On Thu, 27 Jul 2006 17:21:00 +0000 "Jeffrey D. Smith" <[EMAIL PROTECTED]> wrote:
:>From: "Binyamin Dissen" <[EMAIL PROTECTED]> :>Sent: 7/27/2006 10:22 AM :>To: "IBM-MAIN@BAMA.UA.EDU" <IBM-MAIN@BAMA.UA.EDU> :>Subject: Re: Doing a LM of two words on a double word bounday - is it serialized? :>On Thu, 27 Jul 2006 15:54:00 +0000 "Jeffrey D. Smith" :><[EMAIL PROTECTED]> wrote: :>:>LM R2,R3,operand ; no check-point synch, no store. :>:>will concurrently load R2,R3. The values are fetched 8-bytes :>:>in parallel. If another CPU is concurrently storing 8-bytes, :>:>then your CPU will see the entire old 8-bytes or the entire :>:>new 8-bytes and never a mixture of old and new bytes. :>:>Note that if another CPU is storing with a narrower alignment, :>:>like a 4-byte store, then you will get inconsistent results :>:>with either CDS or LM/STM. :>In other words, if CPU1 is doing the LM of the doubleword while CPU2 is using :>PLO changing the two words via CSST (each separately on a word boundary), it :>is undefined whether CPU1 will get a consistent answer? :>I was thinking of using one word as a update-counter and the other a pointer :>to the chain. :>I was going to use PLO to do a compare/swap/store, with the compare/swap on :>the count and a store into the chain (possibly into the head pointer - thus :>both words in the doubleword would be accessed as single words). Obviously I :>would get bad results if the load of the chain header/count did not get a :>consistent result. :>Load word1 followed by PLO CL would obviously get consistent results. :>All CPUs that are fetching/storing must use the same alignment :>and the same serialization mechanism. For updating, they must :>all use CS/CDS/CSG/CDSG, or they must all use PLO with R1 pointing :>at the same lock token (I just use the list header address for R1). :>PLO is designed for discontiguous fetch/store. If it just happens :>that two words are adjacent in the same double word, that won't :>assure an 8-byte concurrent store as viewed by other CPU. It will :>look like two 4-byte concurrent stores separated in time. The PLO :>compare-swap-single/double/triple-store performs the compare :>first and the swap last. All of the accesses performed by PLO :>appear to other CPU to be separated in time. PLO simply uses a :>"red light/green light" hardware semaphore to delay other CPU :>(also using PLO) from accessing the operands. In other words, are you saying in the case above that the LM can get inconsistent results? That L/PLO-CL is required? -- Binyamin Dissen <[EMAIL PROTECTED]> http://www.dissensoftware.com Director, Dissen Software, Bar & Grill - Israel Should you use the mailblocks package and expect a response from me, you should preauthorize the dissensoftware.com domain. I very rarely bother responding to challenge/response systems, especially those from irresponsible companies. ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
