memcpy
-library function.  The default value is 16.
+library function.  The default value is 32.

             reply	other threads:[~2022-11-17 21:55 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-11-17 21:55 Michael Meissner [this message]
  -- strict thread matches above, loose matches on Subject: below --
2022-11-16 20:56 Michael Meissner

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=20221117215541.3668A384F6D3@sourceware.org \
    --to=meissner@gcc.gnu.org \
    --cc=gcc-cvs@gcc.gnu.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).