23 August 1995 Initial announcement of Speak Freely Release 5.0. 24 August 1995 Peter Claus Gutman (pgut01@cs.auckland.az.nz), developer of a very nice encryption library, wrote to suggest his library might prove useful. In the source code for the library, I found a clever 80x86 assembly-language implementation of IDEA, made freely available by its author, who is identified in the source code only as "Bryan". Whoever you are, Bryan, great piece of work! If you, or somebody who knows who you are, happens to read this, let me know so I can give complete attribution. I integrated the assembly language loop into IDEA\IDEA.C, modifying it slightly to work with Microsoft Visual C's inline assembler, so you don't need a separate assembler to take advantage of the optimised code. Whether the assembler or original C code is used depends upon whether USE_ASM is defined, so you can use the original loop for reference or if, for example, your compiler doesn't support inline assembly code or is incompatible with the way Microsoft do it. Enabling the assembly language code increased the speed of IDEA encryption and decryption on my 486/50 machine from 152,000 bytes per second to 242,000 bytes per second--well worth the trouble of integrating the code. 30 August 1995 Completed a massive revision to avoid all packet fragmentation and thus work with WINSOCK drivers such as Trumpet WINSOCK. The changes were so great and ubiquitous there's no point in trying to describe them. In debugging the changes, one of the mysteries that has been dogging me was finally solved--the random hangs, loss of synchronism, failure to release resources, etc. etc. etc. were the result of Windows discarding messages to the main window as a result of overflows of the default 8 message queue. Speak Freely juggles a lot of balls in the air at once, and it's very easy to hit this limit. At initialisation time, we now try to expand the queue to its maximum size of 120 messages or whatever lower maximum the system we're running on supports. Added the "Extended Status" (Propeller Head) dialogue. Made compression modes global rather than per-connection. This means compression only has to be done once, which speeds up party line transmissions. The change is necessary in any case so that packet size can be optimised. 1 September 1995 Update release 5.1. 8 September 1995 CreateSocket() in UTILITY.C contained a "defensive bind()" to address zero as a work-around for some defective WINSOCK implementations. Unfortunately, this work-around causes other TCP/IP stacks, including that built into Windows NT to fail. I made the nugatory bind conditional on a new Options/Workarounds/Always Bind Socket menu item which is, of course, saved in the .INI file. Update release 5.1a. 9 September 1995 Discovered that the reason the socket write was failing on Windows NT and Windows 95 is that Microsoft's built-in WINSOCK, entirely incompatible with Unix and every other WINSOCK I have encountered, refuses sendto() once a datagram socket has been connect()ed. The sole function of connect() on a datagram socket is to specify a default address so subsequent writes can be done with send() (or, on Unix, write()), and there is no prohibition of overriding this default address with a subsequent sendto(). The WINSOCK specification nowhere mentions such a restriction as a Windows-specific change. I modified the socket write code in CONNECT.C and the loop-back socket write in FRAME.C to first try sendto(). If it fails, send() is then tried and if that works all subsequent socket writes for the rest of the session are done using send(). This code has been verified to work on both Windows NT and Windows 95 (first customer shipment edition). Special thanks to John Deters (jad@DHDSC.MN.ORG) who both identified the source of this problem on Windows NT and tested innumerable versions slowly converging toward the actual fix. Added an item to the Propeller Head dialogue to indicate whether sendto() or send() is being used to write to outbound sockets; it's "Sending with" in the "Network" box. Tested with the WINSOCK implementation included with Sun PC-NFS 5.1. Works fine. Update release 5.1b. 10 September 1995 After last week's experience I decided to indulge in some preemptive workarounds for crummy network and sound card drivers which fail in obvious ways which haven't bitten me yet. I expanded the Options/Workarounds menu to include: Audio Assume Half-duplex Assumes the sound card is half-duplex without requiring it to fail an output open while input is open. Accommodates cards which are actually half-duplex but don't indicate this by failing a simultaneous input and output open. Also handles cards which crash the system or application when you try to open them in full-duplex mode. Assume 11025 Samples/sec Assumes the card is capable only of 11025 samples per second mode, not our preferred 8000 samples per second. Permits correct operation on cards which don't fail when opened with a sample rate of 8000 samples per second but which can't actually run at that rate. Network Always Bind Socket As before; bind outbound sockets, even though there's no need to do so. Never Connect Outbound Socket Don't connect() the output sockets. This implies we'll always use sendto() to write to those sockets. Clears "Use send(), Not sendto()" mode if set. Use send(), Not sendto() Always use send() to write to outbound sockets; don't wait for a sendto() to fail first. Accommodates drivers where a sendto() on a connected socket crashes the application or system. Clears "Never Connect Outbound Socket" mode if set. All the workaround modes are saved in the SPEAKFRE.INI file and apply to subsequent executions. The menu items are disabled when a connection is active. As suggested by John Deters (jad@DHDSC.MN.ORG), I added the ability to automatically open an iconised version of Speak Freely whenever a new inbound connection is established. This lets you see the site that's just started talking to you. Since some people might find such an unsolicited pop-up irritating, this only happens if you check the new Options menu item "Look Who's Talking". Tested under Windows 95 final build. Works fine when using the standard built-in WINSOCK, but doesn't resolve host names when Sun PC-NFS is overloaded on top of Windows networking. This appears to be a general problem of this configuration; other programs fail on gethostbyname() in precisely the same way. When you configure Windows 95, be sure to install the TCP/IP driver; if you don't you'll get nowhere fast. 12 September 1995 Sending a stereo .WAV file in ADPCM compression mode crashed the Unix speaker program. The code in READWAVE.C which calculates the number of bytes of .WAV file needed to fill a packet was incorrectly assuming the nBlockAlign field was the size of an individual sample, not the frame of samples for all channels. Fixed. Closing a connection while a .WAV file was being sent orphaned the MMIO handle used to read the file. Fixed in CONNECT.C. 13 September 1995 Added the ability to drop saved connection (.SFX) files in the MDI frame window and thereby open (or activate, if already open) connections to the hosts given in the files. You can drop multiple connection files in a multiple selection and each will be opened. CONNECT.C had its own implementation of DragAcceptFiles() which directly twiddled WS_EX_ACCEPTFILES. It doesn't any more. If a connection file is named on the command line when the program is launched, it is opened once the application is initialised. This permits making an association between the .SFX extension and Speak Freely in the File Manager and launching the program for a given connection by double clicking the connection file. You can specify multiple connection files on the command line, space separated. This allows making a program item icon which opens a collection of connections, a handy thing to put in your StartUp folder. (Suggested by John Gilmore (gnu@toad.com)). John also pointed out that the program wasn't usable without a mouse since the left mouse button was the only way to push to talk. I added logic in CONNECT.C that permits the space bar to be used to toggle push to talk, just as in the Unix mike program. You can cycle between open connections with Ctrl+Tab and use the space bar to select any set to which you wish to transmit. Mouseless users who push to talk with the space bar don't have the benefit of the cursor change to indicate which connections are transmitting. I added a "Transmitting" status indicator in the connection window which appears whenever live audio is being sent to the window. If you make a .WAV sound file with the (nonstandard) sampling rate of 8000 samples per second, it is now played correctly by READWAVE.C, not forced to the closest standard sampling rate of 11025 samples per second. Conversion of stereo .WAV files into mono is still performed for 8000 sample per second files. If the user has the ability to make 8000 sample/sec .WAVs, this reduces file size, improves sound quality, and eliminates CPU overhead when sending such files. .AU files remain the fastest, since they're already mu-law encoded. 14 September 1995 Update release 5.1c. 20 September 1995 Began work on answering machine. Defined structure for data in file, added a new ANSWER.C module with a function to save a sound buffer in an answer file in that format. 25 September 1995 Modified the new connection dialogue handler to allow numeric IP addresses which can't be resolved into host names. If the host name lookup fails, the dotted IP number from inet_ntoa is used as the host name. Good ole' Trumpet Winsock returns an error status if gethostname() is called with a buffer too small to hold the entire name, as opposed to truncating it as Unix does. I changed the two calls in CONNECT.C to get the host name in a temporary buffer, then copy as much as will fit into the sendinghost field of the sound buffer. Added the ability to set the multicast scope with a new item in the Options/Connection dialogue. This item is enabled only if the IP address is a valid multicast group number. Bad ole' Windows 95 WINSOCK returns a WSAEFAULT error if you pass a single byte argument for the IP_MULTICAST_TTL setsockopt() call. This is incompatible with all Unix documentation I have seen. Trumpet works correctly with the single byte argument, and accepts the 2 byte short required by Windows 95. Given the likelihood there's some other WINSOCK that requires a one byte argument, in goes another Options/Workaround/Network item: "Multicast TTL Argument Is char" which does it the Unix way, not as required by Windows 95. 26 September 1995 Added a new Connection/Multicast Groups... dialogue which allows adding and dropping membership in multicast groups. Groups can be specified by DNS-resolvable name or by IP address. A check box controls multicast loop-back of locally sent packets to groups in which this host has added membership. The loopback box is disabled on systems (such as Windows 95) which do not implement the IP_MULTICAST_LOOP setsockopt() option. 1 October 1995 Discovered the multicast tear-down code in the WM_DESTROY message handler of FRAME.C wasn't testing for a NULL multiName[], resulting in bad GlobalFree() calls when we failed to initialise a multicast port. Fixed. FRAME.C wasn't killing the main timeout timer at WM_DESTROY. Fixed. If the attempt to drop a multicast membership at WM_DESTROY time failed, a message box was displayed as a child window of the one frame being destroyed. This is apparently (yet another of the billions and billgatesillions) undocumented no-no--in any case, if you do it, you get an "err USER: Attempt to activate destroyed window" at the time the WM_DESTROY returns. I changed the parent of the message box in this case to be NULL and it seems to be happy now. (In FRAME.C). Finished implementation of the answering machine, ANSWER.C. I'll probably be back before long to make it more message-oriented (select message from a list box of sites and times, individually delete messages, etc.) but at least it now has basic functionality. 