C and MISRA, blues... (arithmetic shifts)
in [Embedded]
|
Prev: STMicroelectronics Introduces ARM9 Microcontroler with 2.1 Mbytes Flash
Next: AC spike go pass XFRM and regulator to reset CPU!!??
From: Nils on 24 Apr 2008 18:12 Grant Edwards schrieb: > Here's something different, but I don't know if it's less ugly. > > int ArithmeticShiftRight (int value, unsigned int shift) > { > switch (shift) > { > case 0: return value; > case 1: return value/2; > case 2: return value/4; > case 3: return value/8; > ... > } > } > I thought about this one as well. With loop-unswitching optimization enabled it wouldn't just give a bit of code-bloat but give perfect code. However, it does not work if value is negative because (-7/2) != (-7>>1). A little known fact and often we can live with it, but I can't. To bad.. > One assumes the compiler is bright enough to optimize value/N to > an arithmetic right shift in the case where N is a constant > power of two. Since the case values are contiguous and > 0-based, that ought to turn into some sort of indexed or > table-driven jump that shouldn't result in more that a couple > instructions of overhead. Speaking of optimizations: I just had a look at the assembly output and GCC generates pretty good code for my function. It does not turn it into a arithmetic shift. Guess that would be asking for to much, but it does some very clever tricks :-) Nils
From: David Brown on 24 Apr 2008 18:17 Nils wrote: > I'm currently rewriting some numerical code for MISRA compliance. > > Signed shifts are not defined by the C-standard, and the code-checker > complaints. Well - no big surprise here. I knew that and did it > nevertheless. Now I have to rewrite. > > > But do you do if you need them anyway? I need *lots* of them, so in > despair I've just created this monster of unreadable code: > How about: int ArithmeticShiftRight (int value, unsigned int shift) { if (value < 0) { return -((int) (((unsigned int) (-value)) >> shift); } else { return ((int) (((unsigned int) (value)) >> shift); } } It's still horrible, but it probably compiles to better code.
From: Wilco Dijkstra on 24 Apr 2008 18:26 "Nils" <n.pipenbrinck(a)cubic.org> wrote in message news:67ccn9F2na9h5U1(a)mid.uni-berlin.de... > I'm currently rewriting some numerical code for MISRA compliance. > > Signed shifts are not defined by the C-standard, and the code-checker complaints. Well - no big surprise here. I knew > that and did it nevertheless. Now I have to rewrite. Sounds like you're going to write some seriously unreadable, unmaintainable and most likely buggy code... Try to explain that one to your manager! The C standard rather stupidly accomodates the merely theoretical case of one's complement (where the obvious implementation of signed shift would round negative numbers differently). MISRA has some good rules but there are a lot of bad ones as well. It's best to pick a subset - the last time I looked at it I threw out more than half of the rules. > Has anyone here seen a C-compiler that does something different than arithmetic shifts on signed integers in the last > 20 years? I've used a lot of compilers over the years, and they all shifted arithmetic. There is no compiler that doesn't correctly implement signed arithmetic. Nobody could get away with creating a compiler that can't compile existing code. Wilco
From: Wilco Dijkstra on 24 Apr 2008 18:28 "David Brown" <david.brown(a)hesbynett.removethisbit.no> wrote in message news:5_GdnfetN7l7m4zVnZ2dneKdnZydnZ2d(a)lyse.net... > Nils wrote: >> I'm currently rewriting some numerical code for MISRA compliance. >> >> Signed shifts are not defined by the C-standard, and the code-checker complaints. Well - no big surprise here. I knew >> that and did it nevertheless. Now I have to rewrite. >> >> >> But do you do if you need them anyway? I need *lots* of them, so in despair I've just created this monster of >> unreadable code: >> > > How about: > > int ArithmeticShiftRight (int value, unsigned int shift) > { > if (value < 0) { > return -((int) (((unsigned int) (-value)) >> shift); > } else { > return ((int) (((unsigned int) (value)) >> shift); > } > } > > It's still horrible, but it probably compiles to better code. But it's incorrect as it doesn't round the same way. Wilco
From: Wilco Dijkstra on 24 Apr 2008 18:33
"Walter Banks" <walter(a)bytecraft.com> wrote in message news:4811043E.B3ACBFAE(a)bytecraft.com... > I am sure you have looked into sign extend and possibly a divide > something like > > if (shift) value = value / (1 << shift); That doesn't round correctly either. > #define asr(shift,value) if (shift) (value = value / (1 << shift)) This has at least 4 different problems... Wilco |