[compiler-rt] [compiler-rt][AArch64] Provide basic implementations of SME memcpy/memmove in case of strictly aligned memory access (PR #138250)

Victor Campos via llvm-commits llvm-commits at lists.llvm.org
Fri May 30 05:26:46 PDT 2025


https://github.com/vhscampos updated https://github.com/llvm/llvm-project/pull/138250

>From cd24ec57a84ed837e3e22d589de15e3ed3691ddb Mon Sep 17 00:00:00 2001
From: Victor Campos <victor.campos at arm.com>
Date: Thu, 1 May 2025 13:31:36 +0100
Subject: [PATCH 1/2] [compiler-rt][AArch64] Provide basic implementations of
 SME memcpy/memmove in case of strictly aligned memory access

The existing implementations, written in assembly, make use of unaligned
accesses for performance reasons. They are not compatible with strict
aligned configurations, i.e. with `-mno-unaligned-access`.

If the functions are used in this scenario, an exception is raised due
to unaligned memory accesses.

This patch reintroduces vanilla implementations for these functions to
be used in strictly aligned configurations. The actual code is largely
based on the code from https://github.com/llvm/llvm-project/pull/77496
---
 .../builtins/aarch64/sme-libc-mem-routines.S  |  6 ++-
 .../lib/builtins/aarch64/sme-libc-routines.c  | 48 ++++++++++++++++++-
 2 files changed, 51 insertions(+), 3 deletions(-)

diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S b/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S
index e736829967c0c..cba4a2cbc4fa9 100644
--- a/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S
@@ -6,6 +6,8 @@
 
 #include "../assembly.h"
 
+#ifdef __ARM_FEATURE_UNALIGNED
+
 //
 //  __arm_sc_memcpy / __arm_sc_memmove
 //
@@ -346,4 +348,6 @@ DEFINE_COMPILERRT_FUNCTION(__arm_sc_memset)
         ret
 END_COMPILERRT_FUNCTION(__arm_sc_memset)
 
-#endif // __aarch64__
+#endif /* defined(__aarch64__) && __ARM_FP != 0 */
+
+#endif /* __ARM_FEATURE_UNALIGNED */
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c b/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c
index 07d6681485556..24b9b3bcb49da 100644
--- a/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c
@@ -1,7 +1,8 @@
 #include <stddef.h>
 
-/* The asm version uses FP registers. Use this on targets without them */
-#if __ARM_FP == 0
+// The asm version uses FP registers and unaligned memory accesses. Use this on
+// targets without them.
+#if __ARM_FP == 0 || !defined(__ARM_FEATURE_UNALIGNED)
 void *__arm_sc_memset(void *dest, int c, size_t n) __arm_streaming_compatible {
   unsigned char *destp = (unsigned char *)dest;
   unsigned char c8 = (unsigned char)c;
@@ -22,3 +23,46 @@ const void *__arm_sc_memchr(const void *src, int c,
 
   return NULL;
 }
+
+#ifndef __ARM_FEATURE_UNALIGNED
+
+static void *memcpy_fwd(void *dest, const void *src,
+                        size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  for (size_t i = 0; i < n; ++i)
+    destp[i] = srcp[i];
+  return dest;
+}
+
+static void *memcpy_rev(void *dest, const void *src,
+                        size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  while (n > 0) {
+    --n;
+    destp[n] = srcp[n];
+  }
+  return dest;
+}
+
+void *__arm_sc_memcpy(void *__restrict dest, const void *__restrict src,
+                      size_t n) __arm_streaming_compatible {
+  return memcpy_fwd(dest, src, n);
+}
+
+void *__arm_sc_memmove(void *dest, const void *src,
+                       size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  if ((srcp > (destp + n)) || (destp > (srcp + n)))
+    return __arm_sc_memcpy(dest, src, n);
+  if (srcp > destp)
+    return memcpy_fwd(dest, src, n);
+  return memcpy_rev(dest, src, n);
+}
+
+#endif /* !defined(__ARM_FEATURE_UNALIGNED) */

>From ca036e8d9befeec7a4bf4e1458ceb00c464f48d7 Mon Sep 17 00:00:00 2001
From: Victor Campos <victor.campos at arm.com>
Date: Fri, 30 May 2025 11:12:29 +0100
Subject: [PATCH 2/2] Changes: - Split functions into separate files. - Select
 which implementation to use based on target features. The   selection is now
 done in CMake.

