https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79159
--- Comment #4 from amker at gcc dot gnu.org ---
Discussed with richi, and conclusion is that vrp issue is hard to fix at the
moment. Easy way out is to investigate why cunrolli peels one additional
iteration than necessary. Note cunrolli computes unrolling number using niter
information, which in turn is inferred from local array bound for this case :
"tmpCorr[9][9]".
The problem is in function record_nonwrapping_iv, as below code:
wide_int min, max;
extreme = fold_convert (unsigned_type, high);
if (TREE_CODE (orig_base) == SSA_NAME
&& TREE_CODE (low) == INTEGER_CST
&& INTEGRAL_TYPE_P (TREE_TYPE (orig_base))
&& get_range_info (orig_base, &min, &max) == VR_RANGE
&& wi::gts_p (min, low))
base = wide_int_to_tree (unsigned_type, min);
else if (TREE_CODE (base) != INTEGER_CST
&& dominated_by_p (CDI_DOMINATORS,
loop->latch, gimple_bb (stmt)))
base = fold_convert (unsigned_type, low);
delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
When analyzing "tmpCorr[i_3][j_4]" in below dump:
<bb 4> [15.00%]:
j_11 = i_3 + 1;
<bb 5> [100.00%]:
# j_4 = PHI <j_11(4), j_13(6)>
if (_1 <= j_4)
goto <bb 8>; [15.00%]
else
goto <bb 6>; [85.00%]
<bb 8> [15.00%]:
goto <bb 3>; [100.00%]
<bb 6> [85.00%]:
_2 = tmpCorr[i_3][j_4];
bar = _2;
j_13 = j_4 + 1;
goto <bb 5>; [100.00%]
The SCEV for j_4 in loop_2 is {j_11, 1}_2, aforementioned code fails to check
that j_11 is larger than 0, instead it uses "low = 0" as starting iteration,
resulting in peeling one additional iteration.
There are two possible fixes here. One is to investigate why evrp doesn't
compute correct range for j_11:
_1: VARYING
_3: VARYING
i_4: [0, +INF]
j_5: [j_13, +INF]
n_12(D): ~[0, 0]
j_13: VARYING <----inaccurate.
j_15: [-2147483647, +INF]
The other fix is to do more SCEV analysis on j_11 in outer loop.
Continue looking which one is better.
Thanks.