2 October 1995 Added keyboard accelerator (CTRL+T) for answering machine, and a new connection menu item that lets you toggle whether incoming messages are recorded without having to pop up the answering machine dialogue. Fixed a bug in which checking or unchecking the record incoming messages box in the answering machine dialogue didn't take effect until you closed the dialogue; now it takes effect immediately. Added code to overwrite the 16 byte session key exchanged via PGP before closing the file on disc. Unfortunately, since we can't transmit and receive the with a pipe, as we do on Unix, there's still a window while PGP is running during which the session key is visible, but at least this keeps it from lying around in unallocated disc space for an indeterminate time. If no answering machine message file was configured, the answering machine dialogue in ANSWER.C called scanMessageFile anyway. Unfortunately, that routine didn't test for answerFile being NULL and proceeded to stomp all over memory. Fixed in ANSWER.C scanMessageFile(). Moved all translatable strings and formats from the .C modules into the string table of the resource file, using the rstring(), rfilter() functions and the Format() macro as intermediaries. Strings that aren't to be translated, such as fopen() mode strings, formats that contain only a field editing code, etc. continue to appear as strings in the source code. Banishing these strings to the resource file reclaimed almost 4K of data space, enough to give us some breathing room should it prove necessary to introduce another static full-size sound buffer for some reason. 3 October 1995 The enabling and disabling of buttons in the answering machine was befuddling Windows' dialogue box keyboard accelerator logic. I added code at the end of a message replay to restore the input focus to the button last pressed or its logical successor if that button has become disabled as a result of the message we just completed. Keyboard accelerators in the answering machine were less than optimally chosen due to renaming of buttons during its development. I rationalised them so the most commonly used buttons have the most obvious keyboard shortcuts. Pressing the Close button in the answering machine gave a debug kernel "err: window destroyed in window callback". Why, I know not. It uses the standard code for modeless dialogues right out of Petzold, which identical code works perfectly in the propeller-head modeless dialogue. Changing the DestroyWindow() to a PostMessage of WM_CLOSE to ourself made the message go away. I changed the propeller-head dialogue in DIALOGS.C to use the same logic. Several modal dialogues needlessly included the system menu in their title bar. Eliminated. (The modal dialogues such as the answering machine and propeller-head continue to display the system menu.) Installed help buttons in all the dialogues, linked to the topic in the help file which describes the dialogue. Moved the names of our help file and the base Windows help file into the resource string table. The Create Key menu item was missing the "..." that indicates it pops up a dialogue. Fixed. The About dialogue was also missing the "...". Fixed. I removed the "How to use help" menu item, which has fallen out of fashion. Changed "Help/Search..." to "Help/Search for Help on..." as used in current Microsoft applications. 4 October 1995 Completed moving all section and item titles for the main .INI file and saved connection files to the string table in the resource file. Whether these should be translated isn't clear: a normal user won't ever examine these files and translating renders them incompatible between different language editions. But the saving in data segment size by elminiating duplication of the section titles alone justifies the work. Added two new string constants kS0[1] = "0" and kS1 = "1" to FRAME.C and changed all references to the explicit constants in profile file I/O to use them. This eliminates redundant string constants in the data space. Found a few string constants I'd missed somehow in READWAVE.C. Banished. Fixed the answering machine to update the host name when a definitive name (one not displayed in parentheses) is seen, replacing any previously displayed name. Added help butttons to all the file open dialogues, linked to the appropriate topics in the help file. Added a pleasant default ring file. I haven't found a suitable (well-recorded and public domain) telephone bell, so I decided to pioneer non-irritating notification of an incoming call with this wind chime derived sound. The original appeared on the CD-ROM (N° 5) accompanying "News Windows" N° 26 (octobre 1995) as the file WINDBELL.WAV. I used Silicon Graphics' soundfiler to convert this from an 11025 kHz PCM stereo file to an 8 kHz monaural .AU file for optimal transmission. Substantial data space was being wasted by repeated constant references to the profile (.INI) file name. I moved this string to the string table in the resource file and changed the code that loads and saves the global configuration to load it once into a string on the stack and reference that temporary copy in all the [Read|Write]PrivateProfile... calls that follow. Added a new MAKEBIN.BAT file in the home directory which builds the binary release archive. Now that the release includes more than the .EXE and .HLP file, something more archival than my fallible memory is needed to make sure 5 October 1995 Remade all screen shots for help file, the addition of the help buttons required updating all the dialogue bitmaps. Added logic to the invocations of PGP in FRAME.C and DIALOGS.C to first try to use the SFPGP.PIF file from the Speak Freely release directory (obtained with GetModuleFileName) and then, if that fails, fall back to call on PGP counting on path search to find it. Going through the PIF allows the user to override the default modes for a WinExec call to a DOS program such as PGP, in particular, to run it in a window, which is much less disruptive of the user's equanimity than blasting out to a DOS prompt. 6 October 1995 Feature release 5.3. 30 October 1995 All of the Look Who's Listening functionality is working, at least if you don't push it into reentrancy into Winsock by trying one of the LWL dialogues while sending or receiving sound. I'll have to go back and review the appropriate locks to keep from befuddling Winsock with actual multitasking. Essentially all the code is in the new module lwl.c. 7 November 1995 Added support for RTP-compatible LPC compression (the Xerox PARC algorithm developed by Ron Frederick). This algorithm does a lot of floating point computation (forget it if you don't have a math coprocessor), and it sometimes mangles sound, especially if you drive the audio input into clipping or have a high-pitched voice. But when it works, it achieves better than 12 to 1 compression, and allows running over 9600 baud lines. The LPC code is in a new lpc subdirectory. 13 November 1995 Added a first cut "broadcast" facility to permit transmission of material to multiple hosts (over a suitably fast, probably local network) without the need to install multicast. The facility is relatively crude but should be adequate for uses applications such as broadcasting meetings across a local network. The site performing the broadcast simply checks Connection/Broadcast. Any audio which arrives while Broadcast is checked is sent to every connected host. All input events are ignored in connection windows while a broadcast is in progress, and remotely initiated connections will not time out during a broadcast. A user can subscribe to a broadcast from a given host by initiating a connection to it and sending a short burst of sound (a second's worth, say). This opens a connection on the broadcasting host to which the broadcast will be sent. A remote host can unsubscribe from the broadcast by sending a similar short burst of sound any time after 10 seconds into the broadcast; the site's connection on the broadcasting host will be closed 10 seconds later. The 10 second delay is to prevent toggling of the broadcast state due to multiple packets being received from the remote site. Whilst broadcasting, the application title indicates "- Broadcasting" and the cursor is always the ear when over a connection window. When broadcasting is toggled off, all connection windows are marked as not being transmitted to and remotely-opened connections resume the timeout process. Using short bursts of sound to subscribe and unsubscribe is ugly but it gets the job done. Once we have a proper RTP packet exchange delimiting the connection, it can be replaced. 14 November 1995 If somebody is already blasting sound at us when Speak Freely is launched, it got all befuddled due to packets arriving before initialisation was fully complete. I changed the WM_CREATE logic in FRAME.C to not enable input on the socket until initialisation is entirely done. 16 November 1995 Feature release 5.5 22 November 1995 People seem to get floating divide by zero errors if they try to use LPC compression on Windows 95. I added a call to _control87() in the initialisation code to disable all floating point error interrupts. This should allow the LPC code to just bumble along with infinities and NANs like it does on Unix, which doesn't seem to do any harm (my suspicion is that this happens only when the LPC code is fed dead silence). This will have no effect anywhere else, since floating point is used only in the LPC code. I made GSM compression the default out of the box. I'm deeply weary of explaining how to enable GSM compression to hundreds of people a day who can't be bothered to read the help file. The Phonebook/Search host didn't default to lwl.fourmilab.ch unless you'd previously made a directory listing. Fixed. The Phonebook/Search box wasn't quite wide enough to hold the longest line of a typical server message and didn't have a horizontal scroll bar. This caused perplexed people to send hundreds of E-mails when they couldn't access the truncated URL the LWL server published. I made the box a few characters wider and enable horizontal scrolling. There was also some ragged logic in the default server for publishing directory entries. Fixed to correctly default to lwl.fourmilab.ch. Update release 5.5a. 28 November 1995 Integrated server-side support for the "show your face" feature. The new file FACE.C contains a dialogue that allows the user to designate a 256 colour .BMP file (the format is verified) as his or her face image and a function invoked from FRAME.C that delivers blocks of the image as requested by a remote host. Processing of face image requests occurs before audio output is acquired (but after creating a connection), so half-duplex systems can still transfer face images while sending audio. 30 November 1995 Integrated client-side handler for "show your face". Face data packets are assembled in FACE.C into an in-memory bitmap in the connection structure. If a complete bitmap is available, the WM_PAINT handler in CONNECT.C for the connection window displays the bitmap instead of the usual status information. When a bitmap is displayed in the connection window, transmit state is indicated by preceding the host name with a small ASCII-art arrow. After adding hundreds of lines of bullshit Windows code trying to swap the palette intelligently when two face images are simultaneously displayed on a colour-mapped display, I decided to exercise that time-proven prerogative of the Windows developer and just give up. The vast majority of users won't connect to more than one person at a time. I fixed it (with even more bullshit code) so that the active window is always shown with the correct palette and inactive windows with the default palette. If you this this is easy to fix, baby, just go and try it before you write me some smart-ass E-mail. Hint: everything you read in the Windows API documentation about palettes is a lie or worse when you start to talk about MDI child windows. There's a sample application on the Developer CD which claims to do this, but a glance at it leads me to estimate at least a week to integrate and test all the crap they went through trying to do what, on any vaguely competently designed windowing system, should be essentially transparent to the application. Users of high-colour and true-colour display boards will be blithely unaware of any problem with multiple simultaneous face images. 