From e258cfcf92f5e31e902fa045b41652f00fcf2521 Mon Sep 17 00:00:00 2001
From: dengjianbo <dengjianbo@loongson.cn>
Date: Thu, 24 Aug 2023 16:50:18 +0800
Subject: [PATCH 09/29] LoongArch: Add ifunc support for strcmp{aligned, lsx}

Based on the glibc microbenchmark, strcmp-aligned implementation could
reduce the runtime 0%-10% for aligned comparison, 10%-20% for unaligned
comparison, strcmp-lsx implemenation could reduce the runtime 0%-50%.

Signed-off-by: Peng Fan <fanpeng@loongson.cn>
Signed-off-by: ticat_fp <fanpeng@loongson.cn>
---
 sysdeps/loongarch/lp64/multiarch/Makefile     |   2 +
 .../lp64/multiarch/ifunc-impl-list.c          |   7 +
 .../loongarch/lp64/multiarch/ifunc-strcmp.h   |  38 ++++
 .../loongarch/lp64/multiarch/strcmp-aligned.S | 179 ++++++++++++++++++
 sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S | 165 ++++++++++++++++
 sysdeps/loongarch/lp64/multiarch/strcmp.c     |  35 ++++
 6 files changed, 426 insertions(+)
 create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
 create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
 create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
 create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp.c

diff --git a/sysdeps/loongarch/lp64/multiarch/Makefile b/sysdeps/loongarch/lp64/multiarch/Makefile
index c4dd3143..d5a500de 100644
--- a/sysdeps/loongarch/lp64/multiarch/Makefile
+++ b/sysdeps/loongarch/lp64/multiarch/Makefile
@@ -12,6 +12,8 @@ sysdep_routines += \
   strchrnul-aligned \
   strchrnul-lsx \
   strchrnul-lasx \
+  strcmp-aligned \
+  strcmp-lsx \
   memcpy-aligned \
   memcpy-unaligned \
   memmove-unaligned \
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
index 7cec0b77..9183b7da 100644
--- a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
@@ -62,6 +62,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, strchrnul, 1, __strchrnul_aligned)
 	      )
 
+  IFUNC_IMPL (i, name, strcmp,
+#if !defined __loongarch_soft_float
+	      IFUNC_IMPL_ADD (array, i, strcmp, SUPPORT_LSX, __strcmp_lsx)
+#endif
+	      IFUNC_IMPL_ADD (array, i, strcmp, 1, __strcmp_aligned)
+	      )
+
   IFUNC_IMPL (i, name, memcpy,
 #if !defined __loongarch_soft_float
               IFUNC_IMPL_ADD (array, i, memcpy, SUPPORT_LASX, __memcpy_lasx)
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h b/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
new file mode 100644
index 00000000..ca26352b
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
@@ -0,0 +1,38 @@
+/* Common definition for strcmp ifunc selection.
+   All versions must be listed in ifunc-impl-list.c.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <ldsodefs.h>
+#include <ifunc-init.h>
+
+#if !defined __loongarch_soft_float
+extern __typeof (REDIRECT_NAME) OPTIMIZE (lsx) attribute_hidden;
+#endif
+
+extern __typeof (REDIRECT_NAME) OPTIMIZE (aligned) attribute_hidden;
+
+static inline void *
+IFUNC_SELECTOR (void)
+{
+#if !defined __loongarch_soft_float
+  if (SUPPORT_LSX)
+    return OPTIMIZE (lsx);
+  else
+#endif
+    return OPTIMIZE (aligned);
+}
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S b/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
new file mode 100644
index 00000000..f5f4f336
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
@@ -0,0 +1,179 @@
+/* Optimized strcmp implementation using basic Loongarch instructions.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library.  If not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc)
+# define STRCMP_NAME __strcmp_aligned
+#else
+# define STRCMP_NAME strcmp
+#endif
+
+LEAF(STRCMP_NAME, 6)
+    lu12i.w     a4, 0x01010
+    andi        a2, a0, 0x7
+    ori         a4, a4, 0x101
+    andi        a3, a1, 0x7
+
+    bstrins.d   a4, a4, 63, 32
+    li.d        t7, -1
+    li.d        t8, 8
+    slli.d      a5, a4, 7
+
+    bne         a2, a3, L(unaligned)
