On Tue, Jan 24, 2017 at 11:12 AM, Bin Cheng <bin.ch...@arm.com> wrote:
> Hi,
> Given test as reported in PR79159:
>
> void foo(float tmpCorr[9][9]);
> float bar;
>
> void finalDigits(int& n)
> {
> float tmpCorr[9][9] = {{0}};
>
> foo(tmpCorr);
> for (int i = 0; i < n; i++) {
> for (int j = i+1; j < n; j++) {
> bar = tmpCorr[i][j];
> }
> }
> }
> Pass cunrolli unrolls the inner loop with unrolling number 9 which is
> inferred from bound of local array definition: "tmpCorr[9][9]". In fact, it
> only needs to be unrolled by 8 times because the starting value of "j" is 1.
> However, loop niter analyzer fails to compute the accurate niter bound
> because cunrolli is before vrp pass and it doesn't know anything about outer
> loop's induction variable in inner loop handling. This patch computes init
> value of induction variable and uses that to improve boundary analysis.
> Bootstrap and test on x86_64 and AArch64. Is it OK?
Ok.
Thanks,
Richard.
> Thanks,
> bin
>
> 2017-01-23 Bin Cheng <bin.ch...@arm.com>
>
> PR tree-optimization/79159
> * tree-ssa-loop-niter.c (get_cst_init_from_scev): New function.
> (record_nonwrapping_iv): Imporve boundary using above function if no
> value range information.
>
> gcc/testsuite/ChangeLog
> 2017-01-23 Bin Cheng <bin.ch...@arm.com>
>
> PR tree-optimization/79159
> * g++.dg/tree-ssa/pr79159.C: New test.
