From: "H.J. Lu" <hjl.tools@gmail.com>
To: Noah Goldstein <goldstein.w.n@gmail.com>
Cc: GNU C Library <libc-alpha@sourceware.org>
Subject: Re: [PATCH v2 5/7] x86: Optimize strcmp-avx2.S
Date: Sun, 9 Jan 2022 17:58:25 -0800 [thread overview]
Message-ID: <CAMe9rOrx5ncAmLRaLs_L3YHnqZOD2vXZjYa+fsG0y9V1Sw5fpA@mail.gmail.com> (raw)
In-Reply-To: <CAFUsyfJhd6P60uCPqg3iiCmgixp-5hH=3jVMqCC0abPZ1eXdHQ@mail.gmail.com>
On Sun, Jan 9, 2022 at 5:06 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Sun, Jan 9, 2022 at 6:41 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Sun, Jan 9, 2022 at 4:31 PM Noah Goldstein via Libc-alpha
> > <libc-alpha@sourceware.org> wrote:
> > >
> > > Optimization are primarily to the loop logic and how the page cross
> > > logic interacts with the loop.
> > >
> > > The page cross logic is at times more expensive for short strings near
> > > the end of a page but not crossing the page. This is done to retest
> > > the page cross conditions with a non-faulty check and to improve the
> > > logic for entering the loop afterwards. This is only particular cases,
> > > however, and is general made up for by more than 10x improvements on
> > > the transition from the page cross -> loop case.
> > >
> > > The non-page cross cases are improved most for smaller sizes [0, 128]
> > > and go about even for (128, 4096]. The loop page cross logic is
> > > improved so some more significant speedup is seen there as well.
> > >
> > > test-strcmp, test-strncmp, test-wcscmp, and test-wcsncmp all pass.
> > >
> > > Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> > > ---
> > > sysdeps/x86_64/multiarch/strcmp-avx2.S | 1590 ++++++++++++++----------
> > > 1 file changed, 939 insertions(+), 651 deletions(-)
> > >
> > > diff --git a/sysdeps/x86_64/multiarch/strcmp-avx2.S b/sysdeps/x86_64/multiarch/strcmp-avx2.S
> > > index 9c73b5899d..28d6a0025a 100644
> > > --- a/sysdeps/x86_64/multiarch/strcmp-avx2.S
> > > +++ b/sysdeps/x86_64/multiarch/strcmp-avx2.S
> > > @@ -26,35 +26,57 @@
> > >
> > > # define PAGE_SIZE 4096
> > >
> > > -/* VEC_SIZE = Number of bytes in a ymm register */
> > > + /* VEC_SIZE = Number of bytes in a ymm register. */
> > > # define VEC_SIZE 32
> > >
> > > -/* Shift for dividing by (VEC_SIZE * 4). */
> > > -# define DIVIDE_BY_VEC_4_SHIFT 7
> > > -# if (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
> > > -# error (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
> > > -# endif
> > > +# define VMOVU vmovdqu
> > > +# define VMOVA vmovdqa
> > >
> > > # ifdef USE_AS_WCSCMP
> > > -/* Compare packed dwords. */
> > > + /* Compare packed dwords. */
> > > # define VPCMPEQ vpcmpeqd
> > > -/* Compare packed dwords and store minimum. */
> > > + /* Compare packed dwords and store minimum. */
> > > # define VPMINU vpminud
> > > -/* 1 dword char == 4 bytes. */
> > > + /* 1 dword char == 4 bytes. */
> > > # define SIZE_OF_CHAR 4
> > > # else
> > > -/* Compare packed bytes. */
> > > + /* Compare packed bytes. */
> > > # define VPCMPEQ vpcmpeqb
> > > -/* Compare packed bytes and store minimum. */
> > > + /* Compare packed bytes and store minimum. */
> > > # define VPMINU vpminub
> > > -/* 1 byte char == 1 byte. */
> > > + /* 1 byte char == 1 byte. */
> > > # define SIZE_OF_CHAR 1
> > > # endif
> > >
> > > +# ifdef USE_AS_STRNCMP
> > > +# define LOOP_REG r9d
> > > +# define LOOP_REG64 r9
> > > +
> > > +# define OFFSET_REG8 r9b
> > > +# define OFFSET_REG r9d
> > > +# define OFFSET_REG64 r9
> > > +# else
> > > +# define LOOP_REG edx
> > > +# define LOOP_REG64 rdx
> > > +
> > > +# define OFFSET_REG8 dl
> > > +# define OFFSET_REG edx
> > > +# define OFFSET_REG64 rdx
> > > +# endif
> > > +
> > > # ifndef VZEROUPPER
> > > # define VZEROUPPER vzeroupper
> > > # endif
> > >
> > > +# if defined USE_AS_STRNCMP
> > > +# define VEC_OFFSET 0
> > > +# else
> > > +# define VEC_OFFSET (-VEC_SIZE)
> > > +# endif
> > > +
> > > +# define xmmZERO xmm15
> > > +# define ymmZERO ymm15
> > > +
> > > # ifndef SECTION
> > > # define SECTION(p) p##.avx
> > > # endif
> > > @@ -79,783 +101,1049 @@
> > > the maximum offset is reached before a difference is found, zero is
> > > returned. */
> > >
> > > - .section SECTION(.text),"ax",@progbits
> > > -ENTRY (STRCMP)
> > > + .section SECTION(.text), "ax", @progbits
> > > +ENTRY(STRCMP)
> > > # ifdef USE_AS_STRNCMP
> > > - /* Check for simple cases (0 or 1) in offset. */
> > > +# ifdef __ILP32__
> > > + /* Clear the upper 32 bits. */
> > > + movl %edx, %rdx
> > > +# endif
> > > cmp $1, %RDX_LP
> > > - je L(char0)
> > > - jb L(zero)
> > > + /* Signed comparison intentional. We use this branch to also
> > > + test cases where length >= 2^63. These very large sizes can be
> > > + handled with strcmp as there is no way for that length to
> > > + actually bound the buffer. */
> > > + jle L(one_or_less)
> > > # ifdef USE_AS_WCSCMP
> > > -# ifndef __ILP32__
> > > movq %rdx, %rcx
> > > - /* Check if length could overflow when multiplied by
> > > - sizeof(wchar_t). Checking top 8 bits will cover all potential
> > > - overflow cases as well as redirect cases where its impossible to
> > > - length to bound a valid memory region. In these cases just use
> > > - 'wcscmp'. */
> > > +
> > > + /* Multiplying length by sizeof(wchar_t) can result in overflow.
