Hi!

This simplification for power of two @1 and @2 folds to false (resp.
to true for the == version) if @1 is larger than @2 (in unsigned
comparison), because @1 & @2 is known to be zero (i.e. for shift count 0)
and for shift count larger than that it will be zero too, either because
@1 << @0 is even larger, or if @0 is too large @1 << @0 overflows to zero.
This is the case of e.g. ((4 << x) & 2) != 0, which is always false.
Now, this PR is about a different problem, if @1 is smaller than @2, say
((1 << x) & 256) != 0, but x has a very narrow type, say unsigned _BitInt(3),
then the largest possible value of x is 7 and ((1 << 7) & 256) is
still 0, 1 << 7 is 128 and so still smaller than 256.
So, if c1 - c2 is larger than maximum value of the shift count type
(resp. for the other case c2 - c1), it will be also always false (resp.
true).
Trying to improve it and using range of x (aka @0) is not needed,
this simplification folds it into @0 != (c1 - c2) and so will be folded
later.  Just the case where c1 - c2 overflows is problematic because
we've lost the details (unless we'd promote both operands or something).
Another possible way to do this would be build_int_cst and check for
the overflow flags, but I think this is shorter.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/16.2?

2026-07-29  Jakub Jelinek  <[email protected]>

        PR tree-optimization/126476
        * match.pd (((C << A) & D) != 0 -> A == 0,
        ((C << A) & D) == 0 -> A != 0): Fold to false/true if
        c1 - c2 resp. c2 - c1 is larger than maximum value of
        A.

        * gcc.dg/torture/bitint-103.c: New test.

--- gcc/match.pd.jj     2026-07-29 17:21:33.893285488 +0200
+++ gcc/match.pd        2026-07-29 17:17:59.855012063 +0200
@@ -5293,7 +5293,11 @@ (define_operator_list SYNC_FETCH_AND_AND
   (cmp (bit_and (lshift integer_pow2p@1 @0) integer_pow2p@2) integer_zerop)
    (with { int c1 = wi::clz (wi::to_wide (@1));
           int c2 = wi::clz (wi::to_wide (@2)); }
-    (if (c1 < c2)
+    (if (c1 < c2
+        /* If c1 - c2 isn't representable in TREE_TYPE (@0), it is also
+           never true, because for any valid x C << x will be smaller
+           than D.  See PR126476.  */
+        || c1 - c2 > wi::to_widest (TYPE_MAX_VALUE (TREE_TYPE (@0))))
      { constant_boolean_node (cmp == NE_EXPR ? false : true, type); }
      (icmp @0 { build_int_cst (TREE_TYPE (@0), c1 - c2); }))))
  (simplify
@@ -5301,7 +5305,8 @@ (define_operator_list SYNC_FETCH_AND_AND
    (if (tree_int_cst_sgn (@1) > 0)
     (with { int c1 = wi::clz (wi::to_wide (@1));
            int c2 = wi::clz (wi::to_wide (@2)); }
-     (if (c1 > c2)
+     (if (c1 > c2
+         || c2 - c1 > wi::to_widest (TYPE_MAX_VALUE (TREE_TYPE (@0))))
       { constant_boolean_node (cmp == NE_EXPR ? false : true, type); }
       (icmp @0 { build_int_cst (TREE_TYPE (@0), c2 - c1); }))))))
 
--- gcc/testsuite/gcc.dg/torture/bitint-103.c.jj        2026-07-29 
17:19:45.916660975 +0200
+++ gcc/testsuite/gcc.dg/torture/bitint-103.c   2026-07-29 17:20:08.750370102 
+0200
@@ -0,0 +1,22 @@
+/* PR tree-optimization/126476 */
+/* { dg-do run { target bitint } } */
+
+[[gnu::noipa]] int
+foo (unsigned _BitInt(4) n)
+{
+  return ((1ULL << n) & (1ULL << 20)) != 0;
+}
+
+[[gnu::noipa]] int
+bar (unsigned _BitInt(4) n)
+{
+  return (((1ULL << 40) >> n) & (1ULL << 20)) != 0;
+}
+
+int
+main ()
+{
+  for (unsigned i = 0; i < 16; i++)
+    if (foo (i) != 0 || bar (i) != 0)
+        __builtin_abort ();
+}

        Jakub

Reply via email to