From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 1803 invoked by alias); 23 Dec 2012 16:49:13 -0000 Received: (qmail 1786 invoked by uid 22791); 23 Dec 2012 16:49:10 -0000 X-SWARE-Spam-Status: No, hits=-3.2 required=5.0 tests=AWL,BAYES_00,DKIM_ADSP_CUSTOM_MED,DKIM_SIGNED,DKIM_VALID,FREEMAIL_FROM,KHOP_RCVD_TRUST,NML_ADSP_CUSTOM_MED,RCVD_IN_DNSWL_LOW,RCVD_IN_HOSTKARMA_YE X-Spam-Check-By: sourceware.org Received: from mail-wg0-f46.google.com (HELO mail-wg0-f46.google.com) (74.125.82.46) by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Sun, 23 Dec 2012 16:49:07 +0000 Received: by mail-wg0-f46.google.com with SMTP id dr13so2864457wgb.25 for ; Sun, 23 Dec 2012 08:49:03 -0800 (PST) X-Received: by 10.180.81.39 with SMTP id w7mr31664483wix.15.1356281343338; Sun, 23 Dec 2012 08:49:03 -0800 (PST) Received: from localhost ([2.28.234.219]) by mx.google.com with ESMTPS id fv2sm39268482wib.4.2012.12.23.08.49.01 (version=TLSv1/SSLv3 cipher=OTHER); Sun, 23 Dec 2012 08:49:02 -0800 (PST) From: Richard Sandiford To: gcc-patches@gcc.gnu.org Mail-Followup-To: gcc-patches@gcc.gnu.org, rdsandiford@googlemail.com Subject: RFA: Fix uint128_t range checking in VRP Date: Sun, 23 Dec 2012 16:49:00 -0000 Message-ID: <87d2y0vhms.fsf@talisman.default> User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/24.1 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Id: List-Archive: List-Post: List-Help: Sender: gcc-patches-owner@gcc.gnu.org X-SW-Source: 2012-12/txt/msg01374.txt.bz2 The minimum uint128_t value is an all-zeros double_int and the maximum value is an all-ones double_int. Truncating these values to int and then sign-extending them gives the same all-zeros and all-ones values, so tree-vrp.c:range_fits_type_p concludes that uint128_t fits in an int. This showed up as a miscompilation of __fractutasf on mips64-linux-gnu, which in turn showed up in convert-float-4.c. There was already code to handle this problem when changing the sign and keeping the precision, but I think the rule applies regardless of precision. Tested on x86_64-linux-gnu and mips64-linux-gnu. OK to install? Richard gcc/ * tree-vrp.c (range_fits_type_p): Require the MSB of the double_int to be clear for sign changes. gcc/testsuite/ * gcc.dg/torture/fp-int-convert-2.c: New test. Index: gcc/tree-vrp.c =================================================================== --- gcc/tree-vrp.c 2012-12-23 11:50:39.046419271 +0000 +++ gcc/tree-vrp.c 2012-12-23 11:51:17.814677008 +0000 @@ -8766,9 +8766,11 @@ range_fits_type_p (value_range_t *vr, un || TREE_CODE (vr->max) != INTEGER_CST) return false; - /* For precision-preserving sign-changes the MSB of the double-int - has to be clear. */ - if (src_precision == precision + /* For sign changes, the MSB of the double_int has to be clear. + An unsigned value with its MSB set cannot be represented by + a signed double_int, while a negative value cannot be represented + by an unsigned double_int. */ + if (TYPE_UNSIGNED (src_type) != unsigned_p && (TREE_INT_CST_HIGH (vr->min) | TREE_INT_CST_HIGH (vr->max)) < 0) return false; Index: gcc/testsuite/gcc.dg/torture/fp-int-convert-2.c =================================================================== --- /dev/null 2012-12-03 19:06:31.446621561 +0000 +++ gcc/testsuite/gcc.dg/torture/fp-int-convert-2.c 2012-12-23 13:46:35.449077882 +0000 @@ -0,0 +1,18 @@ +/* { dg-do run } */ +/* { dg-require-effective-target int128 } */ + +extern void abort (void); + +float __attribute__((noinline)) +f (__uint128_t x) +{ + return x + 1; +} + +int +main (void) +{ + if (f (0xffffffffu) == 0) + abort (); + return 0; +}