---
 compiler-rt/cmake/builtin-config-ix.cmake     |  18 ++
 compiler-rt/lib/builtins/CMakeLists.txt       |  12 +-
 .../aarch64/sme-libc-memcpy-memmove.c         |  55 ++++
 .../builtins/aarch64/sme-libc-memset-memchr.c |  37 +++
 ...utines.S => sme-libc-opt-memcpy-memmove.S} | 133 +--------
 .../aarch64/sme-libc-opt-memset-memchr.S      | 261 ++++++++++++++++++
 .../lib/builtins/aarch64/sme-libc-routines.c  |  68 -----
 7 files changed, 395 insertions(+), 189 deletions(-)
 create mode 100644 compiler-rt/lib/builtins/aarch64/sme-libc-memcpy-memmove.c
 create mode 100644 compiler-rt/lib/builtins/aarch64/sme-libc-memset-memchr.c
 rename compiler-rt/lib/builtins/aarch64/{sme-libc-mem-routines.S => sme-libc-opt-memcpy-memmove.S} (69%)
 create mode 100644 compiler-rt/lib/builtins/aarch64/sme-libc-opt-memset-memchr.S
 delete mode 100644 compiler-rt/lib/builtins/aarch64/sme-libc-routines.c

diff --git a/compiler-rt/cmake/builtin-config-ix.cmake b/compiler-rt/cmake/builtin-config-ix.cmake
index cbb43a5958d2f..b6aea17ec10f2 100644
--- a/compiler-rt/cmake/builtin-config-ix.cmake
+++ b/compiler-rt/cmake/builtin-config-ix.cmake
@@ -50,6 +50,24 @@ void foo(void)  __arm_streaming_compatible {
 }
 ")
 
+builtin_check_c_compiler_source(COMPILER_RT_HAS_ARM_UNALIGNED
+"
+void foo() {
+#ifndef __ARM_FEATURE_UNALIGNED
+#error \"Unaligned accesses unsupported\"
+#endif
+}
+")
+
+builtin_check_c_compiler_source(COMPILER_RT_HAS_ARM_FP
+"
+void foo() {
+#ifndef __ARM_FP
+#error \"No floating-point support\"
+#endif
+}
+")
+
 check_include_files("sys/auxv.h"    COMPILER_RT_HAS_AUXV)
 