> > > + Check if that is possible. All cases where overflow are possible
> > > + are cases where length is large enough that it can never be a
> > > + bound on valid memory so just use wcscmp. */
> > > shrq $56, %rcx
> > > jnz __wcscmp_avx2
> > > +
> > > + leaq (, %rdx, 4), %rdx
> > > # endif
> > > - /* Convert units: from wide to byte char. */
> > > - shl $2, %RDX_LP
> > > -# endif
> > > - /* Register %r11 tracks the maximum offset. */
> > > - mov %RDX_LP, %R11_LP
> > > # endif
> > > + vpxor %xmmZERO, %xmmZERO, %xmmZERO
> > > movl %edi, %eax
> > > - xorl %edx, %edx
> > > - /* Make %xmm7 (%ymm7) all zeros in this function. */
> > > - vpxor %xmm7, %xmm7, %xmm7
> > > orl %esi, %eax
> > > - andl $(PAGE_SIZE - 1), %eax
> > > - cmpl $(PAGE_SIZE - (VEC_SIZE * 4)), %eax
> > > - jg L(cross_page)
> > > - /* Start comparing 4 vectors. */
> > > - vmovdqu (%rdi), %ymm1
> > > - VPCMPEQ (%rsi), %ymm1, %ymm0
> > > - VPMINU %ymm1, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm0, %ymm0
> > > - vpmovmskb %ymm0, %ecx
> > > - testl %ecx, %ecx
> > > - je L(next_3_vectors)
> > > - tzcntl %ecx, %edx
> > > + sall $20, %eax
> > > + /* Check if s1 or s2 may cross a page in next 4x VEC loads. */
> > > + cmpl $((PAGE_SIZE -(VEC_SIZE * 4)) << 20), %eax
> > > + ja L(page_cross)
> > > +
> > > +L(no_page_cross):
> > > + /* Safe to compare 4x vectors. */
> > > + VMOVU (%rdi), %ymm0
> > > + /* 1s where s1 and s2 equal. */
> > > + VPCMPEQ (%rsi), %ymm0, %ymm1
> > > + /* 1s at null CHAR. */
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + /* 1s where s1 and s2 equal AND not null CHAR. */
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > +
> > > + /* All 1s -> keep going, any 0s -> return. */
> > > + vpmovmskb %ymm1, %ecx
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the mismatched index (%rdx) is after the maximum
> > > - offset (%r11). */
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > + cmpq $VEC_SIZE, %rdx
> > > + jbe L(vec_0_test_len)
> > > # endif
> > > +
> > > + /* All 1s represents all equals. incl will overflow to zero in
> > > + all equals case. Otherwise 1s will carry until position of first
> > > + mismatch. */
> > > + incl %ecx
> > > + jz L(more_3x_vec)
> > > +
> > > + .p2align 4,, 4
> > > +L(return_vec_0):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_WCSCMP
> > > + movl (%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (%rdi, %rdx), %ecx
> > > - cmpl (%rsi, %rdx), %ecx
> > > - je L(return)
> > > -L(wcscmp_return):
> > > + cmpl (%rsi, %rcx), %edx
> > > + je L(ret0)
> > > setl %al
> > > negl %eax
> > > orl $1, %eax
> > > -L(return):
> > > # else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > + movzbl (%rdi, %rcx), %eax
> > > + movzbl (%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > # endif
> > > +L(ret0):
> > > L(return_vzeroupper):
> > > ZERO_UPPER_VEC_REGISTERS_RETURN
> > >
> > > - .p2align 4
> > > -L(return_vec_size):
> > > - tzcntl %ecx, %edx
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the mismatched index (%rdx + VEC_SIZE) is after
> > > - the maximum offset (%r11). */
> > > - addq $VEC_SIZE, %rdx
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > + .p2align 4,, 8
> > > +L(vec_0_test_len):
> > > + notl %ecx
> > > + bzhil %edx, %ecx, %eax
> > > + jnz L(return_vec_0)
> > > + /* Align if will cross fetch block. */
> > > + .p2align 4,, 2
> > > +L(ret_zero):
> > > xorl %eax, %eax
> > > - movl (%rdi, %rdx), %ecx
> > > - cmpl (%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > -# else
> > > + VZEROUPPER_RETURN
> > > +
> > > + .p2align 4,, 5
> > > +L(one_or_less):
> > > + jb L(ret_zero)
> > > # ifdef USE_AS_WCSCMP
> > > + /* 'nbe' covers the case where length is negative (large
> > > + unsigned). */
> > > + jnbe __wcscmp_avx2
> > > + movl (%rdi), %edx
> > > xorl %eax, %eax
> > > - movl VEC_SIZE(%rdi, %rdx), %ecx
> > > - cmpl VEC_SIZE(%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > + cmpl (%rsi), %edx
> > > + je L(ret1)
> > > + setl %al
> > > + negl %eax
> > > + orl $1, %eax
> > > # else
> > > - movzbl VEC_SIZE(%rdi, %rdx), %eax
> > > - movzbl VEC_SIZE(%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > + /* 'nbe' covers the case where length is negative (large
> > > + unsigned). */
> > > +
> > > + jnbe __strcmp_avx2
> > > + movzbl (%rdi), %eax
> > > + movzbl (%rsi), %ecx
> > > + subl %ecx, %eax
> > > # endif
> > > +L(ret1):
> > > + ret
> > > # endif
> > > - VZEROUPPER_RETURN
> > >
> > > - .p2align 4
> > > -L(return_2_vec_size):
> > > - tzcntl %ecx, %edx
> > > + .p2align 4,, 10
> > > +L(return_vec_1):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the mismatched index (%rdx + 2 * VEC_SIZE) is
> > > - after the maximum offset (%r11). */
> > > - addq $(VEC_SIZE * 2), %rdx
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > + /* rdx must be > CHAR_PER_VEC so save to subtract w.o fear of
> > > + overflow. */
> > > + addq $-VEC_SIZE, %rdx
> > > + cmpq %rcx, %rdx
> > > + jbe L(ret_zero)
> > > +# endif
> > > +# ifdef USE_AS_WCSCMP
> > > + movl VEC_SIZE(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (%rdi, %rdx), %ecx
> > > - cmpl (%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl VEC_SIZE(%rsi, %rcx), %edx
> > > + je L(ret2)
> > > + setl %al
> > > + negl %eax
> > > + orl $1, %eax
> > > # else
> > > -# ifdef USE_AS_WCSCMP
> > > - xorl %eax, %eax
> > > - movl (VEC_SIZE * 2)(%rdi, %rdx), %ecx
> > > - cmpl (VEC_SIZE * 2)(%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (VEC_SIZE * 2)(%rdi, %rdx), %eax
> > > - movzbl (VEC_SIZE * 2)(%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + movzbl VEC_SIZE(%rdi, %rcx), %eax
> > > + movzbl VEC_SIZE(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > # endif
> > > +L(ret2):
> > > VZEROUPPER_RETURN
> > >
> > > - .p2align 4
> > > -L(return_3_vec_size):
> > > - tzcntl %ecx, %edx
> > > + .p2align 4,, 10
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the mismatched index (%rdx + 3 * VEC_SIZE) is
> > > - after the maximum offset (%r11). */
> > > - addq $(VEC_SIZE * 3), %rdx
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > +L(return_vec_3):
> > > + salq $32, %rcx
> > > +# endif
> > > +
> > > +L(return_vec_2):
> > > +# ifndef USE_AS_STRNCMP
> > > + tzcntl %ecx, %ecx
> > > +# else
> > > + tzcntq %rcx, %rcx
> > > + cmpq %rcx, %rdx
> > > + jbe L(ret_zero)
> > > +# endif
> > > +
> > > +# ifdef USE_AS_WCSCMP
> > > + movl (VEC_SIZE * 2)(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (%rdi, %rdx), %ecx
> > > - cmpl (%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl (VEC_SIZE * 2)(%rsi, %rcx), %edx
> > > + je L(ret3)
> > > + setl %al
> > > + negl %eax
> > > + orl $1, %eax
> > > # else
> > > + movzbl (VEC_SIZE * 2)(%rdi, %rcx), %eax
> > > + movzbl (VEC_SIZE * 2)(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > +# endif
> > > +L(ret3):
> > > + VZEROUPPER_RETURN
> > > +
> > > +# ifndef USE_AS_STRNCMP
> > > + .p2align 4,, 10
> > > +L(return_vec_3):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_WCSCMP
> > > + movl (VEC_SIZE * 3)(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (VEC_SIZE * 3)(%rdi, %rdx), %ecx
> > > - cmpl (VEC_SIZE * 3)(%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > + cmpl (VEC_SIZE * 3)(%rsi, %rcx), %edx
> > > + je L(ret4)
> > > + setl %al
> > > + negl %eax
> > > + orl $1, %eax
> > > # else
> > > - movzbl (VEC_SIZE * 3)(%rdi, %rdx), %eax
> > > - movzbl (VEC_SIZE * 3)(%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > + movzbl (VEC_SIZE * 3)(%rdi, %rcx), %eax
> > > + movzbl (VEC_SIZE * 3)(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > # endif
> > > -# endif
> > > +L(ret4):
> > > VZEROUPPER_RETURN
> > > +# endif
> > > +
> > > + .p2align 4,, 10
> > > +L(more_3x_vec):
> > > + /* Safe to compare 4x vectors. */
> > > + VMOVU VEC_SIZE(%rdi), %ymm0
> > > + VPCMPEQ VEC_SIZE(%rsi), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_1)
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + subq $(VEC_SIZE * 2), %rdx
> > > + jbe L(ret_zero)
> > > +# endif
> > > +
> > > + VMOVU (VEC_SIZE * 2)(%rdi), %ymm0
> > > + VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_2)
> > > +
> > > + VMOVU (VEC_SIZE * 3)(%rdi), %ymm0
> > > + VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_3)
> > >
> > > - .p2align 4
> > > -L(next_3_vectors):
> > > - vmovdqu VEC_SIZE(%rdi), %ymm6
> > > - VPCMPEQ VEC_SIZE(%rsi), %ymm6, %ymm3
> > > - VPMINU %ymm6, %ymm3, %ymm3
> > > - VPCMPEQ %ymm7, %ymm3, %ymm3
> > > - vpmovmskb %ymm3, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(return_vec_size)
> > > - vmovdqu (VEC_SIZE * 2)(%rdi), %ymm5
> > > - vmovdqu (VEC_SIZE * 3)(%rdi), %ymm4
> > > - vmovdqu (VEC_SIZE * 3)(%rsi), %ymm0
> > > - VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm5, %ymm2
> > > - VPMINU %ymm5, %ymm2, %ymm2
> > > - VPCMPEQ %ymm4, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm2, %ymm2
> > > - vpmovmskb %ymm2, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(return_2_vec_size)
> > > - VPMINU %ymm4, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm0, %ymm0
> > > - vpmovmskb %ymm0, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(return_3_vec_size)
> > > -L(main_loop_header):
> > > - leaq (VEC_SIZE * 4)(%rdi), %rdx
> > > - movl $PAGE_SIZE, %ecx
> > > - /* Align load via RAX. */
> > > - andq $-(VEC_SIZE * 4), %rdx
> > > - subq %rdi, %rdx
> > > - leaq (%rdi, %rdx), %rax
> > > # ifdef USE_AS_STRNCMP
> > > - /* Starting from this point, the maximum offset, or simply the
> > > - 'offset', DECREASES by the same amount when base pointers are
> > > - moved forward. Return 0 when:
> > > - 1) On match: offset <= the matched vector index.
> > > - 2) On mistmach, offset is before the mistmatched index.
> > > + cmpq $(VEC_SIZE * 2), %rdx
> > > + jbe L(ret_zero)
> > > +# endif
> > > +
> > > +# ifdef USE_AS_WCSCMP
> > > + /* any non-zero positive value that doesn't inference with 0x1.
