From: "H.J. Lu" <hjl.tools@gmail.com>
To: Noah Goldstein <goldstein.w.n@gmail.com>
Cc: libc-alpha@sourceware.org, carlos@systemhalted.org
Subject: Re: [PATCH v3 2/7] x86: Shrink / minorly optimize strchr-evex and implement with VMM headers
Date: Wed, 19 Oct 2022 09:53:24 -0700 [thread overview]
Message-ID: <CAMe9rOp90hB_PuXr2nq0+Oo+9jdFDshhAKucxEPmWRmtmfV8gw@mail.gmail.com> (raw)
In-Reply-To: <20221019004409.3623395-2-goldstein.w.n@gmail.com>
On Tue, Oct 18, 2022 at 5:44 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> Size Optimizations:
> 1. Condence hot path for better cache-locality.
> - This is most impact for strchrnul where the logic strings with
> len <= VEC_SIZE or with a match in the first VEC no fits entirely
> in the first cache line.
> 2. Reuse common targets in first 4x VEC and after the loop.
> 3. Don't align targets so aggressively if it doesn't change the number
> of fetch blocks it will require and put more care in avoiding the
> case where targets unnecessarily split cache lines.
> 4. Align the loop better for DSB/LSD
> 5. Use more code-size efficient instructions.
> - tzcnt ... -> bsf ...
> - vpcmpb $0 ... -> vpcmpeq ...
> 6. Align labels less aggressively, especially if it doesn't save fetch
> blocks / causes the basic-block to span extra cache-lines.
>
> Code Size Changes:
> strchr-evex.S : -63 bytes
> strchrnul-evex.S: -48 bytes
>
> Net perf changes:
> Reported as geometric mean of all improvements / regressions from N=10
> runs of the benchtests. Value as New Time / Old Time so < 1.0 is
> improvement and 1.0 is regression.
>
> strchr-evex.S (Fixed) : 0.971
> strchr-evex.S (Rand) : 0.932
> strchrnul-evex.S : 0.965
>
> Full results attached in email.
>
> Full check passes on x86-64.
> ---
> sysdeps/x86_64/multiarch/strchr-evex.S | 558 +++++++++++++++----------
> 1 file changed, 340 insertions(+), 218 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/strchr-evex.S b/sysdeps/x86_64/multiarch/strchr-evex.S
> index a1c15c4419..c2a0d112f7 100644
> --- a/sysdeps/x86_64/multiarch/strchr-evex.S
> +++ b/sysdeps/x86_64/multiarch/strchr-evex.S
> @@ -26,48 +26,75 @@
> # define STRCHR __strchr_evex
> # endif
>
> -# define VMOVU vmovdqu64
> -# define VMOVA vmovdqa64
> +# ifndef VEC_SIZE
> +# include "x86-evex256-vecs.h"
> +# endif
>
> # ifdef USE_AS_WCSCHR
> # define VPBROADCAST vpbroadcastd
> -# define VPCMP vpcmpd
> +# define VPCMP vpcmpd
> +# define VPCMPEQ vpcmpeqd
> # define VPTESTN vptestnmd
> +# define VPTEST vptestmd
> # define VPMINU vpminud
> # define CHAR_REG esi
> -# define SHIFT_REG ecx
> +# define SHIFT_REG rcx
> # define CHAR_SIZE 4
> +
> +# define USE_WIDE_CHAR
> # else
> # define VPBROADCAST vpbroadcastb
> -# define VPCMP vpcmpb
> +# define VPCMP vpcmpb
> +# define VPCMPEQ vpcmpeqb
> # define VPTESTN vptestnmb
> +# define VPTEST vptestmb
> # define VPMINU vpminub
