From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 99469 invoked by alias); 6 Apr 2018 10:24:30 -0000 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 Received: (qmail 99021 invoked by uid 89); 6 Apr 2018 10:23:53 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-26.1 required=5.0 tests=BAYES_00,GIT_PATCH_0,GIT_PATCH_1,GIT_PATCH_2,GIT_PATCH_3,KAM_ASCII_DIVIDERS,SPF_PASS,T_RP_MATCHES_RCVD autolearn=ham version=3.3.2 spammy= X-HELO: mx2.suse.de Received: from mx2.suse.de (HELO mx2.suse.de) (195.135.220.15) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Fri, 06 Apr 2018 10:23:52 +0000 Received: from relay2.suse.de (charybdis-ext.suse.de [195.135.220.254]) by mx2.suse.de (Postfix) with ESMTP id 0B254AD29; Fri, 6 Apr 2018 10:23:50 +0000 (UTC) Date: Fri, 06 Apr 2018 10:24:00 -0000 From: Richard Biener To: Alan Modra cc: Tamar Christina , gcc-patches@gcc.gnu.org, nd@arm.com, law@redhat.com, ian@airs.com, bergner@vnet.ibm.com Subject: Re: [PATCH][GCC][mid-end] Fix PR85123 incorrect copies In-Reply-To: <20180405234558.GS3812@bubble.grove.modra.org> Message-ID: References: <20180405122903.GA509@arm.com> <20180405234558.GS3812@bubble.grove.modra.org> User-Agent: Alpine 2.20 (LSU 67 2015-01-07) MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII X-SW-Source: 2018-04/txt/msg00305.txt.bz2 On Fri, 6 Apr 2018, Alan Modra wrote: > On Thu, Apr 05, 2018 at 01:29:06PM +0100, Tamar Christina wrote: > > diff --git a/gcc/expr.c b/gcc/expr.c > > index 00660293f72e5441a6421a280b04c57fca2922b8..7daeb8c91d758edf0b3dc37f6927380b6f3df877 100644 > > --- a/gcc/expr.c > > +++ b/gcc/expr.c > > @@ -2749,7 +2749,7 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src) > > { > > int i, n_regs; > > unsigned HOST_WIDE_INT bitpos, xbitpos, padding_correction = 0, bytes; > > - unsigned int bitsize; > > + unsigned int bitsize = 0; > > rtx *dst_words, dst, x, src_word = NULL_RTX, dst_word = NULL_RTX; > > /* No current ABI uses variable-sized modes to pass a BLKmnode type. */ > > fixed_size_mode mode = as_a (mode_in); > > @@ -2782,7 +2782,7 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src) > > > > n_regs = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; > > dst_words = XALLOCAVEC (rtx, n_regs); > > - bitsize = BITS_PER_WORD; > > + > > if (targetm.slow_unaligned_access (word_mode, TYPE_ALIGN (TREE_TYPE (src)))) > > bitsize = MIN (TYPE_ALIGN (TREE_TYPE (src)), BITS_PER_WORD); > > > > You calculate bitsize here, then override it in the loop? Doesn't > that mean strict align targets will use mis-aligned loads and stores? > > > @@ -2791,6 +2791,17 @@ copy_blkmode_to_reg (machine_mode mode_in, tree src) > > bitpos < bytes * BITS_PER_UNIT; > > bitpos += bitsize, xbitpos += bitsize) > > { > > + /* Find the largest integer mode that can be used to copy all or as > > + many bits as possible of the structure. */ > > + opt_scalar_int_mode mode_iter; > > + FOR_EACH_MODE_IN_CLASS (mode_iter, MODE_INT) > > + if (GET_MODE_BITSIZE (mode_iter.require ()) > > + <= ((bytes * BITS_PER_UNIT) - bitpos) > > + && GET_MODE_BITSIZE (mode_iter.require ()) <= BITS_PER_WORD) > > + bitsize = GET_MODE_BITSIZE (mode_iter.require ()); > > + else > > + break; > > + > > This isn't correct. Consider a 6 byte struct on a 4 byte word, 8 bit > byte, big-endian target when targetm.calls.return_in_msb is false. > > In this scenario, copy_blkmode_to_reg should return two registers, set > as if they had been loaded from two words in memory laid out as > follows (left to right increasing byte addresses): > _______________________ _______________________ > | 0 | 0 | s0 | s1 | | s2 | s3 | s4 | s5 | > ~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~ > > So we will have xbitpos=16 first time around the loop. That means > your new code will attempt to store 32 bits into a bit-field starting > at bit 16 in the first 32-bit register, and of course fail. > > This scenario used to be handled correctly, at least when the struct > wasn't over-aligned. I wonder if it's best to revert the original regression causing patch and look for a proper solution in the GCC 9 timeframe? Thanks, Richard.