> > > */
> > > - subq %rdx, %r11
> > > - jbe L(zero)
> > > -# endif
> > > - addq %rsi, %rdx
> > > - movq %rdx, %rsi
> > > - andl $(PAGE_SIZE - 1), %esi
> > > - /* Number of bytes before page crossing. */
> > > - subq %rsi, %rcx
> > > - /* Number of VEC_SIZE * 4 blocks before page crossing. */
> > > - shrq $DIVIDE_BY_VEC_4_SHIFT, %rcx
> > > - /* ESI: Number of VEC_SIZE * 4 blocks before page crossing. */
> > > - movl %ecx, %esi
> > > - jmp L(loop_start)
> > > + movl $2, %r8d
> > >
> > > +# else
> > > + xorl %r8d, %r8d
> > > +# endif
> > > +
> > > + /* The prepare labels are various entry points from the page
> > > + cross logic. */
> > > +L(prepare_loop):
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + /* Store N + (VEC_SIZE * 4) and place check at the begining of
> > > + the loop. */
> > > + leaq (VEC_SIZE * 2)(%rdi, %rdx), %rdx
> > > +# endif
> > > +L(prepare_loop_no_len):
> > > +
> > > + /* Align s1 and adjust s2 accordingly. */
> > > + subq %rdi, %rsi
> > > + andq $-(VEC_SIZE * 4), %rdi
> > > + addq %rdi, %rsi
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + subq %rdi, %rdx
> > > +# endif
> > > +
> > > +L(prepare_loop_aligned):
> > > + /* eax stores distance from rsi to next page cross. These cases
> > > + need to be handled specially as the 4x loop could potentially
> > > + read memory past the length of s1 or s2 and across a page
> > > + boundary. */
> > > + movl $-(VEC_SIZE * 4), %eax
> > > + subl %esi, %eax
> > > + andl $(PAGE_SIZE - 1), %eax
> > > +
> > > + /* Loop 4x comparisons at a time. */
> > > .p2align 4
> > > L(loop):
> > > +
> > > + /* End condition for strncmp. */
> > > # ifdef USE_AS_STRNCMP
> > > - /* Base pointers are moved forward by 4 * VEC_SIZE. Decrease
> > > - the maximum offset (%r11) by the same amount. */
> > > - subq $(VEC_SIZE * 4), %r11
> > > - jbe L(zero)
> > > -# endif
> > > - addq $(VEC_SIZE * 4), %rax
> > > - addq $(VEC_SIZE * 4), %rdx
> > > -L(loop_start):
> > > - testl %esi, %esi
> > > - leal -1(%esi), %esi
> > > - je L(loop_cross_page)
> > > -L(back_to_loop):
> > > - /* Main loop, comparing 4 vectors are a time. */
> > > - vmovdqa (%rax), %ymm0
> > > - vmovdqa VEC_SIZE(%rax), %ymm3
> > > - VPCMPEQ (%rdx), %ymm0, %ymm4
> > > - VPCMPEQ VEC_SIZE(%rdx), %ymm3, %ymm1
> > > - VPMINU %ymm0, %ymm4, %ymm4
> > > - VPMINU %ymm3, %ymm1, %ymm1
> > > - vmovdqa (VEC_SIZE * 2)(%rax), %ymm2
> > > - VPMINU %ymm1, %ymm4, %ymm0
> > > - vmovdqa (VEC_SIZE * 3)(%rax), %ymm3
> > > - VPCMPEQ (VEC_SIZE * 2)(%rdx), %ymm2, %ymm5
> > > - VPCMPEQ (VEC_SIZE * 3)(%rdx), %ymm3, %ymm6
> > > - VPMINU %ymm2, %ymm5, %ymm5
> > > - VPMINU %ymm3, %ymm6, %ymm6
> > > - VPMINU %ymm5, %ymm0, %ymm0
> > > - VPMINU %ymm6, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm0, %ymm0
> > > -
> > > - /* Test each mask (32 bits) individually because for VEC_SIZE
> > > - == 32 is not possible to OR the four masks and keep all bits
> > > - in a 64-bit integer register, differing from SSE2 strcmp
> > > - where ORing is possible. */
> > > - vpmovmskb %ymm0, %ecx
> > > + subq $(VEC_SIZE * 4), %rdx
> > > + jbe L(ret_zero)
> > > +# endif
> > > +
> > > + subq $-(VEC_SIZE * 4), %rdi
> > > + subq $-(VEC_SIZE * 4), %rsi
> > > +
> > > + /* Check if rsi loads will cross a page boundary. */
> > > + addl $-(VEC_SIZE * 4), %eax
> > > + jnb L(page_cross_during_loop)
> > > +
> > > + /* Loop entry after handling page cross during loop. */
> > > +L(loop_skip_page_cross_check):
> > > + VMOVA (VEC_SIZE * 0)(%rdi), %ymm0
> > > + VMOVA (VEC_SIZE * 1)(%rdi), %ymm2
> > > + VMOVA (VEC_SIZE * 2)(%rdi), %ymm4
> > > + VMOVA (VEC_SIZE * 3)(%rdi), %ymm6
> > > +
> > > + /* ymm1 all 1s where s1 and s2 equal. All 0s otherwise. */
> > > + VPCMPEQ (VEC_SIZE * 0)(%rsi), %ymm0, %ymm1
> > > +
> > > + VPCMPEQ (VEC_SIZE * 1)(%rsi), %ymm2, %ymm3
> > > + VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm4, %ymm5
> > > + VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm6, %ymm7
> > > +
> > > +
> > > + /* If any mismatches or null CHAR then 0 CHAR, otherwise non-
> > > + zero. */
> > > + vpand %ymm0, %ymm1, %ymm1
> > > +
> > > +
> > > + vpand %ymm2, %ymm3, %ymm3
> > > + vpand %ymm4, %ymm5, %ymm5
> > > + vpand %ymm6, %ymm7, %ymm7
> > > +
> > > + VPMINU %ymm1, %ymm3, %ymm3
> > > + VPMINU %ymm5, %ymm7, %ymm7
> > > +
> > > + /* Reduce all 0 CHARs for the 4x VEC into ymm7. */
> > > + VPMINU %ymm3, %ymm7, %ymm7
> > > +
> > > + /* If any 0 CHAR then done. */
> > > + VPCMPEQ %ymm7, %ymmZERO, %ymm7
> > > + vpmovmskb %ymm7, %LOOP_REG
> > > + testl %LOOP_REG, %LOOP_REG
> > > + jz L(loop)
> > > +
> > > + /* Find which VEC has the mismatch of end of string. */
> > > + VPCMPEQ %ymm1, %ymmZERO, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > testl %ecx, %ecx
> > > - je L(loop)
> > > - VPCMPEQ %ymm7, %ymm4, %ymm0
> > > - vpmovmskb %ymm0, %edi
> > > - testl %edi, %edi
> > > - je L(test_vec)
> > > - tzcntl %edi, %ecx
> > > + jnz L(return_vec_0_end)
> > > +
> > > +
> > > + VPCMPEQ %ymm3, %ymmZERO, %ymm3
> > > + vpmovmskb %ymm3, %ecx
> > > + testl %ecx, %ecx
> > > + jnz L(return_vec_1_end)
> > > +
> > > +L(return_vec_2_3_end):
> > > # ifdef USE_AS_STRNCMP
> > > - cmpq %rcx, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > + subq $(VEC_SIZE * 2), %rdx
> > > + jbe L(ret_zero_end)
> > > +# endif
> > > +
> > > + VPCMPEQ %ymm5, %ymmZERO, %ymm5
> > > + vpmovmskb %ymm5, %ecx
> > > + testl %ecx, %ecx
> > > + jnz L(return_vec_2_end)
> > > +
> > > + /* LOOP_REG contains matches for null/mismatch from the loop. If
> > > + VEC 0,1,and 2 all have no null and no mismatches then mismatch
> > > + must entirely be from VEC 3 which is fully represented by
> > > + LOOP_REG. */
> > > + tzcntl %LOOP_REG, %LOOP_REG
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + subl $-(VEC_SIZE), %LOOP_REG
> > > + cmpq %LOOP_REG64, %rdx
> > > + jbe L(ret_zero_end)
> > > +# endif
> > > +
> > > +# ifdef USE_AS_WCSCMP
> > > + movl (VEC_SIZE * 2 - VEC_OFFSET)(%rdi, %LOOP_REG64), %ecx
> > > xorl %eax, %eax
> > > - movl (%rsi, %rcx), %edi
> > > - cmpl (%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl (VEC_SIZE * 2 - VEC_OFFSET)(%rsi, %LOOP_REG64), %ecx
> > > + je L(ret5)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > # else
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > - xorl %eax, %eax
> > > - movl (%rsi, %rcx), %edi
> > > - cmpl (%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + movzbl (VEC_SIZE * 2 - VEC_OFFSET)(%rdi, %LOOP_REG64), %eax
> > > + movzbl (VEC_SIZE * 2 - VEC_OFFSET)(%rsi, %LOOP_REG64), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > +L(ret5):
> > > VZEROUPPER_RETURN
> > >
> > > - .p2align 4
> > > -L(test_vec):
> > > # ifdef USE_AS_STRNCMP
> > > - /* The first vector matched. Return 0 if the maximum offset
> > > - (%r11) <= VEC_SIZE. */
> > > - cmpq $VEC_SIZE, %r11
> > > - jbe L(zero)
> > > + .p2align 4,, 2
> > > +L(ret_zero_end):
> > > + xorl %eax, %eax
> > > + VZEROUPPER_RETURN
> > > # endif
> > > - VPCMPEQ %ymm7, %ymm1, %ymm1
> > > - vpmovmskb %ymm1, %ecx
> > > - testl %ecx, %ecx
> > > - je L(test_2_vec)
> > > - tzcntl %ecx, %edi
> > > +
> > > +
> > > + /* The L(return_vec_N_end) differ from L(return_vec_N) in that
> > > + they use the value of `r8` to negate the return value. This is
> > > + because the page cross logic can swap `rdi` and `rsi`. */
> > > + .p2align 4,, 10
> > > # ifdef USE_AS_STRNCMP