> # define CHAR_REG sil
> -# define SHIFT_REG edx
> +# define SHIFT_REG rdi
> # define CHAR_SIZE 1
> # endif
>
> -# define XMMZERO xmm16
> -
> -# define YMMZERO ymm16
> -# define YMM0 ymm17
> -# define YMM1 ymm18
> -# define YMM2 ymm19
> -# define YMM3 ymm20
> -# define YMM4 ymm21
> -# define YMM5 ymm22
> -# define YMM6 ymm23
> -# define YMM7 ymm24
> -# define YMM8 ymm25
> -
> -# define VEC_SIZE 32
> -# define PAGE_SIZE 4096
> -# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
> -
> - .section .text.evex,"ax",@progbits
> -ENTRY_P2ALIGN (STRCHR, 5)
> - /* Broadcast CHAR to YMM0. */
> - VPBROADCAST %esi, %YMM0
> +# include "reg-macros.h"
> +
> +# if VEC_SIZE == 64
> +# define MASK_GPR rcx
> +# define LOOP_REG rax
> +
> +# define COND_MASK(k_reg) {%k_reg}
> +# else
> +# define MASK_GPR rax
> +# define LOOP_REG rdi
> +
> +# define COND_MASK(k_reg)
> +# endif
> +
> +# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
> +
> +
> +# if CHAR_PER_VEC == 64
> +# define LAST_VEC_OFFSET (VEC_SIZE * 3)
> +# define TESTZ(reg) incq %VGPR_SZ(reg, 64)
> +# else
> +
> +# if CHAR_PER_VEC == 32
> +# define TESTZ(reg) incl %VGPR_SZ(reg, 32)
> +# elif CHAR_PER_VEC == 16
> +# define TESTZ(reg) incw %VGPR_SZ(reg, 16)
> +# else
> +# define TESTZ(reg) incb %VGPR_SZ(reg, 8)
> +# endif
> +
> +# define LAST_VEC_OFFSET (VEC_SIZE * 2)
> +# endif
> +
> +# define VMATCH VMM(0)
> +
> +# define PAGE_SIZE 4096
> +
> + .section SECTION(.text), "ax", @progbits
> +ENTRY_P2ALIGN (STRCHR, 6)
> + /* Broadcast CHAR to VEC_0. */
> + VPBROADCAST %esi, %VMATCH
> movl %edi, %eax
> andl $(PAGE_SIZE - 1), %eax
> /* Check if we cross page boundary with one vector load.
> @@ -75,19 +102,27 @@ ENTRY_P2ALIGN (STRCHR, 5)
> cmpl $(PAGE_SIZE - VEC_SIZE), %eax
> ja L(cross_page_boundary)
>
> +
> /* Check the first VEC_SIZE bytes. Search for both CHAR and the
> null bytes. */
> - VMOVU (%rdi), %YMM1
> -
> + VMOVU (%rdi), %VMM(1)
> /* Leaves only CHARS matching esi as 0. */
> - vpxorq %YMM1, %YMM0, %YMM2
> - VPMINU %YMM2, %YMM1, %YMM2
> - /* Each bit in K0 represents a CHAR or a null byte in YMM1. */
> - VPTESTN %YMM2, %YMM2, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> + vpxorq %VMM(1), %VMATCH, %VMM(2)
> + VPMINU %VMM(2), %VMM(1), %VMM(2)
> + /* Each bit in K0 represents a CHAR or a null byte in VEC_1. */
> + VPTESTN %VMM(2), %VMM(2), %k0
> + KMOV %k0, %VRAX
> +# if VEC_SIZE == 64 && defined USE_AS_STRCHRNUL
> + /* If VEC_SIZE == 64 && STRCHRNUL use bsf to test condition so
> + that all logic for match/null in first VEC first in 1x cache
> + lines. This has a slight cost to larger sizes. */
> + bsf %VRAX, %VRAX
> + jz L(aligned_more)
> +# else
> + test %VRAX, %VRAX
> jz L(aligned_more)
> - tzcntl %eax, %eax
> + bsf %VRAX, %VRAX
> +# endif
> # ifndef USE_AS_STRCHRNUL
> /* Found CHAR or the null byte. */
> cmp (%rdi, %rax, CHAR_SIZE), %CHAR_REG