+    bstrins.d   a0, zero, 2, 0
+    bstrins.d   a1, zero, 2, 0
+    ld.d        t0, a0, 0
+
+    ld.d        t1, a1, 0
+    slli.d      t3, a2, 3
+    sll.d       t2, t7, t3
+    orn         t0, t0, t2
+
+
+    orn         t1, t1, t2
+    sub.d       t2, t0, a4
+    andn        t3, a5, t0
+    and         t2, t2, t3
+
+    bne         t0, t1, L(al_end)
+L(al_loop):
+    bnez        t2, L(ret0)
+    ldx.d       t0, a0, t8
+    ldx.d       t1, a1, t8
+
+    addi.d      t8, t8, 8
+    sub.d       t2, t0, a4
+    andn        t3, a5, t0
+    and         t2, t2, t3
+
+    beq         t0, t1, L(al_loop)
+L(al_end):
+    xor         t3, t0, t1
+    or          t2, t2, t3
+    ctz.d       t3, t2
+
+
+    bstrins.d   t3, zero, 2, 0
+    srl.d       t0, t0, t3
+    srl.d       t1, t1, t3
+    andi        t0, t0, 0xff
+
+    andi        t1, t1, 0xff
+    sub.d       a0, t0, t1
+    jr          ra
+    nop
+
+L(ret0):
+    move        a0, zero
+    jr          ra
+    nop
+    nop
+
+L(unaligned):
+    slt         a6, a3, a2
+    xor         t0, a0, a1
+    maskeqz     t0, t0, a6
+    xor         a0, a0, t0
+
+
+    xor         a1, a1, t0
+    andi        a2, a0, 0x7
+    andi        a3, a1, 0x7
+    bstrins.d   a0, zero, 2, 0
+
+    bstrins.d   a1, zero, 2, 0
+    ld.d        t4, a0, 0
+    ld.d        t1, a1, 0
+    slli.d      a2, a2, 3
+
+    slli.d      a3, a3, 3
+    srl.d       t0, t4, a2
+    srl.d       t1, t1, a3
+    srl.d       t5, t7, a3
+
+    orn         t0, t0, t5
+    orn         t1, t1, t5
+    bne         t0, t1, L(not_equal)
+    sll.d       t5, t7, a2
+
+
+    sub.d       a3, a2, a3
+    orn         t4, t4, t5
+    sub.d       a2, zero, a3
+    sub.d       t2, t4, a4
+
+    andn        t3, a5, t4
+    and         t2, t2, t3
+    bnez        t2, L(find_zero)
+L(un_loop):
+    srl.d       t5, t4, a3
+
+    ldx.d       t4, a0, t8
+    ldx.d       t1, a1, t8
+    addi.d      t8, t8, 8
+    sll.d       t0, t4, a2
+
+    or          t0, t0, t5
+    bne         t0, t1, L(not_equal)
+    sub.d       t2, t4, a4
+    andn        t3, a5, t4
+
+
+    and         t2, t2, t3
+    beqz        t2, L(un_loop)
+L(find_zero):
+    sub.d       t2, t0, a4
+    andn        t3, a5, t0
+
+    and         t2, t2, t3
+    bnez        t2, L(ret0)
+    ldx.d       t1, a1, t8
+    srl.d       t0, t4, a3
+
+L(not_equal):
+    sub.d       t2, t0, a4
+    andn        t3, a5, t0
+    and         t2, t2, t3
+    xor         t3, t0, t1
+
+    or          t2, t2, t3
+L(un_end):
+    ctz.d       t3, t2
+    bstrins.d   t3, zero, 2, 0
+    srl.d       t0, t0, t3
+
+
+    srl.d       t1, t1, t3
+    andi        t0, t0, 0xff
+    andi        t1, t1, 0xff
+    sub.d       t2, t0, t1
+
+
+    sub.d       t3, t1, t0
+    masknez     t0, t2, a6
+    maskeqz     t1, t3, a6
+    or          a0, t0, t1
+
+    jr	ra
+END(STRCMP_NAME)
+
+libc_hidden_builtin_def (STRCMP_NAME)
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S b/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
new file mode 100644
index 00000000..2e177a38
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
@@ -0,0 +1,165 @@
+/* Optimized strcmp implementation using Loongarch LSX instructions.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library.  If not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc) && !defined __loongarch_soft_float
+
+# define STRCMP	__strcmp_lsx
+
+LEAF(STRCMP, 6)
+    pcalau12i       t0, %pc_hi20(L(INDEX))
+    andi            a2, a0, 0xf
+    vld             vr2, t0, %pc_lo12(L(INDEX))
+    andi            a3, a1, 0xf
+
+    bne             a2, a3, L(unaligned)
+    bstrins.d       a0, zero, 3, 0
+    bstrins.d       a1, zero, 3, 0
+    vld             vr0, a0, 0
+
+    vld             vr1, a1, 0
+    vreplgr2vr.b    vr3, a2
+    vslt.b          vr2, vr2, vr3
+    vseq.b          vr3, vr0, vr1
+
