OS/2 Professional

home *** CD-ROM | disk | FTP | other *** search

/ OS/2 Professional / OS2PRO194.ISO / os2 / patches / watcom / c386 / readme.a < prev next >

Wrap

Text File | 1992-05-27 | 21.2 KB | 569 lines

**************** ** C Compiler ** **************** #define L_U(s) L##s would not expand correctly if the argument s contains a backslash character. =========================================================== The 386 compiler was generating segment fixups for variables in the FLAT model when the /d2 option was specified. =========================================================== Based pointer bugs fixed. =========================================================== Under certain circumstances, the C compiler may cause a memory protection fault or exhibit strange behaviour if it runs low on memory. =========================================================== Specifying -mf after -zw would lead to "invalid register modified by pragma" messages. =========================================================== Compiler would hang if a function with a variable defined at an inner scope is inlined with the -oe option. =========================================================== Compiler would hang if a function with a variable defined with extern storage class is inlined with the -oe option. Compiler would page fault if a function is inlined in the second or third parts of a for statement. Real-mode compiler would report "Incomplete type" if an array was declared with a size that is a multiple of 64K. =========================================================== Allow multiple anonymous unions/structs within a struct/union. =========================================================== C/386 accessed "short" and "unsigned short" bit fields as if they were ints. =========================================================== Using symbol names in an in-line pragma sequence could cause the compiler to hang. =========================================================== Functions which have the __saveregs attribute now save and restore any segment registers that are used. =========================================================== L"..." were not handled correctly inside macros. =========================================================== The spawn and exec families of functions are now automatically recognized as being vararg functions, and are called accordingly. =========================================================== ******************** ** Code Generator ** ******************** If the address of a parameter was used as an array, the address would be incorrectly calculated. The following would not work. foo( int x ) { *((BYTE*)(&x)+1) = 1; } The problem did not occur if the parameter used was a pointer. The following would be ok. foo( int *x ) { x[1] = 1; } =========================================================== If the switches -zc -zw -zu -ml were used, and a struct assignment or parameter was used, where the source of the move (or parameter) was a const item, the compiler would generate code that set DS to CS without saving and restoring DS. For example. typedef struct { int i,j,k,l,m; } ST; const ST x = { 1,2,3,4,5 }; foo() { bar( x ); } =========================================================== If an I2000 message (Not enough memory to fully optimize procedure) or an I2001 message (Not enough memory to maintain full peephole) was issued, and constant data was being placed in the code segment (-mf, -zc switches used) then there was the possibility of the data being placed in the middle of code sequences or switch statement jump tables. =========================================================== If a function takes it arguments in the 8087 register (fortran intrinsic functions, C functions with special pragma), and one the arguments contained a ?: operator, or a call to the fortran MIN,MAX functions, the compiler could issue "Internal compiler error 10" and/or go into an infinite loop. =========================================================== Using the /4,/4r,/4s switches could lead to an internal compiler error 40. /3r or /3s could be used as a workaround. =========================================================== When calling a functions returning a structs, the correct value was not always retrieved. This occurred if the return value was assigned elsewhere or returned from the calling function with no intervening control flow or redefinition of its fields. For example value = foo(); printf( "%d\n", value.x ); // prints the wrong value return( value ); value = foo(); printf( "%d\n", value.x ); // prints the wrong value value2 = value; value = foo(); if( value.x ) printf( "%d\n", value.x );// prints the right value return( value ); // due to control flow value = foo(); value.y = 7; printf( "%d\n", value.x ); // prints the right value return( value ); // due to assignment =========================================================== The compiler sometimes assumed that the EAX register was not modified by functions like IF@DFMOD. Use the /d__NO_MATH_OPS as a workaround. =========================================================== Incorrect /d2 debugging information was being generated if /of+ (for the C compiler) was also specified. This showed up by the debugger's "show calls" command not working properly. =========================================================== The code generator might sometimes output references to a line number zero in the debugging information for line numbers. =========================================================== *************** ** C Library ** *************** malloc would not allocate DOS low memory blocks close to 64k in size under DOS4GW. Relaxed to allow 1M maximum. =========================================================== __Init_Argv_ is undefined in Windows 386 3S libraries. =========================================================== In OS/2 2.x versions of fstat() and getftime(), mask off network attribute before checking if it's a disk file. =========================================================== In the 386 math libraries for the /fpc /3s model, the following math functions contained incorrect code. acos asin atan atan2 cos cosh exp fabs fmod log log10 pow sin sinh sqrt tanh =========================================================== Using WILDARGV in /ml /mc or /mh caused a program to crash during startup. The fix was made to INITRTNS.ASM. =========================================================== Flags