> @@ -109,287 +144,374 @@ ENTRY_P2ALIGN (STRCHR, 5)
> # endif
> ret
>
> -
> -
> - .p2align 4,, 10
> -L(first_vec_x4):
> -# ifndef USE_AS_STRCHRNUL
> - /* Check to see if first match was CHAR (k0) or null (k1). */
> - kmovd %k0, %eax
> - tzcntl %eax, %eax
> - kmovd %k1, %ecx
> - /* bzhil will not be 0 if first match was null. */
> - bzhil %eax, %ecx, %ecx
> - jne L(zero)
> -# else
> - /* Combine CHAR and null matches. */
> - kord %k0, %k1, %k0
> - kmovd %k0, %eax
> - tzcntl %eax, %eax
> -# endif
> - /* NB: Multiply sizeof char type (1 or 4) to get the number of
> - bytes. */
> - leaq (VEC_SIZE * 4)(%rdi, %rax, CHAR_SIZE), %rax
> - ret
> -
> # ifndef USE_AS_STRCHRNUL
> L(zero):
> xorl %eax, %eax
> ret
> # endif
>
> -
> - .p2align 4
> + .p2align 4,, 2
> +L(first_vec_x3):
> + subq $-(VEC_SIZE * 2), %rdi
> +# if VEC_SIZE == 32
> + /* Reuse L(first_vec_x3) for last VEC2 only for VEC_SIZE == 32.
> + For VEC_SIZE == 64 the registers don't match. */
> +L(last_vec_x2):
> +# endif
> L(first_vec_x1):
> /* Use bsf here to save 1-byte keeping keeping the block in 1x
> fetch block. eax guranteed non-zero. */
> - bsfl %eax, %eax
> + bsf %VRCX, %VRCX
> # ifndef USE_AS_STRCHRNUL
> - /* Found CHAR or the null byte. */
> - cmp (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %CHAR_REG
> + /* Found CHAR or the null byte. */
> + cmp (VEC_SIZE)(%rdi, %rcx, CHAR_SIZE), %CHAR_REG
> jne L(zero)
> -
> # endif
> /* NB: Multiply sizeof char type (1 or 4) to get the number of
> bytes. */
> - leaq (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %rax
> + leaq (VEC_SIZE)(%rdi, %rcx, CHAR_SIZE), %rax
> ret
>
> - .p2align 4,, 10
> + .p2align 4,, 2
> +L(first_vec_x4):
> + subq $-(VEC_SIZE * 2), %rdi
> L(first_vec_x2):
> # ifndef USE_AS_STRCHRNUL
> /* Check to see if first match was CHAR (k0) or null (k1). */
> - kmovd %k0, %eax
> - tzcntl %eax, %eax
> - kmovd %k1, %ecx
> + KMOV %k0, %VRAX
> + tzcnt %VRAX, %VRAX
> + KMOV %k1, %VRCX
> /* bzhil will not be 0 if first match was null. */
> - bzhil %eax, %ecx, %ecx
> + bzhi %VRAX, %VRCX, %VRCX
> jne L(zero)
> # else
> /* Combine CHAR and null matches. */
> - kord %k0, %k1, %k0
> - kmovd %k0, %eax
> - tzcntl %eax, %eax
> + KOR %k0, %k1, %k0
> + KMOV %k0, %VRAX
> + bsf %VRAX, %VRAX
> # endif
> /* NB: Multiply sizeof char type (1 or 4) to get the number of
> bytes. */
> leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
> ret
>
> - .p2align 4,, 10
> -L(first_vec_x3):
> - /* Use bsf here to save 1-byte keeping keeping the block in 1x
> - fetch block. eax guranteed non-zero. */
> - bsfl %eax, %eax
> -# ifndef USE_AS_STRCHRNUL
> - /* Found CHAR or the null byte. */
> - cmp (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %CHAR_REG
> - jne L(zero)
> +# ifdef USE_AS_STRCHRNUL
> + /* We use this as a hook to get imm8 encoding for the jmp to
