When X is nonzero, X ^ (X - 1) produces a mask of trailing zeros plus
the lowest set bit, so popcount of that expression equals ctz(X) + 1.
Folding to CTZ avoids the blsmsk+popcnt (and a cmove for the zero
case) sequence on targets with a direct CTZ.
gcc/ChangeLog:
PR middle-end/124630
* match.pd (popcount (x ^ (x - 1))): Fold to ctz (x) + 1 when
x is nonzero and CTZ is directly supported.
gcc/testsuite/ChangeLog:
PR middle-end/124630
* gcc.target/i386/pr124630.c: New test.
---
gcc/match.pd | 11 +++++++++++
gcc/testsuite/gcc.target/i386/pr124630.c | 12 ++++++++++++
2 files changed, 23 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/i386/pr124630.c
diff --git a/gcc/match.pd b/gcc/match.pd
index b037b1a2876..10fba2cb788 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -10377,6 +10377,17 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(BUILT_IN_POPCOUNT (convert:type0 @0))
(if (cfn == CFN_BUILT_IN_POPCOUNTLL)
(BUILT_IN_POPCOUNTLL (convert:type0 @0))))))))
+
+/* popcount (X ^ (X - 1)) is CTZ (X) + 1 when X is nonzero. */
+(simplify
+ (POPCOUNT (bit_xor:c tree_expr_nonzero_p@0
+ (plus @0 integer_minus_onep)))
+ (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
+ && direct_internal_fn_supported_p (IFN_CTZ, TREE_TYPE (@0),
+ OPTIMIZE_FOR_SPEED))
+ (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
+ (plus (CTZ:type (convert:utype @0))
+ { build_one_cst (type); }))))
#endif
/* PARITY simplifications. */
diff --git a/gcc/testsuite/gcc.target/i386/pr124630.c
b/gcc/testsuite/gcc.target/i386/pr124630.c
new file mode 100644
index 00000000000..440aede39fa
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr124630.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -mbmi" } */
+/* { dg-final { scan-assembler-not "blsmsk" } } */
+/* { dg-final { scan-assembler-not "cmove" } } */
+
+unsigned foo (unsigned a)
+{
+ if (a != 0)
+ return __builtin_popcount (a ^ (a - 1)) - 1;
+ else
+ return 32;
+}
--
2.34.1
