x86: Small improvements for wcslen

Just a few QOL changes.
    1. Prefer `add` > `lea` as it has high execution units it can run
       on.
    2. Don't break macro-fusion between `test` and `jcc`
    3. Reduce code size by removing gratuitous padding bytes (-90
       bytes).

geometric_mean(N=20) of all benchmarks New / Original: 0.959

All string/memory tests pass.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
(cherry picked from commit 244b415d386487521882debb845a040a4758cb18)

diff --git a/sysdeps/x86_64/wcslen.S b/sysdeps/x86_64/wcslen.S
index 61edea1d14d454c6f69f68144dff0c299cab3236..ad066863a44ea0a56930f761e8c9040c166a41ca 100644 (file)
--- a/sysdeps/x86_64/wcslen.S
+++ b/sysdeps/x86_64/wcslen.S
@@ -41,82 +41,82 @@ ENTRY (__wcslen)
        pxor    %xmm0, %xmm0
 
        lea     32(%rdi), %rax
-       lea     16(%rdi), %rcx
+       addq    $16, %rdi
        and     $-16, %rax
 
        pcmpeqd (%rax), %xmm0
        pmovmskb %xmm0, %edx
        pxor    %xmm1, %xmm1
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm1
        pmovmskb %xmm1, %edx
        pxor    %xmm2, %xmm2
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm2
        pmovmskb %xmm2, %edx
        pxor    %xmm3, %xmm3
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm3
        pmovmskb %xmm3, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm0
        pmovmskb %xmm0, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm1
        pmovmskb %xmm1, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm2
        pmovmskb %xmm2, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm3
        pmovmskb %xmm3, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm0
        pmovmskb %xmm0, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm1
        pmovmskb %xmm1, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm2
        pmovmskb %xmm2, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        pcmpeqd (%rax), %xmm3
        pmovmskb %xmm3, %edx
+       addq    $16, %rax
        test    %edx, %edx
-       lea     16(%rax), %rax
        jnz     L(exit)
 
        and     $-0x40, %rax
@@ -133,104 +133,100 @@ L(aligned_64_loop):
        pminub  %xmm0, %xmm2
        pcmpeqd %xmm3, %xmm2
        pmovmskb %xmm2, %edx
+       addq    $64, %rax
        test    %edx, %edx
-       lea     64(%rax), %rax
        jz      L(aligned_64_loop)
 
        pcmpeqd -64(%rax), %xmm3
        pmovmskb %xmm3, %edx
+    addq       $48, %rdi
        test    %edx, %edx
-       lea     48(%rcx), %rcx
        jnz     L(exit)
 
        pcmpeqd %xmm1, %xmm3
        pmovmskb %xmm3, %edx
+    addq       $-16, %rdi
        test    %edx, %edx
-       lea     -16(%rcx), %rcx
        jnz     L(exit)
 
        pcmpeqd -32(%rax), %xmm3
        pmovmskb %xmm3, %edx
+    addq       $-16, %rdi
        test    %edx, %edx
-       lea     -16(%rcx), %rcx
        jnz     L(exit)
 
        pcmpeqd %xmm6, %xmm3
        pmovmskb %xmm3, %edx
+    addq       $-16, %rdi
        test    %edx, %edx
-       lea     -16(%rcx), %rcx
-       jnz     L(exit)
-
-       jmp     L(aligned_64_loop)
+       jz      L(aligned_64_loop)
 
        .p2align 4
 L(exit):
-       sub     %rcx, %rax
+       sub     %rdi, %rax
        shr     $2, %rax
        test    %dl, %dl
        jz      L(exit_high)
 
-       mov     %dl, %cl
-       and     $15, %cl
+       andl    $15, %edx
        jz      L(exit_1)
        ret
 
-       .p2align 4
+       /* No align here. Naturally aligned % 16 == 1.  */
 L(exit_high):
-       mov     %dh, %ch
-       and     $15, %ch
+       andl    $(15 << 8), %edx
        jz      L(exit_3)
        add     $2, %rax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_1):
        add     $1, %rax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_3):
        add     $3, %rax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail0):
-       xor     %rax, %rax
+       xorl    %eax, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail1):
-       mov     $1, %rax
+       movl    $1, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail2):
-       mov     $2, %rax
+       movl    $2, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail3):
-       mov     $3, %rax
+       movl    $3, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail4):
-       mov     $4, %rax
+       movl    $4, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail5):
-       mov     $5, %rax
+       movl    $5, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail6):
-       mov     $6, %rax
+       movl    $6, %eax
        ret
 
-       .p2align 4
+       .p2align 3
 L(exit_tail7):
-       mov     $7, %rax
+       movl    $7, %eax
        ret
 
 END (__wcslen)