> + L(page_cross_boundary). This allows the hot case of a
> + match/null-term in first VEC to fit entirely in 1 cache
> + line. */
> +L(cross_page_boundary):
> + jmp L(cross_page_boundary_real)
> # endif
> - /* NB: Multiply sizeof char type (1 or 4) to get the number of
> - bytes. */
> - leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
> - ret
>
> .p2align 4
> L(aligned_more):
> +L(cross_page_continue):
> /* Align data to VEC_SIZE. */
> andq $-VEC_SIZE, %rdi
> -L(cross_page_continue):
> - /* Check the next 4 * VEC_SIZE. Only one VEC_SIZE at a time since
> - data is only aligned to VEC_SIZE. Use two alternating methods
> - for checking VEC to balance latency and port contention. */
>
> - /* This method has higher latency but has better port
> - distribution. */
> - VMOVA (VEC_SIZE)(%rdi), %YMM1
> + /* Check the next 4 * VEC_SIZE. Only one VEC_SIZE at a time
> + since data is only aligned to VEC_SIZE. Use two alternating
> + methods for checking VEC to balance latency and port
> + contention. */
> +
> + /* Method(1) with 8c latency:
> + For VEC_SIZE == 32:
> + p0 * 1.83, p1 * 0.83, p5 * 1.33
> + For VEC_SIZE == 64:
> + p0 * 2.50, p1 * 0.00, p5 * 1.50 */
> + VMOVA (VEC_SIZE)(%rdi), %VMM(1)
> /* Leaves only CHARS matching esi as 0. */
> - vpxorq %YMM1, %YMM0, %YMM2
> - VPMINU %YMM2, %YMM1, %YMM2
> - /* Each bit in K0 represents a CHAR or a null byte in YMM1. */
> - VPTESTN %YMM2, %YMM2, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> + vpxorq %VMM(1), %VMATCH, %VMM(2)
> + VPMINU %VMM(2), %VMM(1), %VMM(2)
> + /* Each bit in K0 represents a CHAR or a null byte in VEC_1. */
> + VPTESTN %VMM(2), %VMM(2), %k0
> + KMOV %k0, %VRCX
> + test %VRCX, %VRCX
> jnz L(first_vec_x1)
>
> - /* This method has higher latency but has better port
> - distribution. */
> - VMOVA (VEC_SIZE * 2)(%rdi), %YMM1
> - /* Each bit in K0 represents a CHAR in YMM1. */
> - VPCMP $0, %YMM1, %YMM0, %k0
> - /* Each bit in K1 represents a CHAR in YMM1. */
> - VPTESTN %YMM1, %YMM1, %k1
> - kortestd %k0, %k1
> + /* Method(2) with 6c latency:
> + For VEC_SIZE == 32:
> + p0 * 1.00, p1 * 0.00, p5 * 2.00
> + For VEC_SIZE == 64:
> + p0 * 1.00, p1 * 0.00, p5 * 2.00 */
> + VMOVA (VEC_SIZE * 2)(%rdi), %VMM(1)
> + /* Each bit in K0 represents a CHAR in VEC_1. */
> + VPCMPEQ %VMM(1), %VMATCH, %k0
> + /* Each bit in K1 represents a CHAR in VEC_1. */
> + VPTESTN %VMM(1), %VMM(1), %k1
> + KORTEST %k0, %k1
> jnz L(first_vec_x2)
>
> - VMOVA (VEC_SIZE * 3)(%rdi), %YMM1
> + /* By swapping between Method 1/2 we get more fair port
> + distrubition and better throughput. */
> +
> + VMOVA (VEC_SIZE * 3)(%rdi), %VMM(1)
> /* Leaves only CHARS matching esi as 0. */
> - vpxorq %YMM1, %YMM0, %YMM2
> - VPMINU %YMM2, %YMM1, %YMM2
> - /* Each bit in K0 represents a CHAR or a null byte in YMM1. */
> - VPTESTN %YMM2, %YMM2, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> + vpxorq %VMM(1), %VMATCH, %VMM(2)