 if(ANDROID)
diff --git a/compiler-rt/lib/builtins/CMakeLists.txt b/compiler-rt/lib/builtins/CMakeLists.txt
index 0c986f484bf5e..634a817674f7b 100644
--- a/compiler-rt/lib/builtins/CMakeLists.txt
+++ b/compiler-rt/lib/builtins/CMakeLists.txt
@@ -582,9 +582,17 @@ if (COMPILER_RT_HAS_AARCH64_SME)
     set_source_files_properties(aarch64/arm_apple_sme_abi.s PROPERTIES COMPILE_FLAGS -march=armv8a+sme)
     message(STATUS "AArch64 Apple SME ABI routines enabled")
   elseif (NOT COMPILER_RT_DISABLE_AARCH64_FMV AND COMPILER_RT_HAS_FNO_BUILTIN_FLAG AND COMPILER_RT_AARCH64_FMV_USES_GLOBAL_CONSTRUCTOR)
-    list(APPEND aarch64_SOURCES aarch64/sme-abi.S aarch64/sme-libc-mem-routines.S aarch64/sme-abi-assert.c aarch64/sme-libc-routines.c)
+    if(COMPILER_RT_HAS_ARM_UNALIGNED AND COMPILER_RT_HAS_ARM_FP)
+      list(APPEND aarch64_SOURCES aarch64/sme-libc-opt-memset-memchr.S aarch64/sme-libc-opt-memcpy-memmove.S)
+    elseif(COMPILER_RT_HAS_ARM_UNALIGNED)
+      list(APPEND aarch64_SOURCES aarch64/sme-libc-memset-memchr.c aarch64/sme-libc-opt-memcpy-memmove.S)
+      message(WARNING "AArch64 SME ABI assembly-optimized memset/memchr disabled: target does not have hardware floating-point support.")
+    else()
+      list(APPEND aarch64_SOURCES aarch64/sme-libc-memset-memchr.c aarch64/sme-libc-memcpy-memmove.c)
+      message(WARNING "AArch64 SME ABI assembly-optimized routines disabled: target does not support unaligned accesses.")
+    endif()
     message(STATUS "AArch64 SME ABI routines enabled")
-    set_source_files_properties(aarch64/sme-libc-routines.c PROPERTIES COMPILE_FLAGS "-fno-builtin")
+    set_source_files_properties(aarch64/sme-libc-memset-memchr.c aarch64/sme-libc-memcpy-memmove.c PROPERTIES COMPILE_FLAGS "-fno-builtin")
   else()
     if(COMPILER_RT_DISABLE_AARCH64_FMV)
       message(WARNING "AArch64 SME ABI routines require function multiversioning support.")
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-memcpy-memmove.c b/compiler-rt/lib/builtins/aarch64/sme-libc-memcpy-memmove.c
new file mode 100644
index 0000000000000..9792079964d55
--- /dev/null
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-memcpy-memmove.c
@@ -0,0 +1,55 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file contains basic implementations of Scalable Matrix Extension (SME)
+/// compatible memcpy and memmove functions to be used when their assembly-
+/// optimized counterparts can't.
+///
+//===----------------------------------------------------------------------===//
+
+#include <stddef.h>
+
+static void *__arm_sc_memcpy_fwd(void *dest, const void *src,
+                                 size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  for (size_t i = 0; i < n; ++i)
+    destp[i] = srcp[i];
+  return dest;
+}
+
+static void *__arm_sc_memcpy_rev(void *dest, const void *src,
+                                 size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  while (n > 0) {
+    --n;
+    destp[n] = srcp[n];
+  }
+  return dest;
+}
+
+extern void *__arm_sc_memcpy(void *__restrict dest, const void *__restrict src,
+                             size_t n) __arm_streaming_compatible {
+  return __arm_sc_memcpy_fwd(dest, src, n);
+}
+
+extern void *__arm_sc_memmove(void *dest, const void *src,
+                              size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  const unsigned char *srcp = (const unsigned char *)src;
+
+  if ((srcp > (destp + n)) || (destp > (srcp + n)))
+    return __arm_sc_memcpy(dest, src, n);
+  if (srcp > destp)
+    return __arm_sc_memcpy_fwd(dest, src, n);
+  return __arm_sc_memcpy_rev(dest, src, n);
+}
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-memset-memchr.c b/compiler-rt/lib/builtins/aarch64/sme-libc-memset-memchr.c
new file mode 100644
index 0000000000000..85873c3554fbd
--- /dev/null
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-memset-memchr.c
@@ -0,0 +1,37 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file contains basic implementations of Scalable Matrix Extension (SME)
+/// compatible memset and memchr functions to be used when their assembly-
+/// optimized counterparts can't.
+///
+//===----------------------------------------------------------------------===//
+
+#include <stddef.h>
+
+extern void *__arm_sc_memset(void *dest, int c,
+                             size_t n) __arm_streaming_compatible {
+  unsigned char *destp = (unsigned char *)dest;
+  unsigned char c8 = (unsigned char)c;
+  for (size_t i = 0; i < n; ++i)
+    destp[i] = c8;
+
+  return dest;
+}
+
+extern const void *__arm_sc_memchr(const void *src, int c,
+                                   size_t n) __arm_streaming_compatible {
+  const unsigned char *srcp = (const unsigned char *)src;
+  unsigned char c8 = (unsigned char)c;
+  for (size_t i = 0; i < n; ++i)
+    if (srcp[i] == c8)
+      return &srcp[i];
+
+  return NULL;
+}
\ No newline at end of file
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S b/compiler-rt/lib/builtins/aarch64/sme-libc-opt-memcpy-memmove.S
similarity index 69%
rename from compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S
rename to compiler-rt/lib/builtins/aarch64/sme-libc-opt-memcpy-memmove.S
index cba4a2cbc4fa9..8bc759a29a312 100644
--- a/compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-opt-memcpy-memmove.S
@@ -1,13 +1,23 @@
+//===----------------------------------------------------------------------===//
+//
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-
-// Routines taken from libc/AOR_v20.02/string/aarch64
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file contains assembly-optimized implementations of Scalable Matrix
+/// Extension (SME) compatible memcpy and memmove functions.
+///
+/// These implementations depend on unaligned access support.
+///
+/// Routines taken from libc/AOR_v20.02/string/aarch64.
+///
+//===----------------------------------------------------------------------===//
 
