Update ATOM string routines to latest
Change-Id: I22a600e7f172681cfd38ff73a64e3fd07b284959
Signed-off-by: Lu, Hongjiu <hongjiu.lu@intel.com>
Signed-off-by: Bruce Beare <bruce.j.beare@intel.com>
diff --git a/libc/arch-x86/string/sse2-strlen-atom.S b/libc/arch-x86/string/sse2-strlen-atom.S
new file mode 100644
index 0000000..8911868
--- /dev/null
+++ b/libc/arch-x86/string/sse2-strlen-atom.S
@@ -0,0 +1,369 @@
+#define STRLEN sse2_strlen_atom
+
+#ifndef L
+# define L(label) .L##label
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc .cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc .cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off) .cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg) .cfi_restore reg
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off) .cfi_adjust_cfa_offset off
+#endif
+
+#ifndef cfi_remember_state
+# define cfi_remember_state .cfi_remember_state
+#endif
+
+#ifndef cfi_restore_state
+# define cfi_restore_state .cfi_restore_state
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name) \
+ .type name, @function; \
+ .globl name; \
+ .p2align 4; \
+name: \
+ cfi_startproc
+#endif
+
+#ifndef END
+# define END(name) \
+ cfi_endproc; \
+ .size name, .-name
+#endif
+
+#define CFI_PUSH(REG) \
+ cfi_adjust_cfa_offset (4); \
+ cfi_rel_offset (REG, 0)
+
+#define CFI_POP(REG) \
+ cfi_adjust_cfa_offset (-4); \
+ cfi_restore (REG)
+
+#define PUSH(REG) pushl REG; CFI_PUSH (REG)
+#define POP(REG) popl REG; CFI_POP (REG)
+#define PARMS 4
+#define STR PARMS
+#define ENTRANCE
+#define RETURN ret
+
+ .text
+ENTRY (STRLEN)
+ ENTRANCE
+ mov STR(%esp), %edx
+ xor %eax, %eax
+ cmpb $0, (%edx)
+ jz L(exit_tail0)
+ cmpb $0, 1(%edx)
+ jz L(exit_tail1)
+ cmpb $0, 2(%edx)
+ jz L(exit_tail2)
+ cmpb $0, 3(%edx)
+ jz L(exit_tail3)
+ cmpb $0, 4(%edx)
+ jz L(exit_tail4)
+ cmpb $0, 5(%edx)
+ jz L(exit_tail5)
+ cmpb $0, 6(%edx)
+ jz L(exit_tail6)
+ cmpb $0, 7(%edx)
+ jz L(exit_tail7)
+ cmpb $0, 8(%edx)
+ jz L(exit_tail8)
+ cmpb $0, 9(%edx)
+ jz L(exit_tail9)
+ cmpb $0, 10(%edx)
+ jz L(exit_tail10)
+ cmpb $0, 11(%edx)
+ jz L(exit_tail11)
+ cmpb $0, 12(%edx)
+ jz L(exit_tail12)
+ cmpb $0, 13(%edx)
+ jz L(exit_tail13)
+ cmpb $0, 14(%edx)
+ jz L(exit_tail14)
+ cmpb $0, 15(%edx)
+ jz L(exit_tail15)
+ pxor %xmm0, %xmm0
+ mov %edx, %eax
+ mov %edx, %ecx
+ and $-16, %eax
+ add $16, %ecx
+ add $16, %eax
+
+ pcmpeqb (%eax), %xmm0
+ pmovmskb %xmm0, %edx
+ pxor %xmm1, %xmm1
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm1
+ pmovmskb %xmm1, %edx
+ pxor %xmm2, %xmm2
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+
+ pcmpeqb (%eax), %xmm2
+ pmovmskb %xmm2, %edx
+ pxor %xmm3, %xmm3
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm3
+ pmovmskb %xmm3, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm0
+ pmovmskb %xmm0, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm1
+ pmovmskb %xmm1, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm2
+ pmovmskb %xmm2, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm3
