> On 19 Aug 2026, at 23:09, Wilco Dijkstra <[email protected]> wrote:
>
>
> Improve search_line_sve2() - like the AdvSIMD version, loop until a match is
> found even if the input pointer is unaligned or close to the end. This relies
> on CPP_BUFFER_PADDING >= 256 (maximum SVE vector length) and *end containing a
> match. The resulting code is both simpler and faster.
>
> Performance is ~1-2% better on a huge trace that replays the calls from
> libcpp.
>
> Passes bootstrap, OK for commit?
Ok.
Thanks for improving it.
Kyrill
>
> libcpp/ChangeLog:
>
> * lex.cc (search_line_sve): Improve.
>
> ---
>
> diff --git a/libcpp/lex.cc b/libcpp/lex.cc
> index
> bd35e28ce8f40a33e94f560b510107d49477d44d..047765fe733f028a663fe95c6bc206c9c4ee228f
> 100644
> --- a/libcpp/lex.cc
> +++ b/libcpp/lex.cc
> @@ -758,18 +758,19 @@ done:
> into a byte index, in place of the bitmask the Neon version has to
> build and move to a general register.
>
> - Unlike the Neon version this one needs neither alignment nor a
> - page-crossing test. Full-vector loads run while S is at or below
> - END rounded down to a vector boundary, so they may extend a little
> - past *END into the tail padding; a single predicated vector then
> - covers any remainder. The newline that _cpp_convert_input forces
> - at *END still terminates the scan.
> + Loop until a match is found even if the input pointer is unaligned
> + or close to the end. Since this may overread, it relies on *end
> + containing a match and CPP_BUFFER_PADDING >= 256 (maximum SVE vector
> + length).
>
> The loop consumes svcntb () bytes per iteration, so it scales with the
> implemented vector length. */
>
> +/* Unaligned loads, potentially using padding after the final newline. */
> +static_assert (CPP_BUFFER_PADDING >= 256, "");
> +
> static const uchar * __attribute__ ((target ("+sve2")))
> -search_line_sve2 (const uchar *s, const uchar *end)
> +search_line_sve2 (const uchar *s, const uchar *end ATTRIBUTE_UNUSED)
> {
> /* Order within a segment is irrelevant to MATCH, which tests set
> membership, so this needs no adjustment for big-endian. */
> @@ -777,34 +778,20 @@ search_line_sve2 (const uchar *s, const uchar *end)
> | ((uint32_t) '\\' << 16) | ((uint32_t) '?' << 24));
> const svuint8_t needles = svreinterpret_u8_u32 (svdup_n_u32 (chars));
> const svbool_t all = svptrue_b8 ();
> - const uint64_t vl = svcntb ();
> - svuint8_t data;
> - svbool_t match;
> - uintptr_t limit;
> -
> - /* Unaligned loads, potentially using padding after the final newline. */
> - static_assert (CPP_BUFFER_PADDING >= 256, "");
>
> - /* Align END down to a vector boundary so the loop can consume on
> - average half a vector more near the end. */
> - limit = (uintptr_t) end & -vl;
> + svuint8_t data = svld1_u8 (all, s);
> + svbool_t match = svmatch_u8 (all, data, needles);
> + if (__builtin_expect (svptest_any (all, match), 1))
> + return s + svcntp_b8 (all, svbrkb_b_z (all, match));
>
> - while ((uintptr_t) s <= limit)
> + while (1)
> {
> + s += svcntb ();
> data = svld1_u8 (all, s);
> match = svmatch_u8 (all, data, needles);
> - if (svptest_any (all, match))
> + if (__builtin_expect (svptest_any (all, match), 1))
> return s + svcntp_b8 (all, svbrkb_b_z (all, match));
> - s += vl;
> }
> -
> - /* What is left, up to and including *END, which _cpp_convert_input
> - forces to a newline. That guarantees a match, so no further test is
> - needed. */
> - svbool_t pg = svwhilele_b8_u64 ((uintptr_t) s, (uintptr_t) end);
> - data = svld1_u8 (pg, s);
> - match = svmatch_u8 (pg, data, needles);
> - return s + svcntp_b8 (pg, svbrkb_b_z (pg, match));
> }
>
> static bool lexer_has_sve2;
>