> + VPMINU %VMM(2), %VMM(1), %VMM(2)
> + /* Each bit in K0 represents a CHAR or a null byte in VEC_1. */
> + VPTESTN %VMM(2), %VMM(2), %k0
> + KMOV %k0, %VRCX
> + test %VRCX, %VRCX
> jnz L(first_vec_x3)
>
> - VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> - /* Each bit in K0 represents a CHAR in YMM1. */
> - VPCMP $0, %YMM1, %YMM0, %k0
> - /* Each bit in K1 represents a CHAR in YMM1. */
> - VPTESTN %YMM1, %YMM1, %k1
> - kortestd %k0, %k1
> + VMOVA (VEC_SIZE * 4)(%rdi), %VMM(1)
> + /* Each bit in K0 represents a CHAR in VEC_1. */
> + VPCMPEQ %VMM(1), %VMATCH, %k0
> + /* Each bit in K1 represents a CHAR in VEC_1. */
> + VPTESTN %VMM(1), %VMM(1), %k1
> + KORTEST %k0, %k1
> jnz L(first_vec_x4)
>
> /* Align data to VEC_SIZE * 4 for the loop. */
> +# if VEC_SIZE == 64
> + /* Use rax for the loop reg as it allows to the loop to fit in
> + exactly 2-cache-lines. (more efficient imm32 + gpr
> + encoding). */
> + leaq (VEC_SIZE)(%rdi), %rax
> + /* No partial register stalls on evex512 processors. */
> + xorb %al, %al
> +# else
> + /* For VEC_SIZE == 32 continue using rdi for loop reg so we can
> + reuse more code and save space. */
> addq $VEC_SIZE, %rdi
> andq $-(VEC_SIZE * 4), %rdi
> -
> +# endif
> .p2align 4
> L(loop_4x_vec):
> - /* Check 4x VEC at a time. No penalty to imm32 offset with evex
> - encoding. */
> - VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> - VMOVA (VEC_SIZE * 5)(%rdi), %YMM2
> - VMOVA (VEC_SIZE * 6)(%rdi), %YMM3
> - VMOVA (VEC_SIZE * 7)(%rdi), %YMM4
> -
> - /* For YMM1 and YMM3 use xor to set the CHARs matching esi to
> + /* Check 4x VEC at a time. No penalty for imm32 offset with evex
> + encoding (if offset % VEC_SIZE == 0). */
> + VMOVA (VEC_SIZE * 4)(%LOOP_REG), %VMM(1)
> + VMOVA (VEC_SIZE * 5)(%LOOP_REG), %VMM(2)
> + VMOVA (VEC_SIZE * 6)(%LOOP_REG), %VMM(3)
> + VMOVA (VEC_SIZE * 7)(%LOOP_REG), %VMM(4)
> +
> + /* Collect bits where VEC_1 does NOT match esi. This is later
> + use to mask of results (getting not matches allows us to
> + save an instruction on combining). */
> + VPCMP $4, %VMATCH, %VMM(1), %k1
> +
> + /* Two methods for loop depending on VEC_SIZE. This is because
> + with zmm registers VPMINU can only run on p0 (as opposed to
> + p0/p1 for ymm) so it is less prefered. */
> +# if VEC_SIZE == 32
> + /* For VEC_2 and VEC_3 use xor to set the CHARs matching esi to
> zero. */
> - vpxorq %YMM1, %YMM0, %YMM5
> - /* For YMM2 and YMM4 cmp not equals to CHAR and store result in
> - k register. Its possible to save either 1 or 2 instructions
> - using cmp no equals method for either YMM1 or YMM1 and YMM3
> - respectively but bottleneck on p5 makes it not worth it. */
> - VPCMP $4, %YMM0, %YMM2, %k2
> - vpxorq %YMM3, %YMM0, %YMM7
> - VPCMP $4, %YMM0, %YMM4, %k4
> -
> - /* Use min to select all zeros from either xor or end of string).