1 December 1995 Mycal (mycal@monitor.net) reported that messages using simple (2X) compression were played back at twice normal speed by the answering machine. Fixed. Feature release 5.6. 20 December 1995 Added logic in CONNECT.C to set outputSocketBusy if the send() or sendto() returns less than the number of bytes we attempted to write. The WSAEWOULDBLOCK error status is treated as a truncated buffer and sets outputSocketBusy. All of this is disabled if the new workaround "Disable Output Overflow Recovery" is set just, as always, in case. When outputSocketBusy is set, we're guaranteed by the Winsock spec that we'll receive a the FD_WRITE notification we requested in the call on WSAAsyncSelect for the socket. Right. So in the timer, should we discover the socket has become unblocked for output and the fink didn't tell us, clear outputSocketBusy so things don't hang up. To avoid output overruns, I changed the logic that responds to face image requests to ignore requests received while audio output is active. This should keep face data from pushing an 14.4 modem connection that is barely keeping up with GSM over the edge. The transmission of the face will be resumed by the timeout on the receiving end when the audio transmission is done with no harm done. I also tweaked the timeouts so they're less likely to collide with one another. 23 December 1995 Face bitmap exchange seems to hold the potential for bad medicine when stirred in the pot with multicasting. When sending to a multicast port, CONNECT.C now never offers a face to the subscribers, and face requests received from multicast ports (shouldn't happen, but who knows: it's Windows!) are ignored and the face retrieval status set to Abandoned. Strengthened the .BMP file format verification in FACE.C. We now verify that the bitmap has one plane and <= 256 colours, as must be the case for any compliant bitmap. Added more stringent verification of the format of received bitmaps in FACE.C before passing them on to CONNECT.C to display. This gives us some protection against rogues who let bogus bitmaps through, and weird errors in transmission of the the bitmap that corrupt it. 31 December 1996 Added an explicit +armor=off to all invocations of PGP to guarantee the session key is encrypted in binary mode even if the user has modified the PGP configuration file to make ASCII armour the default. 16 January 1996 Integrated the VOX, Break-In, VOX GSM compression, and anti-hangup code from Dave Hawkes (daveh@cadlink.co.uk). In the process of testing the integrated version, I made the following (perhaps temporary) changes: * For some reason, the fix that places the read-only data in ULAW.C in the code segment causes Speak Freely to crash with a GPF on the first reference to the tables in CONNECT.C, but only when I compile in DEBUG mode. I also needed to include the definition of the new tag CONST_DATA to be able to recompile the ADPCM and LPC libraries, which contain references to the ULAW.C tables. As a stopgap, to get things running, I changed the definition of CONST_DATA back to FAR. I'll look at this in more detail when I get a chance and see if I can't get that data back in the code segment. * The anti-hangup code which calls DefaultMessageLoop() every time a network packet or wave audio input buffer arrives causes unacceptable break-ups of sound on my 486/50; apparently PeekMessage takes too long even in the normal case. I modified MessageLoop() and DefaultMessageLoop() in NETFONE.C to save the time (GetTickCount()) of each pass through the message loop and only run the PeekMessage loop if 10 milliseconds or more (constrained, of course, by the fundamental resolution of the timer) have elapsed. I'm hoping this will run the risk of the delay only when an actual message loop backlog occurs, which runs the graver risk of a lock-up. I've verified that the PeekMessage loop only rarely runs on my machine (usually when I block the window by moving it or pulling down menus), but since I still cannot reproduce the actual lock-up, even when I run my machine at 25 MHz, I can't verify the other side of the equation: whether the lock up is still avoided. * I changed the "VOX" menu item to "Voice Activation" to avoid jargon which might befuddle the radio-naïve. * Changed the title of the "VOX" monitor dialogue to "VOX Monitor". The dialogue isn't wide enough to avoid the "VOX" abbreviation, but adding "Monitor" makes it look a little less stark. It might make sense to go to a horizontal meter to justify more room for a longer title. Some people would like to be able to launch Speak Freely from another application, pointing it a given host with various preset options on the connection. Writing an .SFX file naming the host and specifying the options, then invoking SPEAKFRE.EXE with the .SFX file on the command line permits this, but the .SFX file had to specify both the host name and IP address, forcing the calling application to look up the host name. I modified the newConnection() code in FRAME.C to automatically look up the IP address if an .SFX file specifies only a host name. A typo in the Connection/Save code could have led to nugatory void entries in the .SFX file. Fixed. 17 January 1996 Added a visual indication when VOX is squelching transmission. The ear cursor now changes to an ear with a big X through it (that's the best I can think of right now, but it's better than no indication at all). This makes it a lot easier to evaluate the effect of VOX speed, especially if you don't have a local machine to run tests. Figured what was wrong with the ULAW.C data in the code segment trick. Apparently CONNECT.C and the libraries didn't get recompiled after the definition was changed to CONST_DATA, and continued to reference the Ulaw tables as FAR (as they must). I changed ULAW.C to explicitly place the tables in the code segment, but continue to reference them with a FAR declaration in ULAW.H. 19 January 1996 As suggested by Enoch Wexler (wexler@datasrv.co.il), I added the ability to send and receive Show Your Face images in GIF as well as BMP format. GIFs offer substantial compression compared to even the compressed variants of BMP, which reduces the time it takes to transfer a face image and the likelihood of disrupting audio transmission in the process. Received GIF files are converted in-memory to BMP format by the new module GIFTOBMP which is based on the NETPBM utility GIFTOPNM. GIFTOPNM is copyright 1990, 1991, 1993, by David Koblas (koblas@netcom.com), who notes: Permission to use, copy, modify, and distribute this software and its documentation for any purpose and without fee is hereby granted, provided that the above copyright notice appear in all copies and that both that copyright notice and this permission notice appear in supporting documentation. This software is provided "as is" without express or implied warranty. GIF file decompression requires substantial storage for the LZW decompression buffers and colour map tables. I modified the decompression code to move all large buffers to a dynamically allocated global storage block to avoid overlow of DGROUP and/or the inclusion of static storage which would block execution of multiple instances. The face image selection dialogue in FACE.C was modified to allow selection of GIF images as well as BMP files. Added a new VOX menu item which calls a new function in VOX.C, vox_reset_parameters() to restore all the VOX level adjustment parameters to their original defaults. I also added a Reset button to the VOX Monitor dialogue which does the same thing. In the process, I rearranged the contents of the Monitor dialogue to create enough room to spell out its title. Added a new workaround that totally disables the DefaultMessageLoop lockup-prevention mechanism. While I think my 10 millisecond trigger based on GetTickCount() should be enough, this provides an escape hatch in case it isn't. 23 January 1996 Based on reports that receiving a Ring message can screw up the sound card (for example, muting the microphone), I demoted the previous default call on waveOutSetVolume() in SPEAKER.C which attempts to set maximum output volume when a ring is received to a Workaround which is off by default, "Set Maximum Volume on Ring". Only in the world of Windows would you suppose that something as innocent as a volume control would conceal sharp edges and booby traps awaiting the unsuspecting developer. Based on input from Dave Hawkes, I revised the DefaultMessageLoop code once again. This time it keeps track of the last time the program potentially yielded control to another application (by doing a PeekMessage with the PM_NOYIELD and PM_NOREMOVE flags in the main message loop in NETFONE.C, which seems to return fast enough to avoid pauses, and saving the GetTickCount() value if there is no message in the queue). Then, if DefaultMessageLoop() discovers 350 milliseconds or more have elapsed since the last yield, flushes the message queue using PeekMessage(), which will allow other applications to gain control of the CPU. Dave also pointed out that my pointy-headed code that changes the cursor when VOX muting occurs changed the cursor even if it was outside the window. Fixed. 25 January 1996 The workaround that disables the DefaultMessageLoop() insurance did not actually turn off all traces of the code--the PeekMessage in the main message loop in NETFONE.C still remained. Since I'm sure this will cause unspeakable horrors when it triggers some booby trap Billy-boy has hidden in one of his existing products or is in store for us in the future, I made sure it's disabled when DefaultMessageLoop() is turned off. 6 February 1996 Made the inclusion of encryption conditional on the tag CRYPTO being defined in NETFONE.H. If CRYPTO is not defined, the version number in the About dialogue will have a suffix of " (no crypto)" and the IDEA patent notice will be replaced by an explanation of where to obtain a version including full encryption. The "Encryption" box in the Options/Connection dialogue will contain a more detailed explanation of the no-crypto edition. The Options/Create Key menu item is disabled in no-crypto builds. What's the rationale for this? Simple: a number of CD-ROM publishers and sound card manufacturers are interested in distributing Speak Freely. But since many will be shipping from the U.S. and other countries which attempt to restrict the export of "munitions" like Speak Freely they're afraid, and rightly so, that putting Speak Freely in the box might result in an all-expenses paid extended vacation at Club Fed. The non-crypto version allows them to include Speak Freely without such worries. Once a user has installed Speak Freely and is ready to start using encryption, they can simply follow the instructions in the dialogue boxes (and, soon, the help file) and download a full-encryption version from a site in a country which does not restrict cryptographic software. The non-crypto version can also be posted on bulletin board and commercial online services without risking government-initiated unpleasantness. Note that undefining CRYPTO does not just block access to encryption and decryption; it totally removes the code from the program--the encryption libraries are never referenced and therefore not included by the linker in the executable. Thus there is no risk of a non-crypto build being deemed a munition. The only "cryppish" code that remains is MD5, and it is widely used (for example, in the export edition of Netscape) in non-encryption roles. In non-crypto Speak Freely, the RTP SSRC, timestamp, and packet sequence numbers are generated directly with MD5 rather than the somewhat more random idearand() used in crypto builds. Since we're not going to encrypt the RTP packets anyway, this doesn't compromise anything. As the first step in integrating the RTP support code from the Unix version, replaced RTPACKET.C with the fully-functional one and verified the new rtcp_make_sdes() and rtcp_make_bye() didn't break LWL support. Here's where we stand at the end of first day of the campaign to integrate RTP and VAT support into Speak Freely for Windows. The two procotol translation modules, RTPACKET.C and VATPKT.C and all their support files have been included in the program and fixed to compile without errors or warnings. As noted above, the new RTPACKET.C continues to generate valid packets for the LWL server. Sending both VAT and RTP protocol works for all compression modes, testing with VAT on another machine. VAT correctly recognises the VAT ID and RTCP SDES message we send on the control channel. 