From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: In-Reply-To: <1176262253.8061.60.camel@localhost.localdomain> References: <20070410111508.GA2969@localhost.localdomain> <1176240168.8061.51.camel@localhost.localdomain> <20070411025605.GB2197@localhost.localdomain> <1176262253.8061.60.camel@localhost.localdomain> Mime-Version: 1.0 (Apple Message framework v623) Content-Type: text/plain; charset=US-ASCII; format=flowed Message-Id: From: Segher Boessenkool Subject: Re: [patch 3/3] cell: prevent alignment interrupt on local store Date: Wed, 11 Apr 2007 23:03:28 +0200 To: Benjamin Herrenschmidt Cc: Arnd Bergmann , Akinobu Mita , linuxppc-dev@ozlabs.org, Paul Mackerras , cbe-oss-dev@ozlabs.org List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , >>> I would personally prefer building the entire kernel with -mno-string >>> when cell support is enabled... >> >> I'm about to resend the patch which prevent alignment interrupts on LS >> by using memcpy_fromio() rather than putting a collection of -mno-* >> options >> in Makefile. (only -mno-string is not enough to prevent that as >> Segher said) > > Sure, however we also know for sure that string instructions will suck > a > LOT on a cell ... so for a multi platforms kernel that has cell support > built-in, it makes some sense to not use them. [That's a separate issue and should be handled as a separate patch series.] Not only string insns are slow on the current Cell core, many more insns are, including all dot insns; and those are way more expensive in comparison. Turning off generation of string insns only won't help Cell much. Building a multiplatform kernel with all the insns that are expensive on Cell turned off will hurt a lot on most other CPUs. This isn't just a kernel issue either; the same holds for all of userspace. It seems to me the only way to get good performance on both Cell and all other platforms is to have a separate binary distribution for Cell. Segher