> > > - addq $VEC_SIZE, %rdi
> > > - cmpq %rdi, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > +L(return_vec_1_end):
> > > + salq $32, %rcx
> > > +# endif
> > > +L(return_vec_0_end):
> > > +# ifndef USE_AS_STRNCMP
> > > + tzcntl %ecx, %ecx
> > > +# else
> > > + tzcntq %rcx, %rcx
> > > + cmpq %rcx, %rdx
> > > + jbe L(ret_zero_end)
> > > +# endif
> > > +
> > > +# ifdef USE_AS_WCSCMP
> > > + movl (%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (%rsi, %rdi), %ecx
> > > - cmpl (%rdx, %rdi), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rdi), %eax
> > > - movzbl (%rdx, %rdi), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl (%rsi, %rcx), %edx
> > > + je L(ret6)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > # else
> > > + movzbl (%rdi, %rcx), %eax
> > > + movzbl (%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > +# endif
> > > +L(ret6):
> > > + VZEROUPPER_RETURN
> > > +
> > > +# ifndef USE_AS_STRNCMP
> > > + .p2align 4,, 10
> > > +L(return_vec_1_end):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > + movl VEC_SIZE(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl VEC_SIZE(%rsi, %rdi), %ecx
> > > - cmpl VEC_SIZE(%rdx, %rdi), %ecx
> > > - jne L(wcscmp_return)
> > > + cmpl VEC_SIZE(%rsi, %rcx), %edx
> > > + je L(ret7)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > # else
> > > - movzbl VEC_SIZE(%rax, %rdi), %eax
> > > - movzbl VEC_SIZE(%rdx, %rdi), %edx
> > > - subl %edx, %eax
> > > + movzbl VEC_SIZE(%rdi, %rcx), %eax
> > > + movzbl VEC_SIZE(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > -# endif
> > > +L(ret7):
> > > VZEROUPPER_RETURN
> > > +# endif
> > >
> > > - .p2align 4
> > > -L(test_2_vec):
> > > + .p2align 4,, 10
> > > +L(return_vec_2_end):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_STRNCMP
> > > - /* The first 2 vectors matched. Return 0 if the maximum offset
> > > - (%r11) <= 2 * VEC_SIZE. */
> > > - cmpq $(VEC_SIZE * 2), %r11
> > > - jbe L(zero)
> > > + cmpq %rcx, %rdx
> > > + jbe L(ret_zero_page_cross)
> > > # endif
> > > - VPCMPEQ %ymm7, %ymm5, %ymm5
> > > - vpmovmskb %ymm5, %ecx
> > > - testl %ecx, %ecx
> > > - je L(test_3_vec)
> > > - tzcntl %ecx, %edi
> > > -# ifdef USE_AS_STRNCMP
> > > - addq $(VEC_SIZE * 2), %rdi
> > > - cmpq %rdi, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > +# ifdef USE_AS_WCSCMP
> > > + movl (VEC_SIZE * 2)(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (%rsi, %rdi), %ecx
> > > - cmpl (%rdx, %rdi), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rdi), %eax
> > > - movzbl (%rdx, %rdi), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl (VEC_SIZE * 2)(%rsi, %rcx), %edx
> > > + je L(ret11)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > # else
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > - xorl %eax, %eax
> > > - movl (VEC_SIZE * 2)(%rsi, %rdi), %ecx
> > > - cmpl (VEC_SIZE * 2)(%rdx, %rdi), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (VEC_SIZE * 2)(%rax, %rdi), %eax
> > > - movzbl (VEC_SIZE * 2)(%rdx, %rdi), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + movzbl (VEC_SIZE * 2)(%rdi, %rcx), %eax
> > > + movzbl (VEC_SIZE * 2)(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > +L(ret11):
> > > VZEROUPPER_RETURN
> > >
> > > - .p2align 4
> > > -L(test_3_vec):
> > > +
> > > + /* Page cross in rsi in next 4x VEC. */
> > > +
> > > + /* TODO: Improve logic here. */
> > > + .p2align 4,, 10
> > > +L(page_cross_during_loop):
> > > + /* eax contains [distance_from_page - (VEC_SIZE * 4)]. */
> > > +
> > > + /* Optimistically rsi and rdi and both aligned inwhich case we
> > > + don't need any logic here. */
> > > + cmpl $-(VEC_SIZE * 4), %eax
> > > + /* Don't adjust eax before jumping back to loop and we will
> > > + never hit page cross case again. */
> > > + je L(loop_skip_page_cross_check)
> > > +
> > > + /* Check if we can safely load a VEC. */
> > > + cmpl $-(VEC_SIZE * 3), %eax
> > > + jle L(less_1x_vec_till_page_cross)
> > > +
> > > + VMOVA (%rdi), %ymm0
> > > + VPCMPEQ (%rsi), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_0_end)
> > > +
> > > + /* if distance >= 2x VEC then eax > -(VEC_SIZE * 2). */
> > > + cmpl $-(VEC_SIZE * 2), %eax
> > > + jg L(more_2x_vec_till_page_cross)
> > > +
> > > + .p2align 4,, 4
> > > +L(less_1x_vec_till_page_cross):
> > > + subl $-(VEC_SIZE * 4), %eax
> > > + /* Guranteed safe to read from rdi - VEC_SIZE here. The only
> > > + concerning case is first iteration if incoming s1 was near start
> > > + of a page and s2 near end. If s1 was near the start of the page
> > > + we already aligned up to nearest VEC_SIZE * 4 so gurnateed safe
> > > + to read back -VEC_SIZE. If rdi is truly at the start of a page
> > > + here, it means the previous page (rdi - VEC_SIZE) has already
> > > + been loaded earlier so must be valid. */
> > > + VMOVU -VEC_SIZE(%rdi, %rax), %ymm0
> > > + VPCMPEQ -VEC_SIZE(%rsi, %rax), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > +
> > > + /* Mask of potentially valid bits. The lower bits can be out of
> > > + range comparisons (but safe regarding page crosses). */
> > > + movl $-1, %r10d
> > > + shlxl %esi, %r10d, %r10d
> > > + notl %ecx
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - /* The first 3 vectors matched. Return 0 if the maximum offset
> > > - (%r11) <= 3 * VEC_SIZE. */
> > > - cmpq $(VEC_SIZE * 3), %r11
> > > - jbe L(zero)
> > > -# endif
> > > - VPCMPEQ %ymm7, %ymm6, %ymm6
> > > - vpmovmskb %ymm6, %esi
> > > - tzcntl %esi, %ecx
> > > + cmpq %rax, %rdx
> > > + jbe L(return_page_cross_end_check)
> > > +# endif
> > > + movl %eax, %OFFSET_REG
> > > + addl $(PAGE_SIZE - VEC_SIZE * 4), %eax
> > > +
> > > + andl %r10d, %ecx
> > > + jz L(loop_skip_page_cross_check)
> > > +
> > > + .p2align 4,, 3
> > > +L(return_page_cross_end):
> > > + tzcntl %ecx, %ecx
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - addq $(VEC_SIZE * 3), %rcx
> > > - cmpq %rcx, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > - xorl %eax, %eax
> > > - movl (%rsi, %rcx), %esi
> > > - cmpl (%rdx, %rcx), %esi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + leal -VEC_SIZE(%OFFSET_REG64, %rcx), %ecx
> > > +L(return_page_cross_cmp_mem):
> > > # else
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > + addl %OFFSET_REG, %ecx
> > > +# endif
> > > +# ifdef USE_AS_WCSCMP
> > > + movl VEC_OFFSET(%rdi, %rcx), %edx
> > > xorl %eax, %eax
> > > - movl (VEC_SIZE * 3)(%rsi, %rcx), %esi
> > > - cmpl (VEC_SIZE * 3)(%rdx, %rcx), %esi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (VEC_SIZE * 3)(%rax, %rcx), %eax
> > > - movzbl (VEC_SIZE * 3)(%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + cmpl VEC_OFFSET(%rsi, %rcx), %edx
> > > + je L(ret8)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > +# else
> > > + movzbl VEC_OFFSET(%rdi, %rcx), %eax
> > > + movzbl VEC_OFFSET(%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > +L(ret8):
> > > VZEROUPPER_RETURN
> > >
> > > - .p2align 4
> > > -L(loop_cross_page):
> > > - xorl %r10d, %r10d
> > > - movq %rdx, %rcx
> > > - /* Align load via RDX. We load the extra ECX bytes which should
> > > - be ignored. */
> > > - andl $((VEC_SIZE * 4) - 1), %ecx
> > > - /* R10 is -RCX. */
> > > - subq %rcx, %r10
> > > -
> > > - /* This works only if VEC_SIZE * 2 == 64. */
> > > -# if (VEC_SIZE * 2) != 64
> > > -# error (VEC_SIZE * 2) != 64