7 February 1996 Trying to integrate the LIBDES encryption package need for VAT and RTP encryption blew the data segment, so it's time to run another sweep for excess baryon particles. Using the same CONSTANT_DATA trick as in ULAW.H, I moved the large constant tables in LIBDES\FCRYPT.C and DES\DES.C into the code segment. Result: still over the brink. The biggest memory hog, LPC\LPC.C, was unfortunately not so easily fixed, since its four large floating point analysis vectors are read/write and cannot be hidden in the code segment. I integrated a modified version of the LPC.C from NeVoT, in which the state of the decoder is in a dynamically allocated buffer. I obtain this buffer with GlobalAlloc, getting it out of the static data segment. This of course required FARs all over the place in LPC\LPC.C, but it did the trick. This will also allow, if we decide it's worth doing, maintaining a separate LPC state for each inbound connection. It was intensely irritating to have to constantly answer E-mail from people who tried to build Speak Freely from source code but whose Winsocks not only didn't support multicast, their WINSOCK.H didn't even include the definitions for multicast. I fixed FRAME.C, DIALOGS.C, and NETFONE.H to, if IP_MAX_MEMBERSHIPS isn't defined, silently delete multicast from the build. If the user's Winsock doesn't define the variables we need to generate the code, there's no way he's going to be able to use the feature anyway. The Unix version uses the same trick to adapt to pre-multicast sockets implementations. Swept through the program and added the fProtocol flag to all places Speak Freely protocol packets are generated. This flag helps receivers distinguish Speak Freely packets from VAT and RTP messages. Went a long way toward implementing DES encryption of outbound RTP and VAT packets. It pretty much works--I'll make the final round of tests when DES works in both directions and I can verify correct operation in both directions. 8 February 1996 I modified code in CONNECT.C and FRAME.C to zero the SDES resend timer when the transmit protocol or encryption key is changed. This causes an immediate resend of the SDES/VAT ID in the new mode, which will help the receiver to "sync up" with the change. I discovered that the clever way I integrated VAT and RTP encryption into sendpkt() in CONNECT.C had completely screwed up encryption for Speak Freely protocol. In the process of fixing this, I cleaned up some of the rather tangled logic in that function. Added code to the Options/Connection dialogue to disable the fields and captions for IDEA, PGP, and Key file compression if the protocol is RTP or VAT. These protocols currently specify only DES as a standard mode. Set the "talk spurt" flag for the first packet of a sound file. Added VAT packet translation to sound file output in CONNECT.C. Under certain circumstances, sending a sound file to a connection after sending a ring would just re-send the ring. Fixed in FRAME.C. Disabled direct modem connections. This feature, which fell into the trapdoor called Windows serial port (8250) support, compounded the incoming E-mail pain due to idiots confusing direct dial-up modem connections with SLIP/PPP Internet access. Besides, since Serial I/O is near the top of the Redmond Kiddies' "API of the Year" list, the time it takes to debug it on all the ratty drivers out there exceeds the product life cycle. Users who wish to use Speak Freely as a phone scrambler on direct calls should establish a peer-to-peer TCP/IP connection and use Speak Freely in network mode. Since it will probably take Billy's bozos 20 or 30 years to debug Windows to Windows TCP/IP links, the fact that they'll blame their screwups on other vulnerable applications as well as Speak Freely will deflect a significant percentage of flames, albeit not to the flamers responsible for the mess in the first place. 9 February 1996 Voice activation didn't work with RTP and VAT protocols because load_vox_type_params() in VOX.C didn't know about the new packet sizes used by those protocols. Now it does. Independently, LPC and VOX don't seem to be getting along very well together, regardless of protocol. I'll have to look into this later on. Coming to terms with the fact that I'll be chasing "bugs" in this program as long as there are Kode Kiddies in Redmond, I integrated the hex dump module from the Unix version. It lets me dump packets on the debug stream to see where Winsock wants to go today. Transmission of non-encrypted VAT and RTP packets now seems to be working. That's not to say that DES encryption doesn't work, just that I haven't tested it yet. The initial tests of bouncing VAT messages off the echo server failed due to byte order dependencies in VATPKT.C. These are now fixed; the changes must now be integrated into the Unix stream to handle little-endian boxes. 10 February 1996 Added code to the WM_DESTROY handler in CONNECT.C to transmit an RTP or VAT BYE message to indicate the user has closed the connection. In the process, I added a new sendSessionControl() function which is used by both this logic and the periodic RTCP/VAT ID transmission code in the timer. Discovered that the sockets were getting closed in WM_CLOSE rather than WM_DESTROY, which kept the BYE transmission from working. I moved the socket close to after the BYE is sent in WM_CLOSE. 11 February 1996 Well, I think I finally found out why weird things occasionally happened when you quit Speak Freely with one or more connection windows open. This is really getting too depressing to document, but here we go. Windows, with its unerring instinct for doing things in the most idiotic way possible, sends a WM_DESTROY to the application's outer window procedure, and then later sends WM_DESTROY to each of the child windows. Suppose one or more of those child windows need to do something--send a BYE message, for example--using one of the resources that get freed by the application's outermost WM_DESTROY? Blooie. But don't think you can get away with just sending WM_DESTROY to each of the child windows: nopey, nopey, no. If you try that you fall into the toilet because calling WM_DESTROY doesn't actually make them go away, and as a result they get destroyed twice and all kinds of other horrors ensue. So, once again we are forced into subterfuge by the quintessential inelegance of Windows. We dig out of this particular hole by sending a custom WM_CLEAN_UP_YOUR_ACT message to the child window to tell it we're terminating. The child will then do its regular WM_DESTROY cleanup, including releasing the client data pointer and zeroing the window word to it. When the actual WM_DESTROY arrives, it will discover the client data pointer is zero and avoid executing the cleanup twice. 12 February 1996 Integrated the new LWL\LWL.C library from the Unix version, which allows a separate decoder state for each receiver and contains numerous fixes for subtle coder gotchas such as dividing by zero if total silence is received. This, of course, ran squarely into one of the innumerable floating point code generation bugs in Visual C++ 1.5 (they call it "Visual" because only if you're seeing things would you confuse it for a production compiler for floating-point intensive code). After a fine afternoon of trying various compiler options and workarounds, I found a combination of restructuring of the loop which caused the "Stack overfl" (a very Redmond kind of error message, don't you think?) and optimisation options which got around the error. Until the next "improvement" of the compiler, I'm sure. Found another pair of missing htons() in the LPC packet handlers in VATPKT.C and RTPACKET.C, when stuffing the decoded length into the first two bytes of the sound buffer. Memo to file. Windows' real-time response is so pitiful that machines which are perfectly adequate to run Speak Freely protocol in all compression and encryption modes (a 486/50, for example) can't cope with the smaller packet sizes used by RTP and VAT, particularly on reception. If you want to talk to somebody who can only sent RTP or VAT, you'd better make sure you have enough Intel inside to cope with Billy Boy's idea of process switch latency. 13 February 1996 Integrated a fix to rtpout() in RTPACKET.C from the Unix version. The packet length for outbound RTP ADPCM packets was 2 bytes short, which caused "gravelly" speech and horrible ticking when encryption was enabled. It's possible for the predicted value in the ADPCM coder (ADPCM/ADPCM_U.C) to exceed the range of a signed 16 bit linear sample. Clamping code limits the range when this happens, but needed to declare the unclamped sample as a long rather than int to work on a 16 bit architecture. Fixed. After many, many hours of painful, unremunerated toil I finally figured out what was causing the Debug kernel warnings and fatal errors due to bad pointers at the time we call WSACleanup at application termination time. The essential clue was that it only happens if one or more connection windows have been created by the receipt of a packet from a remote site not already connected. If all connections were created locally it never happened. And, of course, this only happened under the Winsock supplied with Sun PC-NFS 5.1--the problem never occurred under any circumstances on other Winsocks. I'm sure by now you will be shocked and stunned to learn that Sun NFS 5.1 doesn't correctly implement the WSAAsyncGetHostByAddr function. Oh, you can make the call, and you even get back a valid host name. But doing so plants a time bomb which will kill you (at least under the debug kernel) much, much later when you call WSACleanup() right before exiting the program. At that time, depending on where the random pointer inside their so-called WSHELPER points, you get either two invalid global pointer errors or a fatal error due to an object usage count underflow in the (bogus) global block. If the waNetSynchronousGetHostnameAction is set, we eschew the asynchronous request and make the user wait for a blocking gethostbyaddr() which has the merit, at least, of not blowing us away at program termination time. waNetSynchronousGetHostnameAction is set, in turn, based on the workaround waNetSynchronousGetHostname, which can take on the values 0, 1, and 2. If 2, the default, asynchronous host name retrieval is disabled if the Winsock identifies itself in the szDescription field of the WSAData structure returned by WSAStartup() as "Sun Select PC-NFS Windows Sockets Implementation". This automatic selection can be overridden by the user explicitly checking or unchecking the Options / Workarounds / Network / Get &Host Name Synchronously menu item. Thereafter, the user's selection will be used regardless of the identity the Winsock reports. In the process of adding profile variable support for the above workaround, I observed the number of workarounds was about to exceed the rstring() cache in the name of the workarounds section remained. Since, unlike the developers of Microsoft tools, I do not feed off human suffering and take joy in setting booby traps, I modified all the profile read and write code to copy the section name to a stack string variable rather than rely on the pointer within rstring()'s retrieval area to remain valid while all variables in the section are accessed. 