+    vmin.bu         vr3, vr0, vr3
+    vor.v           vr3, vr3, vr2
+    vsetanyeqz.b    fcc0, vr3
+    bcnez           fcc0, L(al_out)
+
+
+L(al_loop):
+    vld             vr0, a0, 16
+    vld             vr1, a1, 16
+    addi.d          a0, a0, 16
+    addi.d          a1, a1, 16
+
+    vseq.b          vr3, vr0, vr1
+    vmin.bu         vr3, vr0, vr3
+    vsetanyeqz.b    fcc0, vr3
+    bceqz           fcc0, L(al_loop)
+
+L(al_out):
+    vseqi.b         vr3, vr3, 0
+    vfrstpi.b       vr3, vr3, 0
+    vshuf.b         vr0, vr0, vr0, vr3
+    vshuf.b         vr1, vr1, vr1, vr3
+
+    vpickve2gr.bu   t0, vr0, 0
+    vpickve2gr.bu   t1, vr1, 0
+    sub.d           a0, t0, t1
+    jr              ra
+
+
+L(unaligned):
+    slt             a4, a3, a2
+    xor             t0, a0, a1
+    maskeqz         t0, t0, a4
+    xor             a0, a0, t0
+
+    xor             a1, a1, t0
+    andi            a2, a0, 0xf
+    andi            a3, a1, 0xf
+    bstrins.d       a0, zero, 3, 0
+
+    bstrins.d       a1, zero, 3, 0
+    vld             vr3, a0, 0
+    vld             vr1, a1, 0
+    vreplgr2vr.b    vr4, a2
+
+    vreplgr2vr.b    vr5, a3
+    vslt.b          vr7, vr2, vr5
+    vsub.b          vr5, vr5, vr4
+    vaddi.bu        vr6, vr2, 16
+
+
+    vsub.b          vr6, vr6, vr5
+    vshuf.b         vr0, vr3, vr3, vr6
+    vor.v           vr0, vr0, vr7
+    vor.v           vr1, vr1, vr7
+
+    vseq.b          vr5, vr0, vr1
+    vsetanyeqz.b    fcc0, vr5
+    bcnez           fcc0, L(not_equal)
+    vslt.b          vr4, vr2, vr4
+
+    vor.v           vr0, vr3, vr4
+    vsetanyeqz.b    fcc0, vr0
+    bcnez           fcc0, L(find_zero)
+    nop
+
+L(un_loop):
+    vld             vr3, a0, 16
+    vld             vr1, a1, 16
+    addi.d          a0, a0, 16
+    addi.d          a1, a1, 16
+
+
+    vshuf.b         vr0, vr3, vr0, vr6
+    vseq.b          vr5, vr0, vr1
+    vsetanyeqz.b    fcc0, vr5
+    bcnez           fcc0, L(not_equal)
+
+    vsetanyeqz.b    fcc0, vr3
+    vor.v           vr0, vr3, vr3
+    bceqz           fcc0, L(un_loop)
+L(find_zero):
+    vmin.bu         vr5, vr1, vr5
+
+    vsetanyeqz.b    fcc0, vr5
+    bcnez           fcc0, L(ret0)
+    vld             vr1, a1, 16
+    vshuf.b         vr0, vr3, vr3, vr6
+
+    vseq.b          vr5, vr0, vr1
+L(not_equal):
+    vmin.bu         vr5, vr0, vr5
+L(un_end):
+    vseqi.b         vr5, vr5, 0
+    vfrstpi.b       vr5, vr5, 0
+
+
+    vshuf.b         vr0, vr0, vr0, vr5
+    vshuf.b         vr1, vr1, vr1, vr5
+    vpickve2gr.bu   t0, vr0, 0
+    vpickve2gr.bu   t1, vr1, 0
+
+    sub.d           t3, t0, t1
+    sub.d           t4, t1, t0
+    masknez         t0, t3, a4
+    maskeqz         t1, t4, a4
+
+    or              a0, t0, t1
+    jr              ra
+L(ret0):
+    move            a0, zero
+    jr              ra
+END(STRCMP)
+
+    .section         .rodata.cst16,"M",@progbits,16
+    .align           4
+L(INDEX):
+    .dword           0x0706050403020100
+    .dword           0x0f0e0d0c0b0a0908
+
+libc_hidden_builtin_def (STRCMP)
+#endif
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp.c b/sysdeps/loongarch/lp64/multiarch/strcmp.c
new file mode 100644
index 00000000..6f249c0b
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp.c
@@ -0,0 +1,35 @@
+/* Multiple versions of strcmp.
+   All versions must be listed in ifunc-impl-list.c.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Define multiple versions only for the definition in libc.  */
+#if IS_IN (libc)
+# define strcmp __redirect_strcmp
+# include <string.h>
+# undef strcmp
+
+# define SYMBOL_NAME strcmp
+# include "ifunc-strcmp.h"
+
+libc_ifunc_redirected (__redirect_strcmp, strcmp, IFUNC_SELECTOR ());
+
+# ifdef SHARED
+__hidden_ver1 (strcmp, __GI_strcmp, __redirect_strcmp)
+  __attribute__ ((visibility ("hidden"))) __attribute_copy__ (strcmp);
+# endif
+#endif
-- 
2.33.0