 #include "../assembly.h"
 
-#ifdef __ARM_FEATURE_UNALIGNED
-
 //
 //  __arm_sc_memcpy / __arm_sc_memmove
 //
@@ -236,118 +246,3 @@ END_COMPILERRT_FUNCTION(__arm_sc_memcpy)
 
 DEFINE_COMPILERRT_FUNCTION_ALIAS(__arm_sc_memmove, __arm_sc_memcpy)
 
-// This version uses FP registers. Use this only on targets with them
-#if defined(__aarch64__) && __ARM_FP != 0
-//
-//  __arm_sc_memset
-//
-
-#define dstin    x0
-#define val      x1
-#define valw     w1
-#define count    x2
-#define dst      x3
-#define dstend2  x4
-#define zva_val  x5
-
-DEFINE_COMPILERRT_FUNCTION(__arm_sc_memset)
-#ifdef __ARM_FEATURE_SVE
-        mov     z0.b, valw
-#else
-        bfi valw, valw, #8, #8
-        bfi valw, valw, #16, #16
-        bfi val, val, #32, #32
-        fmov d0, val
-        fmov v0.d[1], val
-#endif
-        add     dstend2, dstin, count
-
-        cmp     count, 96
-        b.hi    7f  // set_long
-        cmp     count, 16
-        b.hs    4f  // set_medium
-        mov     val, v0.D[0]
-
-        /* Set 0..15 bytes.  */
-        tbz     count, 3, 1f
-        str     val, [dstin]
-        str     val, [dstend2, -8]
-        ret
-        nop
-1:      tbz     count, 2, 2f
-        str     valw, [dstin]
-        str     valw, [dstend2, -4]
-        ret
-2:      cbz     count, 3f
-        strb    valw, [dstin]
-        tbz     count, 1, 3f
-        strh    valw, [dstend2, -2]
-3:      ret
-
-        /* Set 17..96 bytes.  */
-4:  // set_medium
-        str     q0, [dstin]
-        tbnz    count, 6, 6f  // set96
-        str     q0, [dstend2, -16]
-        tbz     count, 5, 5f
-        str     q0, [dstin, 16]
-        str     q0, [dstend2, -32]
-5:      ret
-
-        .p2align 4
-        /* Set 64..96 bytes.  Write 64 bytes from the start and
-           32 bytes from the end.  */
-6:  // set96
-        str     q0, [dstin, 16]
-        stp     q0, q0, [dstin, 32]
-        stp     q0, q0, [dstend2, -32]
-        ret
-
-        .p2align 4
-7:  // set_long
-        and     valw, valw, 255
-        bic     dst, dstin, 15
-        str     q0, [dstin]
-        cmp     count, 160
-        ccmp    valw, 0, 0, hs
-        b.ne    9f  // no_zva
-
-#ifndef SKIP_ZVA_CHECK
-        mrs     zva_val, dczid_el0
-        and     zva_val, zva_val, 31
-        cmp     zva_val, 4              /* ZVA size is 64 bytes.  */
-        b.ne    9f  // no_zva
-#endif
-        str     q0, [dst, 16]
-        stp     q0, q0, [dst, 32]
-        bic     dst, dst, 63
-        sub     count, dstend2, dst      /* Count is now 64 too large.  */
-        sub     count, count, 128       /* Adjust count and bias for loop.  */
-
-        .p2align 4
-8:  // zva_loop
-        add     dst, dst, 64
-        dc      zva, dst
-        subs    count, count, 64
-        b.hi    8b  // zva_loop
-        stp     q0, q0, [dstend2, -64]
-        stp     q0, q0, [dstend2, -32]
-        ret
-
-9:  // no_zva
-        sub     count, dstend2, dst      /* Count is 16 too large.  */
-        sub     dst, dst, 16            /* Dst is biased by -32.  */
-        sub     count, count, 64 + 16   /* Adjust count and bias for loop.  */
-10: // no_zva_loop
-        stp     q0, q0, [dst, 32]
-        stp     q0, q0, [dst, 64]!
-        subs    count, count, 64
-        b.hi    10b  // no_zva_loop
-        stp     q0, q0, [dstend2, -64]
-        stp     q0, q0, [dstend2, -32]
