lib/mem: Bring back assembly versions of x86-64/IA-32 memory routines
This is an attempt to mitigate certain performance regressions. See #557.
diff --git a/common/lib/mem.asm_x86_64 b/common/lib/mem.asm_x86_64
new file mode 100644
index 00000000..f082324f
--- /dev/null
+++ b/common/lib/mem.asm_x86_64
@@ -0,0 +1,57 @@
+section .text
+
+global memcpy
+memcpy:
+ mov rcx, rdx
+ mov rax, rdi
+ rep movsb
+ ret
+
+global memset
+memset:
+ push rdi
+ mov rax, rsi
+ mov rcx, rdx
+ rep stosb
+ pop rax
+ ret
+
+global memmove
+memmove:
+ mov rcx, rdx
+ mov rax, rdi
+
+ cmp rdi, rsi
+ ja .copy_backwards
+
+ rep movsb
+ jmp .done
+
+ .copy_backwards:
+ lea rdi, [rdi+rcx-1]
+ lea rsi, [rsi+rcx-1]
+ std
+ rep movsb
+ cld
+
+ .done:
+ ret
+
+global memcmp
+memcmp:
+ mov rcx, rdx
+ repe cmpsb
+ je .equal
+
+ mov al, byte [rdi-1]
+ sub al, byte [rsi-1]
+ movsx rax, al
+ jmp .done
+
+ .equal:
+ xor eax, eax
+
+ .done:
+ ret
+
+section .note.GNU-stack noalloc noexec nowrite progbits
diff --git a/common/lib/mem.s2.asm_ia32 b/common/lib/mem.s2.asm_ia32
new file mode 100644
index 00000000..a23f685d
--- /dev/null
+++ b/common/lib/mem.s2.asm_ia32
@@ -0,0 +1,77 @@
+section .text
+
+global memcpy
+memcpy:
+ push esi
+ push edi
+ mov eax, dword [esp+12]
+ mov edi, eax
+ mov esi, dword [esp+16]
+ mov ecx, dword [esp+20]
+ rep movsb
+ pop edi
+ pop esi
+ ret
+
+global memset
+memset:
+ push edi
+ mov edx, dword [esp+8]
+ mov edi, edx
+ mov eax, dword [esp+12]
+ mov ecx, dword [esp+16]
+ rep stosb
+ mov eax, edx
+ pop edi
+ ret
+
+global memmove
+memmove:
+ push esi
+ push edi
+ mov eax, dword [esp+12]
+ mov edi, eax
+ mov esi, dword [esp+16]
+ mov ecx, dword [esp+20]
+
+ cmp edi, esi
+ ja .copy_backwards
+
+ rep movsb
+ jmp .done
+
+ .copy_backwards:
+ lea edi, [edi+ecx-1]
+ lea esi, [esi+ecx-1]
+ std
+ rep movsb
+ cld
+
+ .done:
+ pop edi
+ pop esi
+ ret
+
+global memcmp
+memcmp:
+ push esi
+ push edi
+ mov edi, dword [esp+12]
+ mov esi, dword [esp+16]
+ mov ecx, dword [esp+20]
+ repe cmpsb
+ je .equal
+ mov al, byte [edi-1]
+ sub al, byte [esi-1]
+ movsx eax, al
+ jmp .done
+
+ .equal:
+ xor eax, eax
+
+ .done:
+ pop edi
+ pop esi
+ ret
+
+section .note.GNU-stack noalloc noexec nowrite progbits
diff --git a/common/lib/memory.s2.c b/common/lib/memory.s2.c
index a5353b2b..c51bc4c7 100644
--- a/common/lib/memory.s2.c
+++ b/common/lib/memory.s2.c
@@ -1,3 +1,5 @@
+#if !defined (__x86_64__) && !defined (__i386__)
+
#include <stdint.h>
#include <stddef.h>
@@ -51,3 +53,5 @@ int memcmp(const void *s1, const void *s2, size_t n) {
return 0;
}
+
+#endif