14 February 1996 Integrated the new RTPACKET.C, RTPACKET.H, and DESKEY.C from the Unix version. These include the facilities we'll need for parsing SDES packets, recognising BYEs, and creating RTP keys from key strings compliant with RFC 1890. Adjusted packet sizes returned by inputSampleCount() (FRAME.C) for VAT to the maximum permitted within both the experience base of VAT and the 512 byte guaranteed MTU of Winsock. Integrated a fix from the Unix version to guarantee (in our context) that pad bytes added to VAT and RTP packets are zeroed. Modified makeVATid() in VATPKT.C to, as VAT does, prefer the user's full name to the E-mail address if both are available. Added recognition of RTP and VAT SDES/ID packets in FRAME.C. The title of the connection window will now show the user's name, if supplied by the sender. The changeAudioState() function in CONNECT.C also uses the user name, if available, in preference to the host name when it updates the window title to indicate transmit state. RTP and VAT BYE/DONE packets now cause the receive protocol to be reset to PROTOCOL_UNKNOWN. This expedites recognition of a new protocol if the sender switches on the fly. Changed in FRAME.C. Integrated generation of RFC 1890 RTP key and separate old-protocol VAT key. I've still to integrate automatic protocol and key sensing. Added code to encrypt outbound RTP and VAT packets with the appropriate key. The inbound side remains to be done. Nailed another encryption packet size rounding error in CONNECT.C, this time affecting ADPCM encoded outbound packets in VAT protocol. 15 February 1995 I remembered that there was one more place the PC-NFS 5.1 WSAAsyncGetHostByAddr() bug could stab us in the back--in the case where the user enters a numeric IP address in the Connection/New dialogue. I added a gethostbyaddr() alternative to this call if waNetSynchronousGetHostnameAction is set. Finished integrating automatic protocol sensing and key selection for encrypted inbound RTP and VAT packets. The logic was a little less tangled than in the Unix version since we process control and data packets in different callback functions. Did a non-CRYPTO build to make sure all the RTP and VAT changes didn't break something or suck in verboten bits. Sure enough, all of DESKEY.C needed #ifdef CRYPTO, as well as the encryption code in CONNECT.C's sendSessionControl(). Fixed. Implemented the guts of the local loopback facility--it works, but tuning and a nice user interface remain to be done. Why local loopback? So users can debug their audio hardware problems before venturing onto the net, which will be one of the steps in the "beginner's guide to Speak Freely" I'll get around to writing one of these days. Eventually there will be a Help menu item which creates a local echo connection, but for the moment you activate such a connection by making a new connection to "localhost" (or, if your Winsock doesn't know localhost from Casper the Friendly Ghost, 127.0.0.1). Any packets you send are saved in memory until the end of your transmission and then returned, after a short delay, as if echoed by a remote site. This is, then, an echo server that doesn't use the network. Not only does it let users experiment with audio hardware locally, it allows isolating network-induced problems from those which inhere in the CPU or audio hardware. 16 February 1996 Modified changeAudioState to invalidate the connection window without erasing the background. This makes it quicker to repaint the "Transmitting" or blank status when the audio state changes. Added support for Speak Freely SDES messages on the control port. When a Speak Freely connection is open, RTCP SDES messages with a protocol ID of 1 (the old RTP, used by no application I know of) are sent on the control port. These messages allow unambiguous recognition of Speak Freely protocol, transfer of user information, and disconnect notification. The connection window paint code in CONNECT.C now displays the current sending protocol, user name, and E-mail address of the connected user. The connection window is resized depending on the number of lines currently displayed. 17 February 1996 Added the ability to specify a port number for a connection. This required changes all over the place: * A port number can now be entered in the Connection/New dialogue after the host name or IP number, delimited by a slash. * The port number (even if standard) is saved in an .SFX file by Connection/Save / Save As. * A port number, if specified, is restored when a connection file is loaded. If no port number appears in the file, the default of 2074 is used. * The "Connect" button in the Look Who's Listening dialogue now passes both the IP address and port number, separated by a slash, as the known host argment to newConnection in FRAME.C, which was modified to recognise that syntax. * FRAME.C now maintains a list of auxiliary receive sockets, asList. When input arrives, a new function, findPort() searches the list to identify, from the socket number, which port the input arrived from. This is used, if we're creating a new connection based on the input, to set the port to which we'll respond. * CONNECT.C now uses the "port" field in the connection structure as the port to which messages are sent rather than the canned value of 2074. * findClientByHost() in FRAME.C now considers two connections identical only if both the IP address and port numbers are identical. * When creating a new connection with a nonstandard port number, CONNECT.C calls the new function monitorPort() which creates an auxiliary socket pair for that port. If the port is already monitored, the reference count on the auxiliary socket is simply incremented. * When destroying a connection to a nonstandard port, CONNECT.C decrements the reference count on the auxiliary socket and if it's zero closes the socket pair. Note that auxiliary sockets are not bound to a specific host; once a connection is established with a given port, connections from any host can be remotely initiated on that port. This means that if you want to accept connections on a given port as a matter of course, you can do so simply by opening a dummy connection (to a nonexistent address on your subnet, for example) with that port. I'll probably eventually add a separate dialogue that lets you specify ports to monitor automatically but for the moment this gets the job done. Most users will be specifying ports to connect to remote RTP and VAT conferences anyway, not accepting calls on nonstandard ports. Guess what? Sun PC-NFS 5.1 Winsock will not only blow you away if you call WSAAsyncGetHostByAddr(), the blocking version, gethostbyaddr() has a bug in it as well--it forgets to null-terminate the host name in the h_name field, so if you retrieve a host with a name shorter than the last one part of the last host name still sticks out. Working around this by zeroing the host name after you retrieve it is a blatant violation of the Winsock spec which states (section 4.2.1) "The application must never attempt to modify this structure or to free any of its components.". I for one, am not going to add my name to the list of millions who ignore the Winsock spec, so there isn't a damned thing I can do this other than tell people to get a better Winsock. Fortunately, it's purely an ugliness that doesn't do any damage since we're just retrieving the host name to display in the connection window. Oops! In Speak Freely protocol, control channel messages aren't supposed to be encrypted but they were. Fixed. What the world needs now, is lots more workarounds, they're the only thing that drive the bugs to ground.... So, some more anticipatory retaliation: the following are available on the new Options/Workarounds/Protocol submenu. All are, of course, saved in the .INI file and otherwise treated as respectable citizens. No Speak Freely Heartbeat Disable the periodic Speak Freely protocol heartbeat on the control channel. This is primarily intended as a last resort if the (less than 1%) added bandwidth saturates a close to the edge connection, and also in case the control channel packets awake something horrid lurking on the next higher channel. Large RTP Protocol Packets Uses Speak Freely's preferred packet sizes for GSM and LPC compression rather than those typically sent by RTP programs. Most RTP programs were developed on fast workstations with high bandwidth network connectivity. Speak Freely users generally have slower machines and network links which benefit from larger packets. Try this if the person you're talking to reports halting audio in RTP protocol. Disable VAT Protocol Detection VAT protocol will never be automatically selected as a result of receiving a message on the control channel which resembles a VAT control message. Enable this if you never receive VAT protocol messages and are annoyed at how long it takes to identify the protocol of encrypted RTP messages. Disable RTP Protocol Detection RTP protocol will never be automatically selected as a result of receiving a message on the control channel which resembles a RTP control message. Enable this if you never receive RTP protocol messages and are annoyed at how long it takes to identify the protocol of encrypted VAT messages. No Encryption of RTP Control Packets RTP control packets can, according to the standard, be sent either encrypted or in the clear. Most RTP programs I've encountered encrypt their control packets, so this is the default Speak Freely sends (it accepts both encrypted and clear packets). If you set this workaround, control packets are sent in the clear. 