Hi!
We expand __builtin_popcount* (x) == 1 as
x ^ (x - 1) > x - 1, either unconditionally in tree-ssa-math-opts.cc
if we don't have a direct optab support for popcount, or during
expansion where we compare the costs of comparison of the popcount
against one vs. the above expression.
As mentioned in the PR, if we know from ranger that the argument is
not zero, we can emit x & (x - 1) == 0 test which is same number of
GIMPLE statements, but on many targets cheaper (e.g. whenever an AND
instruction can also set flags on whether result was zero or not).
The following patch does that. Bootstrapped/regtested on x86_64-linux
and i686-linux, ok for trunk?
2024-01-03 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/90693
* tree-ssa-math-opts.cc (match_single_bit_test): If
tree_expr_nonzero_p (arg), remember it in the second argument to
IFN_POPCOUNT or lower it as arg & (arg - 1) == 0 rather than
arg ^ (arg - 1) > arg - 1.
* internal-fn.cc (expand_POPCOUNT): If second argument to
IFN_POPCOUNT suggests arg is non-zero, try to expand it as
arg & (arg - 1) == 0 rather than arg ^ (arg - 1) > arg - 1.
* gcc.target/i386/pr90693-2.c: New test.
--- gcc/tree-ssa-math-opts.cc.jj 2024-01-03 11:51:29.581758254 +0100
+++ gcc/tree-ssa-math-opts.cc 2024-01-03 15:18:45.527037903 +0100
@@ -5194,12 +5194,14 @@ match_single_bit_test (gimple_stmt_itera
tree type = TREE_TYPE (arg);
if (!INTEGRAL_TYPE_P (type))
return;
+ bool nonzero_arg = tree_expr_nonzero_p (arg);
if (direct_internal_fn_supported_p (IFN_POPCOUNT, type, OPTIMIZE_FOR_BOTH))
{
/* Tell expand_POPCOUNT the popcount result is only used in equality
comparison with one, so that it can decide based on rtx costs. */
gimple *g = gimple_build_call_internal (IFN_POPCOUNT, 2, arg,
- integer_one_node);
+ nonzero_arg ? integer_zero_node
+ : integer_one_node);
gimple_call_set_lhs (g, gimple_call_lhs (call));
gimple_stmt_iterator gsi2 = gsi_for_stmt (call);
gsi_replace (&gsi2, g, true);
@@ -5209,19 +5211,29 @@ match_single_bit_test (gimple_stmt_itera
gimple *g = gimple_build_assign (argm1, PLUS_EXPR, arg,
build_int_cst (type, -1));
gsi_insert_before (gsi, g, GSI_SAME_STMT);
- g = gimple_build_assign (make_ssa_name (type), BIT_XOR_EXPR, arg, argm1);
+ g = gimple_build_assign (make_ssa_name (type),
+ nonzero_arg ? BIT_AND_EXPR : BIT_XOR_EXPR,
+ arg, argm1);
gsi_insert_before (gsi, g, GSI_SAME_STMT);
+ tree_code cmpcode;
+ if (nonzero_arg)
+ {
+ argm1 = build_zero_cst (type);
+ cmpcode = code;
+ }
+ else
+ cmpcode = code == EQ_EXPR ? GT_EXPR : LE_EXPR;
if (gcond *cond = dyn_cast <gcond *> (stmt))
{
gimple_cond_set_lhs (cond, gimple_assign_lhs (g));
gimple_cond_set_rhs (cond, argm1);
- gimple_cond_set_code (cond, code == EQ_EXPR ? GT_EXPR : LE_EXPR);
+ gimple_cond_set_code (cond, cmpcode);
}
else
{
gimple_assign_set_rhs1 (stmt, gimple_assign_lhs (g));
gimple_assign_set_rhs2 (stmt, argm1);
- gimple_assign_set_rhs_code (stmt, code == EQ_EXPR ? GT_EXPR : LE_EXPR);
+ gimple_assign_set_rhs_code (stmt, cmpcode);
}
update_stmt (stmt);
gimple_stmt_iterator gsi2 = gsi_for_stmt (call);
--- gcc/internal-fn.cc.jj 2024-01-03 11:51:23.252846093 +0100
+++ gcc/internal-fn.cc 2024-01-03 15:34:51.303415845 +0100
@@ -5118,10 +5118,13 @@ expand_POPCOUNT (internal_fn fn, gcall *
/* If .POPCOUNT call has 2 arguments, match_single_bit_test marked it
because the result is only used in an equality comparison against 1.
Use rtx costs in that case to determine if .POPCOUNT (arg) == 1
- or (arg ^ (arg - 1)) > arg - 1 is cheaper. */
+ or (arg ^ (arg - 1)) > arg - 1 is cheaper.
+ If .POPCOUNT second argument is 0, we additionally know that arg
+ is non-zero, so use arg & (arg - 1) == 0 instead. */
bool speed_p = optimize_insn_for_speed_p ();
tree lhs = gimple_call_lhs (stmt);
tree arg = gimple_call_arg (stmt, 0);
+ bool nonzero_arg = integer_zerop (gimple_call_arg (stmt, 1));
tree type = TREE_TYPE (arg);
machine_mode mode = TYPE_MODE (type);
do_pending_stack_adjust ();
@@ -5147,11 +5150,15 @@ expand_POPCOUNT (internal_fn fn, gcall *
1, OPTAB_DIRECT);
if (argm1 == NULL_RTX)
goto fail;
- rtx argxorargm1 = expand_simple_binop (mode, XOR, op0, argm1, NULL_RTX,
- 1, OPTAB_DIRECT);
+ rtx argxorargm1 = expand_simple_binop (mode, nonzero_arg ? AND : XOR, op0,
+ argm1, NULL_RTX, 1, OPTAB_DIRECT);
if (argxorargm1 == NULL_RTX)
goto fail;
- rtx cmp = emit_store_flag (NULL_RTX, GTU, argxorargm1, argm1, mode, 1, 1);
+ rtx cmp;
+ if (nonzero_arg)
+ cmp = emit_store_flag (NULL_RTX, EQ, argxorargm1, const0_rtx, mode, 1, 1);
+ else
+ cmp = emit_store_flag (NULL_RTX, GTU, argxorargm1, argm1, mode, 1, 1);
if (cmp == NULL_RTX)
goto fail;
rtx_insn *cmp_insns = get_insns ();
--- gcc/testsuite/gcc.target/i386/pr90693-2.c.jj 2024-01-03
16:24:54.300981504 +0100
+++ gcc/testsuite/gcc.target/i386/pr90693-2.c 2024-01-03 15:44:54.121005780
+0100
@@ -0,0 +1,33 @@
+/* PR tree-optimization/90693 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mno-abm -mno-popcnt -fdump-tree-optimized" } */
+/* { dg-final { scan-tree-dump-not "POPCOUNT \\\(" "optimized" } } */
+/* { dg-final { scan-tree-dump-not "__builtin_popcount(ll)? \\\(" "optimized"
} } */
+
+int
+foo (unsigned int x)
+{
+ if (x == 0)
+ __builtin_unreachable ();
+ return __builtin_popcount (x) == 1;
+}
+
+int
+bar (unsigned int x)
+{
+ return (x & (x - 1)) == 0;
+}
+
+int
+baz (unsigned long long x)
+{
+ if (x == 0)
+ __builtin_unreachable ();
+ return __builtin_popcountll (x) == 1;
+}
+
+int
+qux (unsigned long long x)
+{
+ return (x & (x - 1)) == 0;
+}
Jakub