+ pmovmskb %xmm3, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm0
+ pmovmskb %xmm0, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm1
+ pmovmskb %xmm1, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm2
+ pmovmskb %xmm2, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm3
+ pmovmskb %xmm3, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm0
+ pmovmskb %xmm0, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm1
+ pmovmskb %xmm1, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm2
+ pmovmskb %xmm2, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ pcmpeqb (%eax), %xmm3
+ pmovmskb %xmm3, %edx
+ test %edx, %edx
+ lea 16(%eax), %eax
+ jnz L(exit)
+
+ and $-0x40, %eax
+ PUSH (%esi)
+ PUSH (%edi)
+ PUSH (%ebx)
+ PUSH (%ebp)
+ xor %ebp, %ebp
+L(aligned_64):
+ pcmpeqb (%eax), %xmm0
+ pcmpeqb 16(%eax), %xmm1
+ pcmpeqb 32(%eax), %xmm2
+ pcmpeqb 48(%eax), %xmm3
+ pmovmskb %xmm0, %edx
+ pmovmskb %xmm1, %esi
+ pmovmskb %xmm2, %edi
+ pmovmskb %xmm3, %ebx
+ or %edx, %ebp
+ or %esi, %ebp
+ or %edi, %ebp
+ or %ebx, %ebp
+ lea 64(%eax), %eax
+ jz L(aligned_64)
+L(48leave):
+ test %edx, %edx
+ jnz L(aligned_64_exit_16)
+ test %esi, %esi
+ jnz L(aligned_64_exit_32)
+ test %edi, %edi
+ jnz L(aligned_64_exit_48)
+ mov %ebx, %edx
+ lea (%eax), %eax
+ jmp L(aligned_64_exit)
+L(aligned_64_exit_48):
+ lea -16(%eax), %eax
+ mov %edi, %edx
+ jmp L(aligned_64_exit)
+L(aligned_64_exit_32):
+ lea -32(%eax), %eax
+ mov %esi, %edx
+ jmp L(aligned_64_exit)
+L(aligned_64_exit_16):
+ lea -48(%eax), %eax
+L(aligned_64_exit):
+ POP (%ebp)
+ POP (%ebx)
+ POP (%edi)
+ POP (%esi)
+L(exit):
+ sub %ecx, %eax
+ test %dl, %dl
+ jz L(exit_high)
+ test $0x01, %dl
+ jnz L(exit_tail0)
+
+ test $0x02, %dl
+ jnz L(exit_tail1)
+
+ test $0x04, %dl
+ jnz L(exit_tail2)
+
+ test $0x08, %dl
+ jnz L(exit_tail3)
+
+ test $0x10, %dl
+ jnz L(exit_tail4)
+
+ test $0x20, %dl
+ jnz L(exit_tail5)
+
+ test $0x40, %dl
+ jnz L(exit_tail6)
+ add $7, %eax
+L(exit_tail0):
+ RETURN
+
+L(exit_high):
+ add $8, %eax
+ test $0x01, %dh
+ jnz L(exit_tail0)
+
+ test $0x02, %dh
+ jnz L(exit_tail1)
+
+ test $0x04, %dh
+ jnz L(exit_tail2)
+
+ test $0x08, %dh
+ jnz L(exit_tail3)
+
+ test $0x10, %dh
+ jnz L(exit_tail4)
+
+ test $0x20, %dh
+ jnz L(exit_tail5)
+
+ test $0x40, %dh
+ jnz L(exit_tail6)
+ add $7, %eax
+ RETURN
+
+ .p2align 4
+L(exit_tail1):
+ add $1, %eax
+ RETURN
+
+L(exit_tail2):
+ add $2, %eax
+ RETURN
+
+L(exit_tail3):
+ add $3, %eax
+ RETURN
+
+L(exit_tail4):
+ add $4, %eax
+ RETURN
+
+L(exit_tail5):
+ add $5, %eax
+ RETURN
+
+L(exit_tail6):
+ add $6, %eax
+ RETURN
+
+L(exit_tail7):
+ add $7, %eax
+ RETURN
+
+L(exit_tail8):
+ add $8, %eax
+ RETURN
+
+L(exit_tail9):
+ add $9, %eax
+ RETURN
+
+L(exit_tail10):
+ add $10, %eax
+ RETURN
+
+L(exit_tail11):
+ add $11, %eax
+ RETURN
+
+L(exit_tail12):
+ add $12, %eax
+ RETURN
+
+L(exit_tail13):
+ add $13, %eax
+ RETURN
+
+L(exit_tail14):
+ add $14, %eax
+ RETURN
+
+L(exit_tail15):
+ add $15, %eax
+ ret
+
+END (STRLEN)