[Bug tree-optimization/102131] [12 Regression] wrong code at -O1 and above on x86_64-linux-gnu

[guojiufu at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102131

--- Comment #6 from Jiu Fu Guo <guojiufu at gcc dot gnu.org> ---
Drafted a patch as below.  With this patch, those cases can pass.

diff --git a/gcc/tree-ssa-loop-niter.c b/gcc/tree-ssa-loop-niter.c
index 7af92d1c893..a400c42919b 100644
--- a/gcc/tree-ssa-loop-niter.c
+++ b/gcc/tree-ssa-loop-niter.c
@@ -1866,6 +1866,24 @@ number_of_iterations_cond (class loop *loop,
              || !iv0->no_overflow || !iv1->no_overflow))
        return false;

+      /* GT/GE has been transformed to LT/LE already.
+       cmp_code could be LT, LE or NE
+
+       For LE/LT transform
+       {iv0.base, iv0.step} LT/LE {iv1.base, iv1.step}
+       to
+       {iv0.base, iv0.step - iv1.step} LT/LE {iv1.base, 0}
+       Negative iv0.step - iv1.step means decreasing until wrap,
+       then the transform is not accurate.
+
+       For example:
+       {1, +, 1} <= {4, +, 3}
+       is not same with
+       {1, +, -2} <= {4, +, 0}
+       */
+      if ((code == LE_EXPR || code == LT_EXPR) && tree_int_cst_sign_bit
(step))
+       return false;
+
       iv0->step = step;
       if (!POINTER_TYPE_P (type))
        iv0->no_overflow = false;