public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Sandiford <richard.sandiford@arm.com>
To: Richard Henderson <rth@redhat.com>
Cc: "gcc-patches\@gcc.gnu.org" <gcc-patches@gcc.gnu.org>,
	 Andreas Krebbel <krebbel@linux.vnet.ibm.com>
Subject: Re: RFA: Fix mode checks for possibly-constant predicates
Date: Wed, 03 Jun 2015 07:23:00 -0000	[thread overview]
Message-ID: <87sia971bv.fsf@e105548-lin.cambridge.arm.com> (raw)
In-Reply-To: <5568A761.20705@redhat.com> (Richard Henderson's message of "Fri,	29 May 2015 18:52:33 +0100")

Richard Henderson <rth@redhat.com> writes:
> On 05/29/2015 10:23 AM, Richard Sandiford wrote:
>> +  /* Check whether the predicate accepts const scalar ints (which always
>> +     have a stored mode of VOIDmode, but logically have a real mode)
>> +     and whether it matches anything besides const scalar ints.  */
>> +  bool matches_const_scalar_int_p = false;
>> +  bool matches_other_p = false;
>> +  for (int i = 0; i < NUM_RTX_CODE; ++i)
>> +    if (p->codes[i])
>> +      switch (i)
>> +	{
>> +	CASE_CONST_SCALAR_INT:
>> +	  matches_const_scalar_int_p = true;
>> +	  break;
>> +
>> +	default:
>> +	  matches_other_p = true;
>> +	  break;
>> +	}
>> +
>> +  /* There's no need for a mode check if the predicate only accepts
>> +     constant integers.  The code checks in the predicate are enough
>> +     to establish that the mode is VOIDmode.
>> +
>> +     Note that the predicate itself should check whether a scalar
>> +     integer is in range of the given mode.  */
>> +  if (!matches_other_p && !p->codes[CONST_DOUBLE])
>> +    return;
>
> I think perhaps it would be cleaner to not use CASE_CONST_SCALAR_INT,
> and then do
>
>   switch (i)
>     {
>     case CONST_INT:
>     case CONST_WIDE_INT:
>       matches_const_scalar_int_p = true;
>       break;
>
>     case CONST_DOUBLE:
>       if (!TARGET_SUPPORTS_WIDE_INT)
>         matches_const_scalar_int_p = true;
>       matches_other_p = true;
>       break;
>
>     default:
>       matches_other_p = true;
>       break;
>     }
>
>   if (!matches_other_p)
>     return;
>
> Otherwise ok.

Ah, yeah, that's better.  Here's what I applied after testing on
x86_64-linux-gnu.

Thanks,
Richard


gcc/
	* genpreds.c (mark_mode_tests): Mark all MATCH_CODEs as
	NO_MODE_TEST.
	(add_mode_tests): Don't add mode tests if the predicate only
	accepts scalar constant integers.  Otherwise, allow the mode
	of "op" to be VOIDmode if the predicate does accept such integers.

Index: gcc/genpreds.c
===================================================================
--- gcc/genpreds.c	2015-06-02 13:32:21.394938060 +0100
+++ gcc/genpreds.c	2015-06-02 13:34:50.377221396 +0100
@@ -218,11 +218,11 @@ needs_variable (rtx exp, const char *var
 
 /* Given an RTL expression EXP, find all subexpressions which we may
    assume to perform mode tests.  Normal MATCH_OPERAND does;
-   MATCH_CODE does if it applies to the whole expression and accepts
-   CONST_INT or CONST_DOUBLE; and we have to assume that MATCH_TEST
-   does not.  These combine in almost-boolean fashion - the only
-   exception is that (not X) must be assumed not to perform a mode
-   test, whether or not X does.
+   MATCH_CODE doesn't as such (although certain codes always have
+   VOIDmode); and we have to assume that MATCH_TEST does not.
+   These combine in almost-boolean fashion - the only exception is
+   that (not X) must be assumed not to perform a mode test, whether
+   or not X does.
 