> > > -# endif
> > > -
> > > - /* Check if the first VEC_SIZE * 2 bytes should be ignored. */
> > > - cmpl $(VEC_SIZE * 2), %ecx
> > > - jge L(loop_cross_page_2_vec)
> > > -
> > > - vmovdqu (%rax, %r10), %ymm2
> > > - vmovdqu VEC_SIZE(%rax, %r10), %ymm3
> > > - VPCMPEQ (%rdx, %r10), %ymm2, %ymm0
> > > - VPCMPEQ VEC_SIZE(%rdx, %r10), %ymm3, %ymm1
> > > - VPMINU %ymm2, %ymm0, %ymm0
> > > - VPMINU %ymm3, %ymm1, %ymm1
> > > - VPCMPEQ %ymm7, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm1, %ymm1
> > > -
> > > - vpmovmskb %ymm0, %edi
> > > - vpmovmskb %ymm1, %esi
> > > -
> > > - salq $32, %rsi
> > > - xorq %rsi, %rdi
> > > -
> > > - /* Since ECX < VEC_SIZE * 2, simply skip the first ECX bytes. */
> > > - shrq %cl, %rdi
> > > -
> > > - testq %rdi, %rdi
> > > - je L(loop_cross_page_2_vec)
> > > - tzcntq %rdi, %rcx
> > > # ifdef USE_AS_STRNCMP
> > > - cmpq %rcx, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > + .p2align 4,, 10
> > > +L(return_page_cross_end_check):
> > > + tzcntl %ecx, %ecx
> > > + leal -VEC_SIZE(%rax, %rcx), %ecx
> > > + cmpl %ecx, %edx
> > > + ja L(return_page_cross_cmp_mem)
> > > xorl %eax, %eax
> > > - movl (%rsi, %rcx), %edi
> > > - cmpl (%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > -# else
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > - xorl %eax, %eax
> > > - movl (%rsi, %rcx), %edi
> > > - cmpl (%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > -# endif
> > > VZEROUPPER_RETURN
> > > +# endif
> > >
> > > - .p2align 4
> > > -L(loop_cross_page_2_vec):
> > > - /* The first VEC_SIZE * 2 bytes match or are ignored. */
> > > - vmovdqu (VEC_SIZE * 2)(%rax, %r10), %ymm2
> > > - vmovdqu (VEC_SIZE * 3)(%rax, %r10), %ymm3
> > > - VPCMPEQ (VEC_SIZE * 2)(%rdx, %r10), %ymm2, %ymm5
> > > - VPMINU %ymm2, %ymm5, %ymm5
> > > - VPCMPEQ (VEC_SIZE * 3)(%rdx, %r10), %ymm3, %ymm6
> > > - VPCMPEQ %ymm7, %ymm5, %ymm5
> > > - VPMINU %ymm3, %ymm6, %ymm6
> > > - VPCMPEQ %ymm7, %ymm6, %ymm6
> > > -
> > > - vpmovmskb %ymm5, %edi
> > > - vpmovmskb %ymm6, %esi
> > > -
> > > - salq $32, %rsi
> > > - xorq %rsi, %rdi
> > >
> > > - xorl %r8d, %r8d
> > > - /* If ECX > VEC_SIZE * 2, skip ECX - (VEC_SIZE * 2) bytes. */
> > > - subl $(VEC_SIZE * 2), %ecx
> > > - jle 1f
> > > - /* Skip ECX bytes. */
> > > - shrq %cl, %rdi
> > > - /* R8 has number of bytes skipped. */
> > > - movl %ecx, %r8d
> > > -1:
> > > - /* Before jumping back to the loop, set ESI to the number of
> > > - VEC_SIZE * 4 blocks before page crossing. */
> > > - movl $(PAGE_SIZE / (VEC_SIZE * 4) - 1), %esi
> > > -
> > > - testq %rdi, %rdi
> > > + .p2align 4,, 10
> > > +L(more_2x_vec_till_page_cross):
> > > + /* If more 2x vec till cross we will complete a full loop
> > > + iteration here. */
> > > +
> > > + VMOVU VEC_SIZE(%rdi), %ymm0
> > > + VPCMPEQ VEC_SIZE(%rsi), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_1_end)
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - /* At this point, if %rdi value is 0, it already tested
> > > - VEC_SIZE*4+%r10 byte starting from %rax. This label
> > > - checks whether strncmp maximum offset reached or not. */
> > > - je L(string_nbyte_offset_check)
> > > -# else
> > > - je L(back_to_loop)
> > > + cmpq $(VEC_SIZE * 2), %rdx
> > > + jbe L(ret_zero_in_loop_page_cross)
> > > # endif
> > > - tzcntq %rdi, %rcx
> > > - addq %r10, %rcx
> > > - /* Adjust for number of bytes skipped. */
> > > - addq %r8, %rcx
> > > +
> > > + subl $-(VEC_SIZE * 4), %eax
> > > +
> > > + /* Safe to include comparisons from lower bytes. */
> > > + VMOVU -(VEC_SIZE * 2)(%rdi, %rax), %ymm0
> > > + VPCMPEQ -(VEC_SIZE * 2)(%rsi, %rax), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_page_cross_0)
> > > +
> > > + VMOVU -(VEC_SIZE * 1)(%rdi, %rax), %ymm0
> > > + VPCMPEQ -(VEC_SIZE * 1)(%rsi, %rax), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > + jnz L(return_vec_page_cross_1)
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - addq $(VEC_SIZE * 2), %rcx
> > > - subq %rcx, %r11
> > > - jbe L(zero)
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > + /* Must check length here as length might proclude reading next
> > > + page. */
> > > + cmpq %rax, %rdx
> > > + jbe L(ret_zero_in_loop_page_cross)
> > > +# endif
> > > +
> > > + /* Finish the loop. */
> > > + VMOVA (VEC_SIZE * 2)(%rdi), %ymm4
> > > + VMOVA (VEC_SIZE * 3)(%rdi), %ymm6
> > > +
> > > + VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm4, %ymm5
> > > + VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm6, %ymm7
> > > + vpand %ymm4, %ymm5, %ymm5
> > > + vpand %ymm6, %ymm7, %ymm7
> > > + VPMINU %ymm5, %ymm7, %ymm7
> > > + VPCMPEQ %ymm7, %ymmZERO, %ymm7
> > > + vpmovmskb %ymm7, %LOOP_REG
> > > + testl %LOOP_REG, %LOOP_REG
> > > + jnz L(return_vec_2_3_end)
> > > +
> > > + /* Best for code size to include ucond-jmp here. Would be faster
> > > + if this case is hot to duplicate the L(return_vec_2_3_end) code
> > > + as fall-through and have jump back to loop on mismatch
> > > + comparison. */
> > > + subq $-(VEC_SIZE * 4), %rdi
> > > + subq $-(VEC_SIZE * 4), %rsi
> > > + addl $(PAGE_SIZE - VEC_SIZE * 8), %eax
> > > +# ifdef USE_AS_STRNCMP
> > > + subq $(VEC_SIZE * 4), %rdx
> > > + ja L(loop_skip_page_cross_check)
> > > +L(ret_zero_in_loop_page_cross):
> > > xorl %eax, %eax
> > > - movl (%rsi, %rcx), %edi
> > > - cmpl (%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rax, %rcx), %eax
> > > - movzbl (%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + VZEROUPPER_RETURN
> > > # else
> > > -# ifdef USE_AS_WCSCMP
> > > - movq %rax, %rsi
> > > - xorl %eax, %eax
> > > - movl (VEC_SIZE * 2)(%rsi, %rcx), %edi
> > > - cmpl (VEC_SIZE * 2)(%rdx, %rcx), %edi
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (VEC_SIZE * 2)(%rax, %rcx), %eax
> > > - movzbl (VEC_SIZE * 2)(%rdx, %rcx), %edx
> > > - subl %edx, %eax
> > > -# endif
> > > + jmp L(loop_skip_page_cross_check)
> > > # endif
> > > - VZEROUPPER_RETURN
> > >
> > > +
> > > + .p2align 4,, 10
> > > +L(return_vec_page_cross_0):
> > > + addl $-VEC_SIZE, %eax
> > > +L(return_vec_page_cross_1):
> > > + tzcntl %ecx, %ecx
> > > # ifdef USE_AS_STRNCMP
> > > -L(string_nbyte_offset_check):
> > > - leaq (VEC_SIZE * 4)(%r10), %r10
> > > - cmpq %r10, %r11
> > > - jbe L(zero)
> > > - jmp L(back_to_loop)
> > > + leal -VEC_SIZE(%rax, %rcx), %ecx
> > > + cmpq %rcx, %rdx
> > > + jbe L(ret_zero_in_loop_page_cross)
> > > +# else
> > > + addl %eax, %ecx
> > > # endif
> > >
> > > - .p2align 4
> > > -L(cross_page_loop):
> > > - /* Check one byte/dword at a time. */
> > > # ifdef USE_AS_WCSCMP
> > > - cmpl %ecx, %eax
> > > + movl VEC_OFFSET(%rdi, %rcx), %edx
> > > + xorl %eax, %eax
> > > + cmpl VEC_OFFSET(%rsi, %rcx), %edx
> > > + je L(ret9)
> > > + setl %al
> > > + negl %eax
> > > + xorl %r8d, %eax
> > > # else
> > > + movzbl VEC_OFFSET(%rdi, %rcx), %eax
> > > + movzbl VEC_OFFSET(%rsi, %rcx), %ecx
> > > subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > - jne L(different)
> > > - addl $SIZE_OF_CHAR, %edx
> > > - cmpl $(VEC_SIZE * 4), %edx
> > > - je L(main_loop_header)
> > > -# ifdef USE_AS_STRNCMP
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > +L(ret9):
> > > + VZEROUPPER_RETURN
> > > +
> > > +
> > > + .p2align 4,, 10
> > > +L(page_cross):
> > > +# ifndef USE_AS_STRNCMP
> > > + /* If both are VEC aligned we don't need any special logic here.