are now removed from stack after int 25 and int 26 (16-bit libraries). =========================================================== Check for -0.0 in sqrt. i.e. Now allow -0.0 as valid argument. =========================================================== Spawn functions did not work with Intel CodeBuilder. =========================================================== __CHP was modifying the condition codes on the 386. =========================================================== EBX was being destroyed by _bios_printer and _bios_serialcom in -3s libraries. =========================================================== ****************** ** Disassembler ** ****************** Object files with an extremely large number of segments in them would hang the disassembler. =========================================================== ************* ** DOS4/GW ** ************* The mouse button state is now supplied to a user-installed mouse callback function in BX. A DOS/16M kernel bug that caused the direction flag to be set during some calls to INT 21h/06 has been fixed. Several bug fixes and improvements since release 1.5: INT 10h functions that take or return values in segment registers are now supported. Page faults caused by the debugger, when VM is on, should no longer cause a crash out to DOS. One program may now spawn another under VM. Page faults caused by the DOS/16M kernel are now handled properly by VM. The mouse mickey count is now supplied to a user-installed mouse callback function in DI and SI. DOS/4GW programs will now run properly under EMM386. DOS function 06 no longer destroys the processor flags in a semi-random fashion. ==== Major functionality change === The 80386 debug registers are now used to monitor for writes within 16 bytes of absolute address zero, providing basic null pointer protection. To enable this, set the DOS4G environment variable to include the string "nullp": SET DOS4G=nullp If the user accesses a location in the range 0p to 15p, a single step trap will be generated. If no debugger is present to get control, DOS4GW will crash with our standard register dump and the message: "Null pointer accessed at " ... This setting is not the default because there is a small performance penalty on 386s for using the debug registers. If a debugger is present, it will get control through the single step/debug interrupt (INT 01). Note: NULL POINTER PROTECTION NOT AVAILABLE UNDER WINDOWS. Several bug fixes and improvments since release 1.4: The register dump at crash time is now more helpful and includes the contents of CR0, 2, and 3. The interrupt enable flag is no longer trashed in Windows enhanced mode. More stack space was made available for DOS4G initialization, which eliminates some random crashes during program load. IRQ 5 handling is improved, allowing PM handlers to handle interrupts from both real and protected mode. The unsupported INT 31h function 0102h (DosMemResize) now sets the carry flag properly. INT 31h function 0600h (Lock Segments) now returns with carry clear outside VM. Various failures to load are now diagnosed. Spawning another DOS/4GW application no longer causes loss of memory. The file handle in AX is now widened on return from INT 21h function 3Ch. The version number and copyright date have been updated. INT 21h functions 6 and 29h now work correctly. A program's exit code is now returned correctly in a Windows enhanced mode DOS box. =========================================================== Interaction with Pharlap based programs and DOS4GW VMM improved. VMM speed significantly increased. The mouse button state is now supplied to a user-installed mouse callback function in BX. A DOS/16M kernel bug that caused the direction flag to be set during some calls to INT 21h/06 has been fixed. =========================================================== ********************** ** Graphics Library ** ********************** The return value for _remapallpalette() was wrong. =========================================================== Chips and Technologies SuperVGA cards were not being identified properly. Also, modes _SVRES16COLOR and _XRES256COLOR were enabled for this mode. =========================================================== The number of text rows on the screen was not always being set properly, affecting _outtext(). =========================================================== There was a problem with the _setwindow() function, some images were not being displayed properly. =========================================================== Multi-lingual character support added to _grtext() for codepages 437, 863 and 850. =========================================================== The _scrolltextwindow() function in SuperVGA modes was using the default plotting action instead of _GPSET. =========================================================== The number of text rows and columns was not always correct with some SuperVGA adapters. =========================================================== The _polygon() function was occassionally blowing the stack causing memory to be trashed. =========================================================== Mixing printf() output with graphics output in SuperVGA modes would sometimes cause objects to appear at the top of the screen instead of the bottom. =========================================================== Added a new routine _GetStackLow() instead of referencing the variable _STACKLOW directly. This routine can be replaced in multitasking environments where the value in _STACKLOW is not valid. =========================================================== ******************** ** C Header Files ** ******************** STDIO.H was missing the prototype for 'fdopen'. =========================================================== All the header files that define 'NULL' were modified to check and see if the macro had already been defined. =========================================================== *************** ** Librarian ** *************** The command ++symbol.dll.exportname didn't work. =========================================================== Librarian was over-estimating the number of pages req'd for the dictionary. Result was libraries larger than necessary (but still valid). =========================================================== ************ ** Linker ** ************ The linker would GP fault when being fed CodeView style