    The mark is the RTL /v flag, which is true for subexpressions which
    do *not* perform mode tests.
@@ -244,10 +244,7 @@ mark_mode_tests (rtx exp)
       break;
 
     case MATCH_CODE:
-      if (XSTR (exp, 1)[0] != '\0'
-	  || (!strstr (XSTR (exp, 0), "const_int")
-	      && !strstr (XSTR (exp, 0), "const_double")))
-	NO_MODE_TEST (exp) = 1;
+      NO_MODE_TEST (exp) = 1;
       break;
 
     case MATCH_TEST:
@@ -313,6 +310,40 @@ add_mode_tests (struct pred_data *p)
   if (p->special)
     return;
 
+  /* Check whether the predicate accepts const scalar ints (which always
+     have a stored mode of VOIDmode, but logically have a real mode)
+     and whether it matches anything besides const scalar ints.  */
+  bool matches_const_scalar_int_p = false;
+  bool matches_other_p = false;
+  for (int i = 0; i < NUM_RTX_CODE; ++i)
+    if (p->codes[i])
+      switch (i)
+	{
+	case CONST_INT:
+	case CONST_WIDE_INT:
+	  matches_const_scalar_int_p = true;
+	  break;
+
+	case CONST_DOUBLE:
+	  if (!TARGET_SUPPORTS_WIDE_INT)
+	    matches_const_scalar_int_p = true;
+	  matches_other_p = true;
+	  break;
+
+	default:
+	  matches_other_p = true;
+	  break;
+	}
+
+  /* There's no need for a mode check if the predicate only accepts
+     constant integers.  The code checks in the predicate are enough
+     to establish that the mode is VOIDmode.
+
+     Note that the predicate itself should check whether a scalar
+     integer is in range of the given mode.  */
+  if (!matches_other_p)
+    return;
+
   mark_mode_tests (p->exp);
 
   /* If the whole expression already tests the mode, we're done.  */
@@ -320,7 +351,11 @@ add_mode_tests (struct pred_data *p)
     return;
 
   match_test_exp = rtx_alloc (MATCH_TEST);
-  XSTR (match_test_exp, 0) = "mode == VOIDmode || GET_MODE (op) == mode";
+  if (matches_const_scalar_int_p)
+    XSTR (match_test_exp, 0) = ("mode == VOIDmode || GET_MODE (op) == mode"
+				" || GET_MODE (op) == VOIDmode");
+  else
+    XSTR (match_test_exp, 0) = "mode == VOIDmode || GET_MODE (op) == mode";
   and_exp = rtx_alloc (AND);
   XEXP (and_exp, 1) = match_test_exp;
 

  reply	other threads:[~2015-06-03  6:28 UTC|newest]

Thread overview: 30+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2015-04-27 10:20 Mostly rewrite genrecog Richard Sandiford
2015-04-28 23:18 ` Jeff Law
2015-04-29 13:59   ` Richard Sandiford
2015-04-29  8:52 ` Eric Botcazou
2015-04-29 13:51   ` Richard Sandiford
2015-04-30 10:44     ` Eric Botcazou
2015-04-30  6:54 ` Bin.Cheng
2015-04-30 11:58   ` Richard Sandiford
2015-04-30  7:17 ` Andreas Schwab
2015-04-30  7:58   ` Richard Sandiford
2015-04-30 12:10     ` Andreas Schwab
2015-04-30 12:33       ` Richard Biener
2015-04-30 16:27         ` Richard Sandiford
2015-05-01 12:42           ` Richard Sandiford
2015-05-01 13:57             ` Jeff Law
2015-05-07  8:59 ` [nvptx] " Thomas Schwinge
2015-05-07  9:14   ` Jakub Jelinek
2015-05-21  8:09     ` Thomas Schwinge
2015-05-21 11:57       ` Bernd Schmidt
2015-05-08  9:10 ` genrecog: Address -Wsign-compare diagnostics (was: Mostly rewrite genrecog) Thomas Schwinge
2015-05-08 14:43   ` genrecog: Address -Wsign-compare diagnostics Richard Sandiford
2015-05-08 18:39   ` Jeff Law
2015-05-16  8:13 ` Mostly rewrite genrecog Andreas Krebbel
2015-05-17 22:05   ` Richard Sandiford
2015-05-22 16:14     ` Andreas Krebbel
2015-05-22 16:32       ` Richard Sandiford
2015-05-29 17:39         ` RFA: Fix mode checks for possibly-constant predicates Richard Sandiford
2015-05-29 19:27           ` Richard Henderson
2015-06-03  7:23             ` Richard Sandiford [this message]
2015-06-03  9:39               ` Richard Sandiford

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=87sia971bv.fsf@e105548-lin.cambridge.arm.com \
    --to=richard.sandiford@arm.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=krebbel@linux.vnet.ibm.com \
    --cc=rth@redhat.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).