-        ret
-END_COMPILERRT_FUNCTION(__arm_sc_memset)
-
-#endif /* defined(__aarch64__) && __ARM_FP != 0 */
-
-#endif /* __ARM_FEATURE_UNALIGNED */
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-opt-memset-memchr.S b/compiler-rt/lib/builtins/aarch64/sme-libc-opt-memset-memchr.S
new file mode 100644
index 0000000000000..b6939f114cbae
--- /dev/null
+++ b/compiler-rt/lib/builtins/aarch64/sme-libc-opt-memset-memchr.S
@@ -0,0 +1,261 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file contains assembly-optimized implementations of Scalable Matrix
+/// Extension (SME) compatible memset and memchr functions.
+///
+/// These implementations depend on unaligned access and floating-point support.
+///
+/// Routines taken from libc/AOR_v20.02/string/aarch64.
+///
+//===----------------------------------------------------------------------===//
+
+#include "../assembly.h"
+
+//
+//  __arm_sc_memset
+//
+
+#define dstin    x0
+#define val      x1
+#define valw     w1
+#define count    x2
+#define dst      x3
+#define dstend2  x4
+#define zva_val  x5
+
+DEFINE_COMPILERRT_FUNCTION(__arm_sc_memset)
+#ifdef __ARM_FEATURE_SVE
+        mov     z0.b, valw
+#else
+        bfi valw, valw, #8, #8
+        bfi valw, valw, #16, #16
+        bfi val, val, #32, #32
+        fmov d0, val
+        fmov v0.d[1], val
+#endif
+        add     dstend2, dstin, count
+
+        cmp     count, 96
+        b.hi    7f  // set_long
+        cmp     count, 16
+        b.hs    4f  // set_medium
+        mov     val, v0.D[0]
+
+        /* Set 0..15 bytes.  */
+        tbz     count, 3, 1f
+        str     val, [dstin]
+        str     val, [dstend2, -8]
+        ret
+        nop
+1:      tbz     count, 2, 2f
+        str     valw, [dstin]
+        str     valw, [dstend2, -4]
+        ret
+2:      cbz     count, 3f
+        strb    valw, [dstin]
+        tbz     count, 1, 3f
+        strh    valw, [dstend2, -2]
+3:      ret
+
+        /* Set 17..96 bytes.  */
+4:  // set_medium
+        str     q0, [dstin]
+        tbnz    count, 6, 6f  // set96
+        str     q0, [dstend2, -16]
+        tbz     count, 5, 5f
+        str     q0, [dstin, 16]
+        str     q0, [dstend2, -32]
+5:      ret
+
+        .p2align 4
+        /* Set 64..96 bytes.  Write 64 bytes from the start and
+           32 bytes from the end.  */
+6:  // set96
+        str     q0, [dstin, 16]
+        stp     q0, q0, [dstin, 32]
+        stp     q0, q0, [dstend2, -32]
+        ret
+
+        .p2align 4
+7:  // set_long
+        and     valw, valw, 255
+        bic     dst, dstin, 15
+        str     q0, [dstin]
+        cmp     count, 160
+        ccmp    valw, 0, 0, hs
+        b.ne    9f  // no_zva
+
+#ifndef SKIP_ZVA_CHECK
+        mrs     zva_val, dczid_el0
+        and     zva_val, zva_val, 31
+        cmp     zva_val, 4              /* ZVA size is 64 bytes.  */
+        b.ne    9f  // no_zva
+#endif
+        str     q0, [dst, 16]
+        stp     q0, q0, [dst, 32]
+        bic     dst, dst, 63
+        sub     count, dstend2, dst      /* Count is now 64 too large.  */
+        sub     count, count, 128       /* Adjust count and bias for loop.  */
+
+        .p2align 4
+8:  // zva_loop
+        add     dst, dst, 64
+        dc      zva, dst
+        subs    count, count, 64
+        b.hi    8b  // zva_loop
+        stp     q0, q0, [dstend2, -64]
