On Sun, 2026-02-08 at 06:17 +0800, Xi Ruoyao wrote:
> On Sat, 2026-02-07 at 11:54 -0700, Jeffrey Law wrote:
> > Robin's patch to fix nonzero bits for the result of a popcount triggered
> > a minor testsuite scan failure for loongarch. Essentially it was
> > expecting to see an ANDI for a zero-extend, but after the change we get
> > a slli with a count 0, which is used for sign extension on loongarch.
> >
> > Both are likely equally performant. So this just adjusts the test to
> > accept both forms. Now one could argue that the result of the popcount
>
> This case is ctz instead of popcount. On LoongArch popcount also has a
> redundant ANDI but due to a different mechanism: LoongArch has no
> scalar popcount so popcount is implemented by moving the value into a
> vector register and invoking vector popcount (if vector extension is
> available). Finally when the SI is moved back from vector reg to
> general reg, we hit https://gcc.gnu.org/PR119013.
>
> This ctz case can be fixed with the SUBREG_PROMOTED_SET trick in
> define_expand, I guess.
Hmm, the trick does not work because combine insist to modify its result
to zero_extend (?!).
The following seems to work for ctz2 and 3:
diff --git a/gcc/config/loongarch/loongarch.md
b/gcc/config/loongarch/loongarch.md
index 77b3298078a..f8693b7f593 100644
--- a/gcc/config/loongarch/loongarch.md
+++ b/gcc/config/loongarch/loongarch.md
@@ -1452,6 +1452,15 @@ (define_insn "ctz<mode>2"
[(set_attr "type" "clz")
(set_attr "mode" "<MODE>")])
+(define_insn "*ctzsi2_extend"
+ [(set (match_operand:DI 0 "register_operand" "=r")
+ (any_extend:DI
+ (ctz:SI (match_operand:SI 1 "register_operand" "r"))))]
+ "TARGET_64BIT"
+ "ctz.w\t%0,%1"
+ [(set_attr "type" "clz")
+ (set_attr "mode" "SI")])
+
;;
;; ....................
;;
But for vtz 1 and 4 it seems the andi is already in the tree:
int ctz1 (unsigned int x)
{
int _4;
int _5;
<bb 2> [local count: 1073741824]:
_4 = .CTZ (x_1(D), 32);
_5 = _4 & 31;
return _5;
}
int ctz4 (long unsigned int x)
{
int _4;
int _5;
<bb 2> [local count: 1073741824]:
_4 = .CTZ (x_1(D), 64);
_5 = _4 & 63;
return _5;
}
Is there an existing bug report for them?
--
Xi Ruoyao <[email protected]>