> - */
> - VPMINU %YMM1, %YMM5, %YMM1
> - VPMINU %YMM3, %YMM7, %YMM3
> + vpxorq %VMM(2), %VMATCH, %VMM(6)
> + vpxorq %VMM(3), %VMATCH, %VMM(7)
>
> - /* Use min + zeromask to select for zeros. Since k2 and k4 will
> - have 0 as positions that matched with CHAR which will set
> - zero in the corresponding destination bytes in YMM2 / YMM4.
> - */
> - VPMINU %YMM1, %YMM2, %YMM2{%k2}{z}
> - VPMINU %YMM3, %YMM4, %YMM4
> - VPMINU %YMM2, %YMM4, %YMM4{%k4}{z}
> -
> - VPTESTN %YMM4, %YMM4, %k1
> - kmovd %k1, %ecx
> - subq $-(VEC_SIZE * 4), %rdi
> - testl %ecx, %ecx
> + /* Find non-matches in VEC_4 while combining with non-matches
> + from VEC_1. NB: Try and use masked predicate execution on
> + instructions that have mask result as it has no latency
> + penalty. */
> + VPCMP $4, %VMATCH, %VMM(4), %k4{%k1}
> +
> + /* Combined zeros from VEC_1 / VEC_2 (search for null term). */
> + VPMINU %VMM(1), %VMM(2), %VMM(2)
> +
> + /* Use min to select all zeros from either xor or end of
> + string). */
> + VPMINU %VMM(3), %VMM(7), %VMM(3)
> + VPMINU %VMM(2), %VMM(6), %VMM(2)
> +
> + /* Combined zeros from VEC_2 / VEC_3 (search for null term). */
> + VPMINU %VMM(3), %VMM(4), %VMM(4)
> +
> + /* Combined zeros from VEC_2 / VEC_4 (this has all null term and
> + esi matches for VEC_2 / VEC_3). */
> + VPMINU %VMM(2), %VMM(4), %VMM(4)
> +# else
> + /* Collect non-matches for VEC_2. */
> + VPCMP $4, %VMM(2), %VMATCH, %k2
> +
> + /* Combined zeros from VEC_1 / VEC_2 (search for null term). */
> + VPMINU %VMM(1), %VMM(2), %VMM(2)
> +
> + /* Find non-matches in VEC_3/VEC_4 while combining with non-
> + matches from VEC_1/VEC_2 respectively. */
> + VPCMP $4, %VMM(3), %VMATCH, %k3{%k1}
> + VPCMP $4, %VMM(4), %VMATCH, %k4{%k2}
> +
> + /* Finish combining zeros in all VECs. */
> + VPMINU %VMM(3), %VMM(4), %VMM(4)
> +
> + /* Combine in esi matches for VEC_3 (if there was a match with
> + esi, the corresponding bit in %k3 is zero so the
> + VPMINU_MASKZ will have a zero in the result). NB: This make
> + the VPMINU 3c latency. The only way to avoid it is to
> + createa a 12c dependency chain on all the `VPCMP $4, ...`
> + which has higher total latency. */
> + VPMINU %VMM(2), %VMM(4), %VMM(4){%k3}{z}
> +# endif
> + VPTEST %VMM(4), %VMM(4), %k0{%k4}
> + KMOV %k0, %VRDX
> + subq $-(VEC_SIZE * 4), %LOOP_REG
> +
> + /* TESTZ is inc using the proper register width depending on
> + CHAR_PER_VEC. An esi match or null-term match leaves a zero-
> + bit in rdx so inc won't overflow and won't be zero. */
> + TESTZ (rdx)
> jz L(loop_4x_vec)
>
> - VPTESTN %YMM1, %YMM1, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(last_vec_x1)
> + VPTEST %VMM(1), %VMM(1), %k0{%k1}
> + KMOV %k0, %VGPR(MASK_GPR)
> + TESTZ (MASK_GPR)
> +# if VEC_SIZE == 32
> + /* We can reuse the return code in page_cross logic for VEC_SIZE
> + == 32. */
> + jnz L(last_vec_x1_vec_size32)