+        stp     q0, q0, [dstend2, -32]
+        ret
+
+9:  // no_zva
+        sub     count, dstend2, dst      /* Count is 16 too large.  */
+        sub     dst, dst, 16            /* Dst is biased by -32.  */
+        sub     count, count, 64 + 16   /* Adjust count and bias for loop.  */
+10: // no_zva_loop
+        stp     q0, q0, [dst, 32]
+        stp     q0, q0, [dst, 64]!
+        subs    count, count, 64
+        b.hi    10b  // no_zva_loop
+        stp     q0, q0, [dstend2, -64]
+        stp     q0, q0, [dstend2, -32]
+        ret
+END_COMPILERRT_FUNCTION(__arm_sc_memset)
+
+//
+//  __arm_sc_memchr
+//
+
+#define srcin		x0
+#define chrin		w1
+#define cntin		x2
+
+#define result		x0
+
+#define src		x3
+#define	tmp		x4
+#define wtmp2		w5
+#define synd		x6
+#define soff		x9
+#define cntrem		x10
+
+#define vrepchr		v0
+#define vdata1		v1
+#define vdata2		v2
+#define vhas_chr1	v3
+#define vhas_chr2	v4
+#define vrepmask	v5
+#define vend		v6
+
+/*
+ * Core algorithm:
+ *
+ * For each 32-byte chunk we calculate a 64-bit syndrome value, with two bits
+ * per byte. For each tuple, bit 0 is set if the relevant byte matched the
+ * requested character and bit 1 is not used (faster than using a 32bit
+ * syndrome). Since the bits in the syndrome reflect exactly the order in which
+ * things occur in the original string, counting trailing zeros allows to
+ * identify exactly which byte has matched.
+ */
+
+DEFINE_COMPILERRT_FUNCTION(__arm_sc_memchr)
+	/* Do not dereference srcin if no bytes to compare.  */
+	cbz	cntin, 4f
+	/*
+	 * Magic constant 0x40100401 allows us to identify which lane matches
+	 * the requested byte.
+	 */
+	mov	wtmp2, #0x0401
+	movk	wtmp2, #0x4010, lsl #16
+	dup	vrepchr.16b, chrin
+	/* Work with aligned 32-byte chunks */
+	bic	src, srcin, #31
+	dup	vrepmask.4s, wtmp2
+	ands	soff, srcin, #31
+	and	cntrem, cntin, #31
+	b.eq	0f
+
+	/*
+	 * Input string is not 32-byte aligned. We calculate the syndrome
+	 * value for the aligned 32 bytes block containing the first bytes
+	 * and mask the irrelevant part.
+	 */
+
+	ld1	{vdata1.16b, vdata2.16b}, [src], #32
+	sub	tmp, soff, #32
+	adds	cntin, cntin, tmp
+	cmeq	vhas_chr1.16b, vdata1.16b, vrepchr.16b
+	cmeq	vhas_chr2.16b, vdata2.16b, vrepchr.16b
+	and	vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b
+	and	vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b
+	addp	vend.16b, vhas_chr1.16b, vhas_chr2.16b		/* 256->128 */
+	addp	vend.16b, vend.16b, vend.16b			/* 128->64 */
+	mov	synd, vend.d[0]
+	/* Clear the soff*2 lower bits */
+	lsl	tmp, soff, #1
+	lsr	synd, synd, tmp
+	lsl	synd, synd, tmp
+	/* The first block can also be the last */
+	b.ls	2f
+	/* Have we found something already? */
+	cbnz	synd, 3f
+
+0: // loop
+	ld1	{vdata1.16b, vdata2.16b}, [src], #32
+	subs	cntin, cntin, #32
+	cmeq	vhas_chr1.16b, vdata1.16b, vrepchr.16b
+	cmeq	vhas_chr2.16b, vdata2.16b, vrepchr.16b
+	/* If we're out of data we finish regardless of the result */
+	b.ls	1f
+	/* Use a fast check for the termination condition */
+	orr	vend.16b, vhas_chr1.16b, vhas_chr2.16b
+	addp	vend.2d, vend.2d, vend.2d
+	mov	synd, vend.d[0]
+	/* We're not out of data, loop if we haven't found the character */
+	cbz	synd, 0b
+
+1: // end
+	/* Termination condition found, let's calculate the syndrome value */
+	and	vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b