> > > + Only valid for strcmp where stop condition is guranteed to be
> > > + reachable by just reading memory. */
> > > + testl $((VEC_SIZE - 1) << 20), %eax
> > > + jz L(no_page_cross)
> > > # endif
> > > +
> > > + movl %edi, %eax
> > > + movl %esi, %ecx
> > > + andl $(PAGE_SIZE - 1), %eax
> > > + andl $(PAGE_SIZE - 1), %ecx
> > > +
> > > + xorl %OFFSET_REG, %OFFSET_REG
> > > +
> > > + /* Check which is closer to page cross, s1 or s2. */
> > > + cmpl %eax, %ecx
> > > + jg L(page_cross_s2)
> > > +
> > > + /* The previous page cross check has false positives. Check for
> > > + true positive as page cross logic is very expensive. */
> > > + subl $(PAGE_SIZE - VEC_SIZE * 4), %eax
> > > + jbe L(no_page_cross)
> > > +
> > > + /* Set r8 to not interfere with normal return value (rdi and rsi
> > > + did not swap). */
> > > # ifdef USE_AS_WCSCMP
> > > - movl (%rdi, %rdx), %eax
> > > - movl (%rsi, %rdx), %ecx
> > > + /* any non-zero positive value that doesn't inference with 0x1.
> > > + */
> > > + movl $2, %r8d
> > > # else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %ecx
> > > + xorl %r8d, %r8d
> > > # endif
> > > - /* Check null char. */
> > > - testl %eax, %eax
> > > - jne L(cross_page_loop)
> > > - /* Since %eax == 0, subtract is OK for both SIGNED and UNSIGNED
> > > - comparisons. */
> > > - subl %ecx, %eax
> > > -# ifndef USE_AS_WCSCMP
> > > -L(different):
> > > +
> > > + /* Check if less than 1x VEC till page cross. */
> > > + subl $(VEC_SIZE * 3), %eax
> > > + jg L(less_1x_vec_till_page)
> > > +
> > > + /* If more than 1x VEC till page cross, loop throuh safely
> > > + loadable memory until within 1x VEC of page cross. */
> > > +
> > > + .p2align 4,, 10
> > > +L(page_cross_loop):
> > > +
> > > + VMOVU (%rdi, %OFFSET_REG64), %ymm0
> > > + VPCMPEQ (%rsi, %OFFSET_REG64), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incl %ecx
> > > +
> > > + jnz L(check_ret_vec_page_cross)
> > > + addl $VEC_SIZE, %OFFSET_REG
> > > +# ifdef USE_AS_STRNCMP
> > > + cmpq %OFFSET_REG64, %rdx
> > > + jbe L(ret_zero_page_cross)
> > > # endif
> > > - VZEROUPPER_RETURN
> > > + addl $VEC_SIZE, %eax
> > > + jl L(page_cross_loop)
> > > +
> > > + subl %eax, %OFFSET_REG
> > > + /* OFFSET_REG has distance to page cross - VEC_SIZE. Guranteed
> > > + to not cross page so is safe to load. Since we have already
> > > + loaded at least 1 VEC from rsi it is also guranteed to be safe.
> > > + */
> > > +
> > > + VMOVU (%rdi, %OFFSET_REG64), %ymm0
> > > + VPCMPEQ (%rsi, %OFFSET_REG64), %ymm0, %ymm1
> > > + VPCMPEQ %ymm0, %ymmZERO, %ymm2
> > > + vpandn %ymm1, %ymm2, %ymm1
> > > + vpmovmskb %ymm1, %ecx
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + leal VEC_SIZE(%OFFSET_REG64), %eax
> > > + cmpq %rax, %rdx
> > > + jbe L(check_ret_vec_page_cross2)
> > > + addq %rdi, %rdx
> > > +# endif
> > > + incl %ecx
> > > + jz L(prepare_loop_no_len)
> > >
> > > + .p2align 4,, 4
> > > +L(ret_vec_page_cross):
> > > +# ifndef USE_AS_STRNCMP
> > > +L(check_ret_vec_page_cross):
> > > +# endif
> > > + tzcntl %ecx, %ecx
> > > + addl %OFFSET_REG, %ecx
> > > +L(ret_vec_page_cross_cont):
> > > # ifdef USE_AS_WCSCMP
> > > - .p2align 4
> > > -L(different):
> > > - /* Use movl to avoid modifying EFLAGS. */
> > > - movl $0, %eax
> > > + movl (%rdi, %rcx), %edx
> > > + xorl %eax, %eax
> > > + cmpl (%rsi, %rcx), %edx
> > > + je L(ret12)
> > > setl %al
> > > negl %eax
> > > - orl $1, %eax
> > > - VZEROUPPER_RETURN
> > > + xorl %r8d, %eax
> > > +# else
> > > + movzbl (%rdi, %rcx), %eax
> > > + movzbl (%rsi, %rcx), %ecx
> > > + subl %ecx, %eax
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > # endif
> > > +L(ret12):
> > > + VZEROUPPER_RETURN
> > >
> > > # ifdef USE_AS_STRNCMP
> > > - .p2align 4
> > > -L(zero):
> > > + .p2align 4,, 10
> > > +L(check_ret_vec_page_cross2):
> > > + incl %ecx
> > > +L(check_ret_vec_page_cross):
> > > + tzcntl %ecx, %ecx
> > > + addl %OFFSET_REG, %ecx
> > > + cmpq %rcx, %rdx
> > > + ja L(ret_vec_page_cross_cont)
> > > + .p2align 4,, 2
> > > +L(ret_zero_page_cross):
> > > xorl %eax, %eax
> > > VZEROUPPER_RETURN
> > > +# endif
> > >
> > > - .p2align 4
> > > -L(char0):
> > > -# ifdef USE_AS_WCSCMP
> > > - xorl %eax, %eax
> > > - movl (%rdi), %ecx
> > > - cmpl (%rsi), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rsi), %ecx
> > > - movzbl (%rdi), %eax
> > > - subl %ecx, %eax
> > > -# endif
> > > - VZEROUPPER_RETURN
> > > + .p2align 4,, 4
> > > +L(page_cross_s2):
> > > + /* Ensure this is a true page cross. */
> > > + subl $(PAGE_SIZE - VEC_SIZE * 4), %ecx
> > > + jbe L(no_page_cross)
> > > +
> > > +
> > > + movl %ecx, %eax
> > > + movq %rdi, %rcx
> > > + movq %rsi, %rdi
> > > + movq %rcx, %rsi
> > > +
> > > + /* set r8 to negate return value as rdi and rsi swapped. */
> > > +# ifdef USE_AS_WCSCMP
> > > + movl $-4, %r8d
> > > +# else
> > > + movl $-1, %r8d
> > > # endif
> > > + xorl %OFFSET_REG, %OFFSET_REG
> > >
> > > - .p2align 4
> > > -L(last_vector):
> > > - addq %rdx, %rdi
> > > - addq %rdx, %rsi
> > > + /* Check if more than 1x VEC till page cross. */
> > > + subl $(VEC_SIZE * 3), %eax
> > > + jle L(page_cross_loop)
> > > +
> > > + .p2align 4,, 6
> > > +L(less_1x_vec_till_page):
> > > + /* Find largest load size we can use. */
> > > + cmpl $16, %eax
> > > + ja L(less_16_till_page)
> > > +
> > > + VMOVU (%rdi), %xmm0
> > > + VPCMPEQ (%rsi), %xmm0, %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incw %cx
> > > + jnz L(check_ret_vec_page_cross)
> > > + movl $16, %OFFSET_REG
> > > # ifdef USE_AS_STRNCMP
> > > - subq %rdx, %r11
> > > + cmpq %OFFSET_REG64, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case0)
> > > + subl %eax, %OFFSET_REG
> > > +# else
> > > + /* Explicit check for 16 byte alignment. */
> > > + subl %eax, %OFFSET_REG
> > > + jz L(prepare_loop)
> > > # endif
> > > - tzcntl %ecx, %edx
> > > +
> > > + VMOVU (%rdi, %OFFSET_REG64), %xmm0
> > > + VPCMPEQ (%rsi, %OFFSET_REG64), %xmm0, %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incw %cx
> > > + jnz L(check_ret_vec_page_cross)
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > + addl $16, %OFFSET_REG
> > > + subq %OFFSET_REG64, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case0)
> > > + subq $-(VEC_SIZE * 4), %rdx
> > > +
> > > + leaq -(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq -(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > +# else
> > > + leaq (16 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq (16 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > # endif
> > > -# ifdef USE_AS_WCSCMP
> > > + jmp L(prepare_loop_aligned)
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + .p2align 4,, 2
> > > +L(ret_zero_page_cross_slow_case0):
> > > xorl %eax, %eax
> > > - movl (%rdi, %rdx), %ecx
> > > - cmpl (%rsi, %rdx), %ecx
> > > - jne L(wcscmp_return)
> > > -# else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %edx
> > > - subl %edx, %eax
> > > + ret
> > > # endif
> > > - VZEROUPPER_RETURN
> > >
> > > - /* Comparing on page boundary region requires special treatment:
> > > - It must done one vector at the time, starting with the wider
> > > - ymm vector if possible, if not, with xmm. If fetching 16 bytes
> > > - (xmm) still passes the boundary, byte comparison must be done.