19 February 1996 After further deliberations, I decided not to implement automatic protocol switching to the protocol received from the active window, although much of the infrastructure to do so is in place. The reason is that simply adding the new dimension of multiple protocols has the potential for inducing further confusion among users who don't understand the distinction between the compression mode received and that used in transmission. Trying to explain all the possible conditions one could get into with automatic protocol switching is probably futile. I may eventually put in a warning that pops up if the user tries to transmit to a connection which has sent us packets in a different protocol than the current transmit protocol. Protocol mismatch is never a problem when communicating with other copies of Speak Freely, since it auto-senses the protocol. Since initially relatively few users will be talking to other programs, those cutting-edge users are probably best encouraged to operate in "manual transmission" mode to avoid confusion. The gimmick that forces immediate transmission of the identity message on the control channel (rather than waiting for the next timer interval) wasn't doing so when the protocol is set to Speak Freely. Fixed in FRAME.C. Drat! When I added the port number criterion to decide whether a connection was already open, I forgot to handle local loopback. Fixed. Added direct pointers from the Help menu to the FAQ and Mailing List sections of the Help file, and to create local loopback connection directly. Direct access to local loopback required a tweak in CONNECT.C to not attempt to turn the loopback IP address into a host name. 20 February 1996 sendSessionControl() in CONNECT.C wasn't incrementing packetsSent for the extended status dialogue. Fixed. loop_sendto() in LOOPBACK.C wasn't returning SOCKET_ERROR and setting the last error code as it should if it can't allocate the loopback buffer. Fixed. To eliminate the choppiness that afflicted local loopback, particularly with the small packets sent by VAT protocol, I modified loopback replay to adopt a strategy of keeping the output queue stuffed with packets up to a limit of 10, and refilling the queue to that length every 10 milliseconds (Hah!! More like when Windows gets around to us.) rather than attempting to time each packet to a time resolution Windows just can't handle. (If I used the multimedia timer, it probably could, but that requires interrupt code call-backs into a DLL and imposes restrictions on what we can do from the call-back that Speak Freely couldn't live with.) Naturally, this straightforward approach walked squarely into the jaws of disaster. The message loop insurance code was causing the timer code to be re-entered while it was playing back loopback packets, setting off a spectacular riot of recursion. I disabled the message loop insurance for loopback packet playback. Since we're strictly controlling the rate of packet arrival from loopback and the length of the stream is limited anyway, we don't really need the message loop insurance, which is intended to keep packets arriving from the network from hanging us. 21 February 1996 I added some code to the MM_WIM_DATA message handler in FRAME.C to soften the impact of the anti-lockup code on outbound audio quality. If a machine is right on the ragged edge of being able to compress in real time (a 486/50 sending GSM, for example), occasional Windows-induced delays will trigger the anti-lockup code and cause a sound packet to be dropped. I added code that allows recovery from one re-entry to the message handler by saving the packet and processing it immediately after the already-underway packet. Re-entries while an already saved packet awaits processing continue to discard packets. The anti-lockup code in MM_WIM_DATA and socketInput did not increment the appropriate PacketLost counters. Fixed. Added an item to the Help menu that points people directly to the echo server topic in the help file. Building on Dave Hawkes' insight that the Windows wave audio output pause and restart could be used to implement a de-jittering replay delay with no buffering logic within Speak Freely, I implemented a first cut at de-jittering. Any input packet from the network which causes us to acquire audio output is considered the start of a "talk spurt". (Once we've transitioned to RTP, we can use the packet header bit for this, but we have to get there somehow.) When such a packet is received, if the jitterBuf has a nonzero replay delay in milliseconds, a timer with that expiration is launched to trigger the replay and wave audio output is paused. Wave audio output is restarted when the timer expires, or if the number of packets queued for replay exceeeds half the number of messages in the input queue (detected in SPEAKER.C). A new Options/Jitter Compensation menu item allows specifying the initial delay for a talk spurt. The longer the delay, the greater the suppression of jitter, but at the cost of a greater time parallax between the reception of the packet and its being played on the speaker. When shutting down audio input, the final partial packet of sound before the shutdown could be lost. Fixing this little buglet naturally required massive changes to how audio input is torn down, since Windows likes to return packets from the queue in any old order at shutdown time, but imposes on the application a rigid order in which the API must be called. The terminateWaveInput() function in FRAME.C now actually does no such thing. In fact, it just resets audio input, causing any partial packet and the rest of the input queue to be returned to the message loop. Code for the MM_WIM_DATA message now processes any partial packets, padding them if necessary to the length prescribed by the protocol (with the correct pad depending on whether audio is 8 or 16 bits--gosh this is fun!) and sends any non-zero-length packets. If termination is underway, packets are unprepared and released, and an allocated packet counter decremented. When that counter goes to zero, wave audio is finally actually closed. The above fix of course broke how Options/Break Input manages the transition between input and output mode for half-duplex audio hardware. Fixed (I think, pending reports to the contrary from the field). 22 February 1995 One more tweak to Break Input--if inputPaused is set, the WM_MIM_DATA handler in FRAME.C now immediately discards any partial packets that are returned during input termination. This speeds up the transition to playing the packets arriving from the socket. 26 February 1996 Port numbers greater than 32767 were not accepted in Connection/New due to being scanned as a signed short rather than unsigned. Fixed. The supposedly private bits used by the answering machine to mark the start of a transmission conflicted with the fProtocol flag bit, resulting in each Speak Freely protocol packet being considered the start of a separate message by the answering machine. Fixed. 29 February 1996 Dodging another intracardial dagger from our south-of-the-equator purveyors of what purport to be WINSOCK drivers introduces another layer of unnecessary and unwarranted complexity. The WINSOCK spec allows mutant windsuckers to abort any call that "re-enters" WINSOCK with a WSAEINPROGRESS call. Fine: what would you expect from the "vision of the future is a reboot in the face forever" people? But could you imagine, even in your wildest fantasy, that the most innocent of all socket calls, the one which places a socket in non-blocking mode, could itself blow off if a so-called blocking call (and many calls so-deemed have no non-blocking variants) is in progress? So, I turned the code that sends the heartbeat to the Look Who's Listening server inside out to cope with this crap, and thereby avoid the dreaded "Operation already in progress" puke-o-rama which, on some WINSOCKs poisons all future network accesses. I am sure this will have ugly consequences on other buggy platforms which will become apparent in the weeks and months to come. Failure to look up the host name corresponding to an IP address after a connection was made displayed an error dialogue. Little did I know that 95% of all Windows 95 users do not have a valid domain name server configured, and each and every one of them E-mail me when this message appears. Warning message deleted; don't keep those cards and letters coming. 1 March 1996 Integrated the changes to VATPKT.C from the Unix version to recognise IDLIST packets as valid to to provide, in the future, the ability to include a conference ID in the packets we send. Integrated the fixes from the Unix version into FRAME.C to recognise VAT IDLIST (3) packets and correctly parse the user names therein. This allows us to connect in VAT protocol to CU-SeeMe reflectors. As part of this change, the user name field (uname) in the connection structure was made a dynamically allocated buffer to permit long lists of participants in a conference. I modified the connection creation logic in controlInput() in FRAME.C to never create a new VAT protocol connection unless an ID (1) message is received. This keeps IDLISTS (3) which rain in from conferences you've just left from re-opening the conference connection. Also, it was inelegant to open a connection based on a VAT BYE from the blue. A remote VAT connection will be opened only upon the receipt of an unencrypted ID packet. Note that we still have a problem with VAT conference Lazarus connections which result from VAT packets which arrive on the data port and are mistaken for pre-6.0 Speak Freely sound packets--they won't be played, but they'll still open the connection. This will go away once we've gotten everybody on 6.0 and require control port session control unless a special workaround is set to communicate with older versions. Added logic in the connection window WM_PAINT handler in CONNECT.C to distinguish a VAT IDLIST (and one of these days, a multi-party RTCP SDES) from a simple ID and list the users in the conference one one per line. After an afternoon of flailing around that sacrified the requisite number of neurons, I finally figured out a way to handle a user quitting the program while audio output is active which does not lead to sudden death. This involves the usual flags, mushy timers, countdowns and activity tests which Windows requires to do even the simplest things, and is much too depressing to discuss here. If you must see it, start at the WM_CLOSE handler in FRAME.C and follow the trail of slime through the rest of that file. Echo, voice on demand, and reflector servers all have a tendency to create "Lazarus connections" which, seconds after you close them, pop back into existence when a packet comes back from the other end. To prevent this, I added a special anti-Lazarus mechanism in FRAME.C and CONNECT.C. When the user closes a connection window, the WM_CLEAN_UP_YOUR_ACT message handler in CONNECT.C saves the IP address of the host in a new global Lazarus and sets the timeout counter LazarusLong to LazarusLength (15 seconds as presently configured), which is decremented by the main one-second timer in FRAME.C. If a packet arrives from the last connection window to be closed while LazarusLong has not yet counted down to zero, is it discarded by code in socketInput and controlInput in FRAME.C before it causes the connection to be reestablished. This provides a "decent interval" to allow postmortem packets arriving from the remote host to be discarded without re-opening the connection window. To avoid the ignominy of shipping a release containing OutputDebugString diagnostic output, I added an updated version of the check for debug output in a production build that's used in Home Planet. The new version provides a primate-readable description of the error that points to the offending line and doesn't interfere with compilation of the rest of the file. 2 March 1996 Fixed a place in isHalfDuplex() in FRAME.C where in the case of an error in the process of determining whether audio is half duplex, the wave format buffer could be freed twice. For some reason, trying to compile the program on a Pentium with twice as much free RAM as the 486, the resource compiler dies in the middle of windowsx.h with "Out of far heap". I excluded this file from resource compiler builds, and the little pointy head now deigns to work. The VAT IDLIST packet parsing in FRAME.C's controlInput() had an off-by-one error when the name string length was a multiple of 4 and the terminating '\0' fell into the next 4 byte segment. Fixed. One final gratuitous Gatesian gutshot this fine Saturday night--Virtual (if you confuse it for a real compiler, you're seeing things) C 1.52c generates bad code for the VAT IDLIST packet parser in controlInput() (FRAME.C) when full optimisation is selected. (I'm sure, gentle reader, this will surprise you, having come this far with me down the rathole.) So, I #pragma'ed off all optimisation in that function, wishing there were were some way I could #fragma the "making it all too much" crowd spewing their ghastly gigabytes into an industry I was once proud to be a part of. 6.0-Alpha 4 prerelease. 3 March 1996 Integrated lots of fixes all over the place for 32 bit compile problems. I made these changes in a very conservative manner--what the compiler sees when compiling the 16 bit version should be identical to the code before the fixes were installed. 