+	and	vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b
+	addp	vend.16b, vhas_chr1.16b, vhas_chr2.16b		/* 256->128 */
+	addp	vend.16b, vend.16b, vend.16b			/* 128->64 */
+	mov	synd, vend.d[0]
+	/* Only do the clear for the last possible block */
+	b.hi	3f
+
+2: // masklast
+	/* Clear the (32 - ((cntrem + soff) % 32)) * 2 upper bits */
+	add	tmp, cntrem, soff
+	and	tmp, tmp, #31
+	sub	tmp, tmp, #32
+	neg	tmp, tmp, lsl #1
+	lsl	synd, synd, tmp
+	lsr	synd, synd, tmp
+
+3: // tail
+	/* Count the trailing zeros using bit reversing */
+	rbit	synd, synd
+	/* Compensate the last post-increment */
+	sub	src, src, #32
+	/* Check that we have found a character */
+	cmp	synd, #0
+	/* And count the leading zeros */
+	clz	synd, synd
+	/* Compute the potential result */
+	add	result, src, synd, lsr #1
+	/* Select result or NULL */
+	csel	result, xzr, result, eq
+	ret
+
+4: // zero_length
+	mov	result, #0
+	ret
+END_COMPILERRT_FUNCTION(__arm_sc_memchr)
+
diff --git a/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c b/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c
deleted file mode 100644
index 24b9b3bcb49da..0000000000000
--- a/compiler-rt/lib/builtins/aarch64/sme-libc-routines.c
+++ /dev/null
@@ -1,68 +0,0 @@
-#include <stddef.h>
-
-// The asm version uses FP registers and unaligned memory accesses. Use this on
-// targets without them.
-#if __ARM_FP == 0 || !defined(__ARM_FEATURE_UNALIGNED)
-void *__arm_sc_memset(void *dest, int c, size_t n) __arm_streaming_compatible {
-  unsigned char *destp = (unsigned char *)dest;
-  unsigned char c8 = (unsigned char)c;
-  for (size_t i = 0; i < n; ++i)
-    destp[i] = c8;
-
-  return dest;
-}
-#endif
-
-const void *__arm_sc_memchr(const void *src, int c,
-                            size_t n) __arm_streaming_compatible {
-  const unsigned char *srcp = (const unsigned char *)src;
-  unsigned char c8 = (unsigned char)c;
-  for (size_t i = 0; i < n; ++i)
-    if (srcp[i] == c8)
-      return &srcp[i];
-
-  return NULL;
-}
-
-#ifndef __ARM_FEATURE_UNALIGNED
-
-static void *memcpy_fwd(void *dest, const void *src,
-                        size_t n) __arm_streaming_compatible {
-  unsigned char *destp = (unsigned char *)dest;
-  const unsigned char *srcp = (const unsigned char *)src;
-
-  for (size_t i = 0; i < n; ++i)
-    destp[i] = srcp[i];
-  return dest;
-}
-
-static void *memcpy_rev(void *dest, const void *src,
-                        size_t n) __arm_streaming_compatible {
-  unsigned char *destp = (unsigned char *)dest;
-  const unsigned char *srcp = (const unsigned char *)src;
-
-  while (n > 0) {
-    --n;
-    destp[n] = srcp[n];
-  }
-  return dest;
-}
-
-void *__arm_sc_memcpy(void *__restrict dest, const void *__restrict src,
-                      size_t n) __arm_streaming_compatible {
-  return memcpy_fwd(dest, src, n);
-}
-
-void *__arm_sc_memmove(void *dest, const void *src,
-                       size_t n) __arm_streaming_compatible {
-  unsigned char *destp = (unsigned char *)dest;
-  const unsigned char *srcp = (const unsigned char *)src;
-
-  if ((srcp > (destp + n)) || (destp > (srcp + n)))
-    return __arm_sc_memcpy(dest, src, n);
-  if (srcp > destp)
-    return memcpy_fwd(dest, src, n);
-  return memcpy_rev(dest, src, n);
-}
-
-#endif /* !defined(__ARM_FEATURE_UNALIGNED) */



More information about the llvm-commits mailing list