This is in reference to the thread "Optimizer broke code".
I rewrote the function due to finding an unrelated bug, because I wanted
to remove a feature, and because I wanted to make another effort at
using string functions to simplify it. Despite the function testing ok,
once again, in Win32 GCC, it failed in the same way on the ARM chip as
the previous function. The new function is below.
To summarize what was occurring, a string passed into the word wrap
function lcd_DisplayStr is processed, which then calls the LCD display
function lcd_DspStr. The string passed into lcd_DspStr is always random
Global -O2 optimization causes lcd_DisplayStr to malfunction. If I
compile the entire project with -O2, but only the problematic function
with "-O2 -fno-forward-propagate", it works. I looked up what
forward-propagate is but I don't understand what it's doing. I guess
for now I'll just leave it this way. Oddly enough, this is the third
version of this function, and the first one works as designed, which
isn't that drastically different from #2 and #2.
This has been a strange experience. I'm no C code expert, but I've
written a fair amount of it for my own stuff, and never ran across a
problem like this before.
int lcd_DisplayStr(unsignedchar* instr, int maxlines, int maxwidth, int startaddr, constunsignedshort *font, int ellipsis, int invert)
unsignedchar tempc[maxwidth + 1];
line_count = 0;
while((line_count < maxlines) && (*instr != 0))
memset(&tempc, 0, (maxwidth + 1));
strncpy(&tempc, instr, maxwidth);
strx = strrchr(&tempc, ' '); //look for the last space in output stringif((strlen(instr)) <= (strlen(&tempc))) //if the remaining instr will fit into one line, move input string pointer past already displayed text
instr = instr + strlen(instr);
elseif((strx != 0) && ((strx - &tempc) > (maxwidth - FRAGMENT))) //decide if word fragment long enough to leave on present line
instr = instr + (strx - &tempc);
*strx = 0; //terminate output string
instr = instr + maxwidth;
while(*instr == ' ') //eliminate leading spaces on subsequent output strings
while((strlen(&tempc)) < maxwidth) //fill up rest of &tempc line with spaces
if((line_count == maxlines) && (*instr != 0) && (ellipsis == 1)) //if entire instr can't be displayed, and ellipsis enabled, display "..."
*(&tempc + (maxwidth - 3)) = 0;
lcd_DspStr(&tempc, startaddr, invert, font);
startaddr = startaddr + (*(font + 1) * LCD_GRAPHIC_LINE);
Just for my curiosity:
What version of gcc are you using?
The option "-fno-forward-propagate" is not very common on newer gcc's
(even google does not find it in any gcc-option list). The more recent
gcc's do know "-f-forward-propagate", but this is enabled in all
optimization states EXEPT -O2.
Oliver wrote:> Just for my curiosity:>> What version of gcc are you using?>> The option "-fno-forward-propagate" is not very common on newer gcc's> (even google does not find it in any gcc-option list). The more recent> gcc's do know "-f-forward-propagate", but this is enabled in all> optimization states EXEPT -O2.>> Oliver
I'm using CodeSourcery G++ Lite 2011.03-42 (the most recent version), or
gcc 4.4.1, for ARM.
I found almost 80,000 results for "-fno-forward-propagate" on Google,
but a little trick; you have to drop the first hyphen, otherwise that
tells Google to not look for it. I found that -O2 is enabling this
optimization by doing "arm-none-eabi-gcc -Q -O2 --help=optimizer".
In any case, adding "-fno-forward-propagate" fixes the problem, but I'd
love to know why. Why out of all the other code is it picking on this
Oliver wrote:>>I also found that -fforward-propagate is enabled by default on -O, -O2,>>-O3, -Os.>> Yes, you are right. I mixed up the lines.>> Oliver
I think this could classify as "the optimizer broke my code", despite
this generally being very unlikely.
>I think this could classify as "the optimizer broke my code"
Sure. But ist still very likely, that this is caused by your code, and
very unlikely, that this is caused by the optimizer.
Anyhow, I compiled your code piece with optimization -O2 into a little
example, and, beside the compiler threw a bunch of warnings related to
pointer signdness problems, it run flawlessly in the gdb-simulator.
(Using the yagarto toolchain with gcc 4.5.2).
If you could build a short example, which can be used to reproduce the
fault, this would be helpfull.
Oliver wrote:>>I think this could classify as "the optimizer broke my code">> Sure. But ist still very likely, that this is caused by your code, and> very unlikely, that this is caused by the optimizer.>> Anyhow, I compiled your code piece with optimization -O2 into a little> example, and, beside the compiler threw a bunch of warnings related to> pointer signdness problems, it run flawlessly in the gdb-simulator.> (Using the yagarto toolchain with gcc 4.5.2).>> If you could build a short example, which can be used to reproduce the> fault, this would be helpfull.>> Oliver
I could have pointer signedness problems, but my compiler doesn't show
this, although I do like to eliminate all warnings even if they are
harmless. I am getting a warning related to a #define being redefined,
so warnings are enabled. The #define problem is an unrelated issue
involving third party code which I am still tracking down (CPU
Anyway, simply calling the function I posted results in it passing a few
characters of garbage to lcd_DspStr(). The target does have periodic
interrupts (but no RTOS) that could be overwriting part of the
stack...actually recent findings lead me to think this is the problem.
But no other code is affected, and there's a lot of other things going
on that should be getting corrupted, and the stack is used heavily. I
suppose you would have to have an interrupt in your test to more
accurately simulate it.
I've never been able to get ARM code simulation to work using gdb. It
would be very useful, but I just gave up. I could put an example
together, but including the interrupts won't simulate because it's CPU
dependent, correct? I'm just not sure how to provide an example that
can be simulated without a lot of work on your side.
gdb does not simulate Interrups, or any other ARM-hardware. To debug
this you need a JTAG debugger for the real hardware, or a commercial
siumulator. With this it should not be a big problem to find the point
in the code where tempc gets screwed up.
You will get the warnings, if you compile with -Wall, which is a good
idea to use as the standard warning flag. Just to see, what else could
be suspicous, add -Wextra, and if you ever plan to write standard-C, try
-pedantic (although this is not recommeded even by the gcc authors).
Klaus Wachtler wrote:> Do you really think -O3 could help to improve this?
I don't think so, and I think the same option causing me problems is
enabled with -O3 also.
Awhile ago I did some comparisons between the different optimization
levels, and found that -O3 resulted in larger code, and if I recall
correctly, worse or the same performance. So I stuck with -O2 in this