debugging information from the Zortech C++ compiler rather than ignoring it. =========================================================== When generating OS/2 LE or LX (flat) style executables, the linker would not place the contents of the OPTION DESCRIPTION string into the executable image. =========================================================== Linker would GP fault when the very first token started with a weird character (e.g. =). =========================================================== Overlay manager wasn't giving the debugger the correct information about segment locations. =========================================================== Relocation handling system in the 16-bit linker could not handle segments much larger than 16 megabytes when linking for os/2 flat model executables. =========================================================== Fill in 0x300 as the expected version number for windows. Windows will not automatically load static references to dlls if this is not set! =========================================================== The previous big change to the linker's relocation handling system caused it to GP fault in a few boundary cases. =========================================================== Overlapping ledatas in an object file could cause the linker to try and pad negative amounts of nulls. =========================================================== A few minor modifications to match extensions and bug fixes to the extended memory library. =========================================================== ************************ ** Windows Supervisor ** ************************ stopped supervisor from allowing aliasing of NULL pointers. =========================================================== callback routines allocated from 32-bit DLL's were not getting a proper thunking layer set up. =========================================================== supervisor was calling SetBitmapBits instead of GetBitmapBits for its cover of GetBitmapBits =========================================================== supervisor was not maintaining a complete list of alias selectors, and if the 32-bit application's memory block resized and the base address of the block changed, some alias selectors could have invalid base addresses, causing a page fault. ************ ** WVIDEO ** ************ Allow the OS/2 debugger to open up to 40 files (useful when lots of the DLL's being used have debugging information). =========================================================== When the global debugging information section exceeded 64K in size, there was a possiblity that the OS/2, Windows, and QNX versions of the debugger would access violate when attempting to read it in. =========================================================== If a program loaded a DLL and then hit a breakpoint then the Windows and OS/2 debuggers would not properly skip over the breakpoint when execution was resumed. =========================================================== The SHIFT-TAB key did not work properly in the Windows hosted debugger. =========================================================== Using print/window with structures could cause the debugger to corrupt memory. =========================================================== Watch points did not work correctly in the NLM debugger. =========================================================== Debugging a PharLap application that installed the floating point emulator did not properly disable the emulator when premeaturely terminating the application with the debugger. =========================================================== Automatically disable use of the debug registers in an OS/2 2.0 DOS box, since it doesn't allow us to modify them. =========================================================== Extremely large (in excess of 400 characters) command lines in debugger invoke files would act as if random characters were missing from the file. =========================================================== Some problems with the Novell remote link dropping packets were resolved. =========================================================== Garbage characters might appear at the end of some help topics. =========================================================== If not enough relocation values were entered for the NEW/SYMBOL command, and the debugger started prompting for new values, the symbolic information would not be correctly mapped. =========================================================== In the help window, double clicking on a word that started at the left edge of the window would not pick up the first letter when searching for the topic. E.g., if the word was "foobar", the topic searched for would be "oobar". =========================================================== You can now debug PharLap applications in an OS/2 DOS box. =========================================================== While debugging an overlaid application, the debugger may not have been able to examine the memory for a section if it was not already loaded when WVIDEO was entered. =========================================================== The memory address shown in the assembly window for a mov ax,[bp+di] instruction would not be correct if bp + di was greater than 64K. =========================================================== The Windows STD.DLL file was accidently compiled with some internal debugging support enabled. This could cause Windows to hang or the debugger to GP fault if the system did not have a monochrome display card attached. =========================================================== The debugger would get confused if the physically first line number information for a module was not also the numerically lowest line number for the file (#line directives could cause this not to be true). =========================================================== ********** ** WOMP ** ********** Check for 64k debugging segments was applied to all segments. =========================================================== ************** ** Profiler ** ************** Fixed problem with WPROF not detecting the presence of a mouse in all circumstances. Fixed problem with systems that do not have the "." directory as the first entry in a sub-directory. =========================================================== WSAMPOS2.EXE now can sample PM applications properly. ===========================================================