> > > - */
> > > - .p2align 4
> > > -L(cross_page):
> > > - /* Try one ymm vector at a time. */
> > > - cmpl $(PAGE_SIZE - VEC_SIZE), %eax
> > > - jg L(cross_page_1_vector)
> > > -L(loop_1_vector):
> > > - vmovdqu (%rdi, %rdx), %ymm1
> > > - VPCMPEQ (%rsi, %rdx), %ymm1, %ymm0
> > > - VPMINU %ymm1, %ymm0, %ymm0
> > > - VPCMPEQ %ymm7, %ymm0, %ymm0
> > > - vpmovmskb %ymm0, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(last_vector)
> > >
> > > - addl $VEC_SIZE, %edx
> > > + .p2align 4,, 10
> > > +L(less_16_till_page):
> > > + /* Find largest load size we can use. */
> > > + cmpl $24, %eax
> > > + ja L(less_8_till_page)
> > >
> > > - addl $VEC_SIZE, %eax
> > > -# ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the current offset (%rdx) >= the maximum offset
> > > - (%r11). */
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# endif
> > > - cmpl $(PAGE_SIZE - VEC_SIZE), %eax
> > > - jle L(loop_1_vector)
> > > -L(cross_page_1_vector):
> > > - /* Less than 32 bytes to check, try one xmm vector. */
> > > - cmpl $(PAGE_SIZE - 16), %eax
> > > - jg L(cross_page_1_xmm)
> > > - vmovdqu (%rdi, %rdx), %xmm1
> > > - VPCMPEQ (%rsi, %rdx), %xmm1, %xmm0
> > > - VPMINU %xmm1, %xmm0, %xmm0
> > > - VPCMPEQ %xmm7, %xmm0, %xmm0
> > > - vpmovmskb %xmm0, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(last_vector)
> > > + vmovq (%rdi), %xmm0
> > > + vmovq (%rsi), %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + VPCMPEQ %xmm1, %xmm0, %xmm1
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incb %cl
> > > + jnz L(check_ret_vec_page_cross)
> > >
> > > - addl $16, %edx
> > > -# ifndef USE_AS_WCSCMP
> > > - addl $16, %eax
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + cmpq $8, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case0)
> > > # endif
> > > + movl $24, %OFFSET_REG
> > > + /* Explicit check for 16 byte alignment. */
> > > + subl %eax, %OFFSET_REG
> > > +
> > > +
> > > +
> > > + vmovq (%rdi, %OFFSET_REG64), %xmm0
> > > + vmovq (%rsi, %OFFSET_REG64), %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + VPCMPEQ %xmm1, %xmm0, %xmm1
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + incb %cl
> > > + jnz L(check_ret_vec_page_cross)
> > > +
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the current offset (%rdx) >= the maximum offset
> > > - (%r11). */
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# endif
> > > -
> > > -L(cross_page_1_xmm):
> > > -# ifndef USE_AS_WCSCMP
> > > - /* Less than 16 bytes to check, try 8 byte vector. NB: No need
> > > - for wcscmp nor wcsncmp since wide char is 4 bytes. */
> > > - cmpl $(PAGE_SIZE - 8), %eax
> > > - jg L(cross_page_8bytes)
> > > - vmovq (%rdi, %rdx), %xmm1
> > > - vmovq (%rsi, %rdx), %xmm0
> > > - VPCMPEQ %xmm0, %xmm1, %xmm0
> > > - VPMINU %xmm1, %xmm0, %xmm0
> > > - VPCMPEQ %xmm7, %xmm0, %xmm0
> > > - vpmovmskb %xmm0, %ecx
> > > - /* Only last 8 bits are valid. */
> > > - andl $0xff, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(last_vector)
> > > + addl $8, %OFFSET_REG
> > > + subq %OFFSET_REG64, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case0)
> > > + subq $-(VEC_SIZE * 4), %rdx
> > >
> > > - addl $8, %edx
> > > - addl $8, %eax
> > > + leaq -(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq -(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > +# else
> > > + leaq (8 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq (8 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > +# endif
> > > + jmp L(prepare_loop_aligned)
> > > +
> > > +
> > > + .p2align 4,, 10
> > > +L(less_8_till_page):
> > > +# ifdef USE_AS_WCSCMP
> > > + /* If using wchar then this is the only check before we reach
> > > + the page boundary. */
> > > + movl (%rdi), %eax
> > > + movl (%rsi), %ecx
> > > + cmpl %ecx, %eax
> > > + jnz L(ret_less_8_wcs)
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the current offset (%rdx) >= the maximum offset
> > > - (%r11). */
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > + addq %rdi, %rdx
> > > + /* We already checked for len <= 1 so cannot hit that case here.
> > > + */
> > > # endif
> > > + testl %eax, %eax
> > > + jnz L(prepare_loop_no_len)
> > > + ret
> > >
> > > -L(cross_page_8bytes):
> > > - /* Less than 8 bytes to check, try 4 byte vector. */
> > > - cmpl $(PAGE_SIZE - 4), %eax
> > > - jg L(cross_page_4bytes)
> > > - vmovd (%rdi, %rdx), %xmm1
> > > - vmovd (%rsi, %rdx), %xmm0
> > > - VPCMPEQ %xmm0, %xmm1, %xmm0
> > > - VPMINU %xmm1, %xmm0, %xmm0
> > > - VPCMPEQ %xmm7, %xmm0, %xmm0
> > > - vpmovmskb %xmm0, %ecx
> > > - /* Only last 4 bits are valid. */
> > > - andl $0xf, %ecx
> > > - testl %ecx, %ecx
> > > - jne L(last_vector)
> > > + .p2align 4,, 8
> > > +L(ret_less_8_wcs):
> > > + setl %OFFSET_REG8
> > > + negl %OFFSET_REG
> > > + movl %OFFSET_REG, %eax
> > > + xorl %r8d, %eax
> > > + ret
> > > +
> > > +# else
> > > +
> > > + /* Find largest load size we can use. */
> > > + cmpl $28, %eax
> > > + ja L(less_4_till_page)
> > > +
> > > + vmovd (%rdi), %xmm0
> > > + vmovd (%rsi), %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + VPCMPEQ %xmm1, %xmm0, %xmm1
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + subl $0xf, %ecx
> > > + jnz L(check_ret_vec_page_cross)
> > >
> > > - addl $4, %edx
> > > # ifdef USE_AS_STRNCMP
> > > - /* Return 0 if the current offset (%rdx) >= the maximum offset
> > > - (%r11). */
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > + cmpq $4, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case1)
> > > # endif
> > > + movl $28, %OFFSET_REG
> > > + /* Explicit check for 16 byte alignment. */
> > > + subl %eax, %OFFSET_REG
> > >
> > > -L(cross_page_4bytes):
> > > -# endif
> > > - /* Less than 4 bytes to check, try one byte/dword at a time. */
> > > -# ifdef USE_AS_STRNCMP
> > > - cmpq %r11, %rdx
> > > - jae L(zero)
> > > -# endif
> > > -# ifdef USE_AS_WCSCMP
> > > - movl (%rdi, %rdx), %eax
> > > - movl (%rsi, %rdx), %ecx
> > > -# else
> > > - movzbl (%rdi, %rdx), %eax
> > > - movzbl (%rsi, %rdx), %ecx
> > > -# endif
> > > - testl %eax, %eax
> > > - jne L(cross_page_loop)
> > > +
> > > +
> > > + vmovd (%rdi, %OFFSET_REG64), %xmm0
> > > + vmovd (%rsi, %OFFSET_REG64), %xmm1