4 March 1996 People were having so much trouble getting the automatic adaptive VOX to work that I disabled it and replaced it with a manual set-the-level yourself mechanism. The VOX Monitor dialogue now allows you move the red threshold indicator with the scroll bar with complete freedom, while (if audio input is live) showing the VU meter as before. Slow, Medium, and Fast continue to regulate the number of samples of silence which must be seen before transmission mutes. There is no need with a manual VOX adjustment for a Reset facility, so the menu item and button in the monitor dialogue for that function were removed. I also disabled VOX GSM compression mode, since I'm afraid explaining its interaction with manual VOX would only confuse people. All the code is still there for adaptive VOX and VOX GSM compression--if you compile with VOX_GSM defined, it will all come back. Just as I suspected, the workaround for Trumpet Winsock's WSAEINPROGRESS bugs in contacting the Look Who's Listening server walked right into the jaws of another flaky Winsock--this time the one that comes with Windows 95. You apparently can't count on it to always notify you when a socket is closed, even if you've requested such notification via WSAAsyncSelect(). This led to timeout warning messages, attempts to make socket calls on already-closed sockets, and other horrors. Throwing up my virtual hands in disgust, I ripped out all message-based event sequencing of the LWL socket in FRAME.C and replaced it with timer logic. This is a crazy way of doing what should be trivial, but it's the only way to guarantee (to the extent one can ever use that word in conjunction with something associated with Windows) we won't be nailed by timing windows, lost notifications, or other flaky behaviour on the part of Winsock. 5 March 1996 Well, that didn't work on a production build under PC-NFS, because a blocking connect (to a slow-to-respond LWL server) could interfere with data transmission. So (and I have a very, very bad feeling about this), I made LWL transmission entirely non-blocking. When we receive the FD_CONNECT notification, that triggers the send() and sets the timer to close the socket 8 seconds later (since we don't dare count on linger mode to work properly). Wanna bet we need a timer to back up the FD_CONNECT notification in case Winsock forgets to send us one? We'll see. Added some additional paranoia in LWL.C to make sure no traffic is sent to the LWL server while a non-blocking transmission is in progress. 6.0-Alpha 5 prerelease. 7 March 1996 It's reported that connecting to a VAT conference with 35 people active causes a "runtime error 202 at 0001:1E". I don't have the vaguest idea what is causing this, what the error number means, or even where the error message is coming from, and I can't reproduce it since I don't have access to multicast to get on such a conference. I am not going to let nonsense like this deny the 6.0 update to tens of thousands of users who will never go near anything like this. I just hammered a test in CONNECT.C that limits the number of participants displayed to 8 and puts an ellipsis at the end of the list if there are more than that. Added code to multicastJoin() in FRAME.C to join or drop any auxiliary sockets to the current multicast list, as well as the default port sockets. Modified monitorPort in CONNECT.C to call multicastJoin to drop and then reacquire the multicast memberships whenever a new auxiliary socket is created. 8 March 1996 Once again we see how the poor design of Windows turns what is conceptually an easy task into a snowdrift, nay polar caps, of tangled and tricky code that one is never really sure will work everywhere. This time it's double clicking in the connection window to begin continuous transmission (which a lot more people will be doing now that we have VOX and Break Input). Remember back on the 21st of February when I redid the logic of how audio input is switched off so as dodge various bullets in the multimedia complex? Well guess what...ever since then double clicking hasn't worked (I didn't notice this since I generally use the space bar). Now, the double click logic in CONNECT.C couldn't have been simpler or more compliant with the Windows interface guidelines--the second click merely extends the scope of what the first one does, and does so simply by not switching off input on the button up event following a double click. Ahhh, but recall that you can't just turn audio input on and off like a light switch. It takes a while for the input buffers to rattle through the message queue and all the assorted dust to settle. So when we received the double click and subsequent button up event, audio input was still in the process of being terminated (as a result of the first button up event), and that's no time to go and re-open it. What to do? Well, here comes another brutal hack necessitated by irrational Microsoft design. When we process an WM_LBUTTONUP in CONNECT.C, we no longer close audio input to the connection. Instead, we set a timer (does this sound familiar?) running to expire GetDoubleClickTime() after the button was released. If we see a WM_LBUTTONDBLCLK an its subsequent WM_LBUTTONUP before the timer expires, we revoke the time with KillTimer() and leave audio input active. If the timer goes off without our having seen further mouse action, audio input to the connection is shut off at that time through the expedient of faking a space bar input to the message loop. This is spectacularly ugly but it gets the job done. 12 March 1996 I fired up the Large Software Collider for another day of flailing away with Visual C++ 4.0 under Windows 95. I discovered that the last single remaining piece of non-paranoid code in RTPACKET.C, the part that innocently stores the 16 bit length into RTCP SDES items, ran afoul of Visual C++'s talent for invariably making the wrong choice when given the slightest latitude for doing so. In this case we had a structure with several bit fields which added up to 16 bits followed by an unsigned short. Visual C++ padded the bit fields (which we are not use anyway, since it would store them in backwards order) to a 32 bit boundary before generating the short. I changed the code to use hard-coded byte numbers and casts everywhere, essentially treating Visual C++ as an assembler, which is approximately its intellectual level. Integrated recent fixes from the 16 bit development stream to bring the source streams into sync. From now on the 32 bit version on the LSC will be the primary stream, with the 16 bit version automatically recompiled from the common source. In the process of being led down the garden path by a bug in BoundsChecker "Compile-Time Instrumentation" (GlobalFreePtr() in windowsx.h in WIN32 generates bogus bad handle messages), I cleaned up some code belch in FRAME.C by eliminating the dynamic allocation and free of the PCMWAVEFORMAT structure in various functions. The structure is small; there's no reason not to just put it on the stack and get rid of all the error tests, frees on various paths, etc. ugly etc. Simple file I/O is so ubquitous in programs that it was an irresistable target for the Giant Rat of Redmond. Now you're supposed to open files using CreateFile function (elegant choice of name, don't you think), which takes 6 pages to document in the API book. The old offically sanctioned way of dealing with files, _lopen(), _lread(), etc., which worked perfectly well, are still there, but they've gratuitously deleted the definition of READ_WRITE, which was how you were supposed to specify that mode. I examined WINDOWS.H for Windows 3.1 and discovered that OF_READWRITE, the new officially sanctioned name, has the same numerical value and is defined in 3.1's WINDOWS.H, so I just changed the references in CONNECT.C and FRAME.C to use that symbol in order to compile on both 16 and 32 bit without an ugly #ifdef. This, and removing an incorrect earlier fix which called _sopen (yet another entirely incompatible way of doing I/O) instead of _lopen, got sound files and ring working on 32 bit. 14 March 1996 The Look Who's Listening dialogues weren't working because they were testing for edit control notifications using hard-coded Win 3.1 notification codes. I changed them to use WM_COMMAND_NOTIFY, which fetches the notification code from the correct place for both 16 and 32 bit builds. Fixes were needed in both LWL.C and DIALOGS.C. The handling of the WM_MDIACTIVATE message in COMMAND.C had to be revised due to the "packing" change in its arguments between Win 16 and 32. After threading through the lies and disinformation provided by what purports to be a guide to porting code to 32 bit Windows (found on the current MSDN CD-ROM), I fixed the code to work correctly on both 16 and 32 bit. Found and "fixed" (in other words uglified) two more places in RTPACKET.C where the straightforward code from the RTP standard ran afoul of Visual C++ 4.0's eccentric ideas about structure alignment (which, as far as I know, no other 32 bit compiler seems to share). When I disabled modem support, I forgot that CRC.C was used only by the modem and didn't disable it as well, which wasted a little data segment space. Fixed. To avoid future disasters stemming from unwarranted trust in Microsoft compilers, I #ifdef-ed out the definitions of the RTP and RTCP packet structures in RTP.H, leaving them for documentation only. Guess what? Another bunch of field references popped out in LWL.C and RTPACKET.C. Most were harmless (at least on VC 4.0), but the ones that weren't explained why LWL searches weren't working. So, as far as I can tell, LWL publish and search are now both working again. BoundsChecker was nattering about how we terminate the main MDI frame window at exit time. I added a handler for WM_NCDESTROY to zero the hwndMDIClient handle so a (harmless) "bad window handle" is not passed to DefFrameProc() in FRAME.C. Who you gonna call? BoundsChecker!!! Staked a memory leak in the processing of VAT IDLIST packets in FRAME.C. If an IDLIST packet was identical to the last displayed, it was never released. Fixed. Memo to file: BoundsChecker will report several memory leaks of buffers allocated by malloc() by the various libraries. These are nothing to worry about, as they are allocated on the local heap and do not leave resources in use after Speak Freely exits. There is no reason to complicate the termination code purely to appease BoundsChecker. 15 March 1996 When I made the WAVEFORMAT structures automatic rather than dynamically allocated, I forgot to delete the error message for allocation failure. It's gone. If something blew up in processing at WM_CREATE time, it's possible the WM_DESTROY handler could try to walk through the MDI windows using a NULL MDI client window handle. Fixed. The WM_CREATE handler, onCreate() in FRAME.C, returned 0 in case of error, not -1 as specified by the Windows API book. Heaven knows what you are really supposed to do since returning 0, the documented value for success, actually terminates the application. So, I return 1 just like I used to, which has the merit of working. Under WIN32, our WinMain() function in NETFONE.C is not informed if a copy of Speak Freely is already executing. In order to implement the trick of double clicking an .SFX file to open a connection in an existing copy of Speak Freely, I had to add a FindWindow() call which searches for a preexisting window with our class name to detect if Speak Freely is already running. Because socket ports and audio hardware are non-shared resources, it is not possible to run more than one copy of Speak Freely at a time. Connections established on the command line with non-standard ports were not working because the auxiliary socket must be created after the hwndMDIFrame window creation is complete. I moved the command line processing from the main window onCreate() in FRAME.C back to INIT.C so it's done after the frame window is created and its handle therefore available. Added an indication to the title in the About dialogue as to whether this is a 16 or 32 bit version. If something screwed up in the process of creating a connection which displayed an alert, it was possible for the not-yet-initialised connection to receive a timer message which caused it to try to send a heartbeat with the protocol not yet specified, which referenced a NULL pointer. Fixed. 