Hi Ayal, Thanks for the inputs, I will try this on GCC 4.0, which sounds quite interesting.
regards, Vasanth On 6/15/05, Ayal Zaks <[EMAIL PROTECTED]> wrote: > > > > > > Vasanth <[EMAIL PROTECTED]> > > > I am using powerpc-eabi-gcc (3.4.1) and trying to retarget it for a > > fully pipelined FPU. I have a DFA model for the FPU. I am looking at > > the code produced for a simple FIR algorithm (a loop iterating over an > > array, with a multiply-add operation per iteration). (I am not using > > the fused-madd) > > > > for (i = 0; i < 64; i++) > > accum = z[i] * h[i]; > > > > I have the FIR loop partially unrolled, yet am not seeing the multiply > > from say, iteration i+1, overlapping with the multiply from iteration > > i. From the scheduling dumps, I do see that the compiler knows that > > each use of the multiply is incurring the full latency of the multiply > > instead of having reduced latency by pipelining in software. The adds > > are also completely linked by data flow and the compiler does not seem > > to be using temporary registers to be able to exploit executing some > > of the adds in parallel. Hence, each add is stalled on the previous > > add. > > > > fadds f5,f0,f8 > > fadds f4,f5,f6 > > fadds f2,f4,f11 > > fadds f1,f2,f3 > > fadds f11,f1,f13 > > > > The register pressure is not very high. Registers f15-f31 are not used at > all. > > To break the linkage between the adds, try to keep the original loop > (instead of partially unrolling it yourself) and use -funroll-loops > -fvariable-expansion-in-unroller --param > max-variable-expansions-in-unroller=8 (or some other number greater than 1 > but small enough to avoid spills). > > (see http://gcc.gnu.org/ml/gcc/2004-09/msg01554.html) > > This too was introduced in GCC 4.0. > Ayal. > > > > > > My question is, am I expecting the wrong version of GCC to be doing > > this. I saw the following thread about SMS. > > > > http://gcc.gnu.org/ml/gcc/2003-09/msg00954.html > > > > that seems relevant. Would GCC 4.x be a better version for my > > requirement? If not, any ideas would be greatly appreciated. > > > > thanks in advance, > > Vasanth > >