> > > + VPCMPEQ %xmm0, %xmmZERO, %xmm2
> > > + VPCMPEQ %xmm1, %xmm0, %xmm1
> > > + vpandn %xmm1, %xmm2, %xmm1
> > > + vpmovmskb %ymm1, %ecx
> > > + subl $0xf, %ecx
> > > + jnz L(check_ret_vec_page_cross)
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + addl $4, %OFFSET_REG
> > > + subq %OFFSET_REG64, %rdx
> > > + jbe L(ret_zero_page_cross_slow_case1)
> > > + subq $-(VEC_SIZE * 4), %rdx
> > > +
> > > + leaq -(VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq -(VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > +# else
> > > + leaq (4 - VEC_SIZE * 4)(%rdi, %OFFSET_REG64), %rdi
> > > + leaq (4 - VEC_SIZE * 4)(%rsi, %OFFSET_REG64), %rsi
> > > +# endif
> > > + jmp L(prepare_loop_aligned)
> > > +
> > > +# ifdef USE_AS_STRNCMP
> > > + .p2align 4,, 2
> > > +L(ret_zero_page_cross_slow_case1):
> > > + xorl %eax, %eax
> > > + ret
> > > +# endif
> > > +
> > > + .p2align 4,, 10
> > > +L(less_4_till_page):
> > > + subq %rdi, %rsi
> > > + /* Extremely slow byte comparison loop. */
> > > +L(less_4_loop):
> > > + movzbl (%rdi), %eax
> > > + movzbl (%rsi, %rdi), %ecx
> > > subl %ecx, %eax
> > > - VZEROUPPER_RETURN
> > > -END (STRCMP)
> > > + jnz L(ret_less_4_loop)
> > > + testl %ecx, %ecx
> > > + jz L(ret_zero_4_loop)
> > > +# ifdef USE_AS_STRNCMP
> > > + decq %rdx
> > > + jz L(ret_zero_4_loop)
> > > +# endif
> > > + incq %rdi
> > > + /* end condition is reach page boundary (rdi is aligned). */
> > > + testl $31, %edi
> > > + jnz L(less_4_loop)
> > > + leaq -(VEC_SIZE * 4)(%rdi, %rsi), %rsi
> > > + addq $-(VEC_SIZE * 4), %rdi
> > > +# ifdef USE_AS_STRNCMP
> > > + subq $-(VEC_SIZE * 4), %rdx
> > > +# endif
> > > + jmp L(prepare_loop_aligned)
> > > +
> > > +L(ret_zero_4_loop):
> > > + xorl %eax, %eax
> > > + ret
> > > +L(ret_less_4_loop):
> > > + xorl %r8d, %eax
> > > + subl %r8d, %eax
> > > + ret
> > > +# endif
> > > +END(STRCMP)
> > > #endif
> > > --
> > > 2.25.1
> > >
> >
> > LGTM.
>
> Should I wait until 2.36 release to push the optimized versions?
Yes, please.
> There are alot of edge cases with these functions and last time we
> tried to improve them in:
>
> commit c46e9afb2df5fc9e39ff4d13777e4b4c26e04e55
> Author: H.J. Lu <hjl.tools@gmail.com>
> Date: Fri Oct 29 12:40:20 2021 -0700
>
> x86-64: Improve EVEX strcmp with masked load
>
>
> We missed a case:
> https://bugzilla.redhat.com/show_bug.cgi?id=2026399#c19
Is it a correctness bug?
> >
> > Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
> >
> > Thanks.
> >
> > --
> > H.J.
Thanks.
--
H.J.
next prev parent reply other threads:[~2022-01-10 1:59 UTC|newest]
Thread overview: 59+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-01-09 12:29 [PATCH v1 1/5] x86: Optimize strcmp-avx2.S and fix for [BZ# 28755] Noah Goldstein
2022-01-09 12:29 ` [PATCH v1 2/5] x86: Optimize strcmp-evex.S " Noah Goldstein
2022-01-09 12:29 ` [PATCH v1 3/5] string: remove stupid_[strcmp, strncmp, wcscmp, wcsncmp] Noah Goldstein
2022-01-09 12:29 ` [PATCH v1 4/5] string: Improve coverage in test-strcmp.c and test-strncmp.c Noah Goldstein
2022-01-09 12:29 ` [PATCH v1 5/5] benchtests: Add more coverage for strcmp and strncmp benchmarks Noah Goldstein
2022-01-09 12:35 ` [PATCH v1 1/5] x86: Optimize strcmp-avx2.S and fix for [BZ# 28755] Noah Goldstein
2022-01-09 14:07 ` H.J. Lu
2022-01-10 0:29 ` Noah Goldstein
2022-01-10 0:27 ` [PATCH v2 1/7] x86: Fix __wcsncmp_avx2 in strcmp-avx2.S " Noah Goldstein
2022-01-10 0:27 ` [PATCH v2 2/7] x86: Fix __wcsncmp_evex in strcmp-evex.S " Noah Goldstein
2022-01-10 0:35 ` H.J. Lu
2022-01-10 0:27 ` [PATCH v2 3/7] string/test-str*cmp: remove stupid_[strcmp, strncmp, wcscmp, wcsncmp] Noah Goldstein
2022-01-10 0:37 ` H.J. Lu
2022-01-10 0:27 ` [PATCH v2 4/7] string: Improve coverage in test-strcmp.c and test-strncmp.c Noah Goldstein
2022-01-10 0:38 ` H.J. Lu
2022-01-10 2:51 ` Noah Goldstein
2022-01-10 0:27 ` [PATCH v2 5/7] x86: Optimize strcmp-avx2.S Noah Goldstein
2022-01-10 0:41 ` H.J. Lu
2022-01-10 1:06 ` Noah Goldstein
2022-01-10 1:58 ` H.J. Lu [this message]
2022-01-10 2:54 ` Noah Goldstein
2022-01-10 0:27 ` [PATCH v2 6/7] x86: Optimize strcmp-evex.S Noah Goldstein
2022-01-10 0:41 ` H.J. Lu
2022-01-10 0:27 ` [PATCH v2 7/7] benchtests: Add more coverage for strcmp and strncmp benchmarks Noah Goldstein
2022-01-10 0:34 ` [PATCH v2 1/7] x86: Fix __wcsncmp_avx2 in strcmp-avx2.S [BZ# 28755] H.J. Lu
2022-01-10 21:35 ` [PATCH v3 " Noah Goldstein
2022-01-10 21:35 ` [PATCH v3 2/7] x86: Fix __wcsncmp_evex in strcmp-evex.S " Noah Goldstein
2022-01-11 2:15 ` H.J. Lu
2022-01-26 22:04 ` H.J. Lu
2022-04-29 22:05 ` Sunil Pandey
2022-01-10 21:35 ` [PATCH v3 3/7] string/test-str*cmp: remove stupid_[strcmp, strncmp, wcscmp, wcsncmp] Noah Goldstein
2022-01-10 21:35 ` [PATCH v3 4/7] string: Improve coverage in test-strcmp.c and test-strncmp.c Noah Goldstein
2022-01-10 21:35 ` [PATCH v3 5/7] x86: Optimize strcmp-avx2.S Noah Goldstein
2022-02-14 14:10 ` Andreas Schwab
2022-02-14 18:23 ` H.J. Lu
2022-02-14 19:16 ` Andreas Schwab
2022-02-14 19:30 ` H.J. Lu
2022-02-14 19:35 ` Andreas Schwab
2022-02-14 20:59 ` H.J. Lu
2022-02-14 21:10 ` H.J. Lu
2022-02-15 11:11 ` Andreas Schwab
2022-02-15 12:55 ` Andreas Schwab
2022-02-15 12:58 ` Noah Goldstein
2022-02-15 13:09 ` Noah Goldstein
2022-02-15 13:32 ` Noah Goldstein
2022-02-15 13:37 ` Noah Goldstein
2022-02-15 16:33 ` Noah Goldstein
2022-02-14 23:42 ` Noah Goldstein
2022-02-15 10:43 ` Andreas Schwab
2022-02-15 11:22 ` Andreas Schwab
2022-02-15 11:28 ` Noah Goldstein
2022-02-15 12:24 ` Andreas Schwab
2022-01-10 21:35 ` [PATCH v3 6/7] x86: Optimize strcmp-evex.S Noah Goldstein
2022-01-10 21:35 ` [PATCH v3 7/7] benchtests: Add more coverage for strcmp and strncmp benchmarks Noah Goldstein
2022-01-11 2:15 ` [PATCH v3 1/7] x86: Fix __wcsncmp_avx2 in strcmp-avx2.S [BZ# 28755] H.J. Lu
2022-01-26 22:05 ` H.J. Lu
2022-01-27 4:29 ` H.J. Lu
2022-01-27 5:10 ` H.J. Lu
2022-01-27 5:52 ` Noah Goldstein
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAMe9rOrx5ncAmLRaLs_L3YHnqZOD2vXZjYa+fsG0y9V1Sw5fpA@mail.gmail.com \
--to=hjl.tools@gmail.com \
--cc=goldstein.w.n@gmail.com \
--cc=libc-alpha@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).