16 March 1996 Transmission of face images to remote hosts was broken in 32 bit because I accidentally used _lseek instead of _llseek in FACE.C to position the read pointer in the face file. You can get away with this on Windows 3.1, but not in 32 bit. Fixed. Modified the code that paints connection windows in CONNECT.C and the MDI child window class definition to adopt the "new look" of Windows 95 dialogue boxes. The old connection window looked too garish alongside "grey flannel" Windows 95. This only happens for 32 bit builds. Turned off some no-longer-necessary debug mode code from SPEAKER.C which could conflict with the display of faces and participants in a multi-party conference. No call to desdone() was made at application termination time, which orphaned several buffers allocated by desinit(). Since these were local storage allocated by malloc(), no harm was done, but I added a call to desdone() to get rid of the natters from BoundsChecker. Discovered that several of the libraries were being explicitly loaded from their home subdirectories rather than being dynamically selected based on the configuration. This explained why DES encryption wasn't working for Speak Freely protocol--the current library was never loaded! I fixed the search paths so that the current configuration libraries will always be used. Deleted the "Bounds Checker" configurations which led to gnarly ~n ZIP names. If I decide to try Bounds Checker compile-time instrumentation again (like, when it really works), I'll use a portable directory name like BOUNDER. The code which protects against OutputDebugString()s left in release version ran afoul of WIN32's defining this as a macro for its own evil purposes. I added a #undef to get rid of of any preexisting macro so our version will prevail without warning messages. 17 March 1996 Can you believe it? If you make an association between a file type and your application and double click on a long file name like "D:\Connections\Jean-Pascal Bauer.sfx" your application gets launched with the file name on the command line not quoted! Blown away is twenty years of Unix and MS-DOS convention that command line arguments are separated by spaces. Thanks, Billy boy. Okay, we have to handle this somehow, so I changed the command line parser in INIT.C and FRAME.C to, on 32 bit builds only, use a comma as the delimiter between multiple connection files on the command line. I am quite certain that Murkysoft will add quotes around long file names in a future "upgrade", at which time every application which programmed around the current idiocy will stop working. I suppose one could anticipate this dagger from the future and parse quoted file names in the present, but I am not going to stoop to speculative defensive programming. I discovered the reason we were getting (harmless) "Can't open include file messages" when attempting to re-scan dependencies was the zany way Build/Settings/Resources parses the list of include directories. I twiddled the list delimiters until it generated a correct command line to eliminate the natter. All the help buttons in the dialogues were broken due to the negative ID_HELP (used to avoid having a different help ID for every dialogue, thank you very much resource compiler) being widened to an int without sign extension in the WM_COMMAND handler. I added a cast to (short) to restore proper sign extension in both 16 and 32 bit builds. Made a new SF32.HPJ file in the help directory that allows recompilation with the Visual C 4.0 HCW help compiler. Remade the help file screen shot bitmaps for Windows 95 appearance. These are kept in a new HELP\BMP32 directory which is searched first by SF32.HPJ. 18 March 1996 Added a new Help/Performance Benchmark dialogue which displays performance of the various compression and encryption modes as a percentage of the real-time sample rate. The goal is to help people decide which modes their computer can handle and to evaluate different optimisations, compiler code generation options, etc. Integrated the NSA LPC-10 codec into a new LPC-10 subdirectory and added code to SPEAKER.C to support decompressing packets received in LPC-10. Transmission isn't implemented yet. 19 March 1996 Two references to multicastJoin in CONNECT.C weren't disabled when IP_MAX_MEMBERSHIPS isn't defined, indicating a brain-dead Winsock that doesn't support multicast. LPC-10 compression on output is now integrated into CONNECT.C. Added LPC-10 compression to the performance benchmark. 21 March 1996 Integrated a fix from the Unix version for bad decryption when both a key file and another encryption mode that requires padding to an 8 byte frame. Since the padding has been done before the key file encryption is performed, key file decryption in SPEAKER.C needs to pad to guarantee the entire frame for subsequent decryption is properly decoded. Integrated the new (slightly) improved Simple compression module, RATE.C, from the Unix version. 25 March 1996 Hitting Esc or clicking the "X" button didn't close the Performance Benchmark dialogue because there wasn't an IDCANCEL case in the dialogue procedure. Fixed. Added a Help button to the Benchmark dialogue with appropriate link to the topic in the help file. 6.1-Alpha 1 prerelease. 3 April 1996 The new simple compression code (RATE.C) broke the answering machine's replay of simple compressed packets. For various historical reasons, all poor, simple compression was handled differently than all other types in SPEAKER.C: simple compressed packets were not expanded in place as for GSM, LPC, etc., but left compressed when written to the answering machine and decompressed again when played. This was fine prior to RATE.C, since no state information was needed to perform the decompression, but with the new algorithm the connection data structure must be available to decompress and it is not, of course, around when one plays back a message from the answering machine. I modified the simple decompression code in SPEAKER.C to work like all other forms of compression and expand the sound buffer in place, then removed the hack from ANSWER.C which allowed the fComp2X bit to pass through to the answering machine file in answerSave(). (Reported by Marc de Groot, marc@immersive.com). Deleted some long-obsolete debugging code in SPEAKER.C, already turned off with #ifdef OBSOLETE. A race condition existed in ANSWER.C which could lead to a General Protection Fault from an access through a null pointer if the answering machine was either popped up or closed while audio was being received from the network. There was a possibility that answerSave() would decide the answering machine dialogue was up because hDlgAnswer was non-NULL, then try to add the new message to msgTable before the dialogue procedure got a chance to allocate the table. I added a test in answerSave() to only consider the answering machine dialogue up after it has finished its WM_INITDIALOG processing and the message table is allocated. (Reported by Marc de Groot.) Sending with PGP encryption failed when creating the session key temporary file due to an incompatible change in the GetTempFileName() function in Win32. Now you have to call GetTempPath() to find out where to stick the temp file. The obvious upward-compatible way to handle the 16 bit Windows API call clearly eluded the "Monkey-C, Monkey-doo" back at the slug ranch. (Reported by John Deters, jad@dsddhc.com). The same temporary directory diddling had to be done in pgpSetSessionKey() in FRAME.C handle the New Official Temporary Directory Nomenclature Mechanism. 4 April 1996 Okay, I get it now...the reason the MS-DOS window hangs with the annoying alert box is that PGP was being run directly rather then through SFPGP.PIF, which is marked "Close on exit". The PIF was not being used, in turn, thanks to the fact that under Visual C 4, the program you're working on is executed from the WinDebug or WinRel (or whatever) directory, not the project directory where SFPGP.PIF lives. This shouldn't be a problem in the production version since the .EXE and .PIF will be in the same archive. To facilitate testing of the development version, I copied SFPGP.PIF into both debug and release directories. Automatic PGP encryption and decryption of session keys now seems to work OK. After a day chasing the mysterious floating point error on a 486DX-33 when running the LPC decompression phase of the benchmark (reported by Bill Oatman (BillO@lpa.com), and correlating what was seen in Speak Freely with the sporadic reports of floating point errors in Home Planet which appear on 486DX machines as soon as a user installs Windows 95 (even if Home Planet has been running perfectly on the same machine for years under Windows 3.x), I am now persuaded there is something wrong in the management of the coprocessor/FPU control word related to Windows 95, which perhaps manifests itself only on slower (and maybe therefore earlier step revision) 486DX machines. Please allow me to explain.... According to Mushysoft's documentation for all C libraries dating back to Visual C 1, when a program receives control, all floating point exceptions are masked; in other words the FPU generates a NaN, Infinity, etc. instead of causing an interrupt. Here's the statement from the documentation for the _control87() function in the Visual C 4 online help: Note The run-time libraries mask all floating-point exceptions by default. Pretty unambiguous, huh? But if you believe that, then there is no way you can get a floating point exception interrupt at all, unless you use _control87() to deliberately unmask one or more exceptions which, I can assure you, I'm not doing. I did build a version of Speak Freely which unmasked exceptions and discovered that the exception which popped out of LPC.C's uncompressing phase was a completely harmless denormalised and/or underflow where the default (exception masked) behaviour of the FPU is precisely what is intended. But this particular folder of the Hex Files now contains a number of unambiguous reports from a variety of people, affecting two different applications, which contradict the expectation of no interrupts. Naturally, this problem does not manifest itself on any machine I have tested on, including a 486DX. A search of Mickeysoft's product support database for all the obvious keywords, examination of the known bug list for Visual C++ 4.0, and errata for the 486 on Intel's Web site came up blank on anything seemingly related to a problem of this nature. What to do? Well, like every other time something completely impossible happens on a Mangysoft platform, there's little one can do other than pile on defensive workarounds and add instrumentation to pin down the bug. In this case, I added a call on _control87() in the application initialisation code in WinMain() (NETFONE.C) which explicitly masks all floating point exceptions, even though they're supposed to already be masked when we receive control. In addition, I added temporary code, executed when you start the performance benchmark, which verifies that the FPU is in the correct modes according to the value returned by _control87(0, 0). If it isn't, a warning is issued and an attempt made to reset the modes to all masked. If this fails to do so, a second warning is issued. The output from these debug message boxes should, on a machine which exhibits the failure, point much more directly at the source of the problem.