> +# else
> + jnz L(last_vec_x1_vec_size64)
> +# endif
> +
>
> - VPTESTN %YMM2, %YMM2, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> + /* COND_MASK integates the esi matches for VEC_SIZE == 64. For
> + VEC_SIZE == 32 they are already integrated. */
> + VPTEST %VMM(2), %VMM(2), %k0 COND_MASK(k2)
> + KMOV %k0, %VRCX
> + TESTZ (rcx)
> jnz L(last_vec_x2)
>
> - VPTESTN %YMM3, %YMM3, %k0
> - kmovd %k0, %eax
> - /* Combine YMM3 matches (eax) with YMM4 matches (ecx). */
> -# ifdef USE_AS_WCSCHR
> - sall $8, %ecx
> - orl %ecx, %eax
> - bsfl %eax, %eax
> + VPTEST %VMM(3), %VMM(3), %k0 COND_MASK(k3)
> + KMOV %k0, %VRCX
> +# if CHAR_PER_VEC == 64
> + TESTZ (rcx)
> + jnz L(last_vec_x3)
> # else
> - salq $32, %rcx
> - orq %rcx, %rax
> - bsfq %rax, %rax
> + salq $CHAR_PER_VEC, %rdx
> + TESTZ (rcx)
> + orq %rcx, %rdx
> # endif
> +
> + bsfq %rdx, %rdx
> +
> # ifndef USE_AS_STRCHRNUL
> /* Check if match was CHAR or null. */
> - cmp (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %CHAR_REG
> + cmp (LAST_VEC_OFFSET)(%LOOP_REG, %rdx, CHAR_SIZE), %CHAR_REG
> jne L(zero_end)
> # endif
> /* NB: Multiply sizeof char type (1 or 4) to get the number of
> bytes. */
> - leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
> + leaq (LAST_VEC_OFFSET)(%LOOP_REG, %rdx, CHAR_SIZE), %rax
> ret
>
> - .p2align 4,, 8
> -L(last_vec_x1):
> - bsfl %eax, %eax
> -# ifdef USE_AS_WCSCHR
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes.
> - */
> - leaq (%rdi, %rax, CHAR_SIZE), %rax
> -# else
> - addq %rdi, %rax
> +# ifndef USE_AS_STRCHRNUL
> +L(zero_end):
> + xorl %eax, %eax
> + ret
> # endif
>
> -# ifndef USE_AS_STRCHRNUL
> +
> + /* Seperate return label for last VEC1 because for VEC_SIZE ==
> + 32 we can reuse return code in L(page_cross) but VEC_SIZE ==
> + 64 has mismatched registers. */
> +# if VEC_SIZE == 64
> + .p2align 4,, 8
> +L(last_vec_x1_vec_size64):
> + bsf %VRCX, %VRCX
> +# ifndef USE_AS_STRCHRNUL
> /* Check if match was null. */
> - cmp (%rax), %CHAR_REG
> + cmp (%rax, %rcx, CHAR_SIZE), %CHAR_REG
> jne L(zero_end)
> -# endif
> -
> +# endif
> +# ifdef USE_AS_WCSCHR
> + /* NB: Multiply wchar_t count by 4 to get the number of bytes.
> + */
> + leaq (%rax, %rcx, CHAR_SIZE), %rax
> +# else
> + addq %rcx, %rax
> +# endif
> ret
>
> + /* Since we can't combine the last 2x matches for CHAR_PER_VEC
> + == 64 we need return label for last VEC3. */
> +# if CHAR_PER_VEC == 64
> .p2align 4,, 8
> +L(last_vec_x3):
> + addq $VEC_SIZE, %LOOP_REG
> +# endif
> +
> + /* Duplicate L(last_vec_x2) for VEC_SIZE == 64 because we can't
> + reuse L(first_vec_x3) due to register mismatch. */
> L(last_vec_x2):
> - bsfl %eax, %eax
> -# ifndef USE_AS_STRCHRNUL
> + bsf %VGPR(MASK_GPR), %VGPR(MASK_GPR)
> +# ifndef USE_AS_STRCHRNUL
> /* Check if match was null. */
> - cmp (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %CHAR_REG
> + cmp (VEC_SIZE * 1)(%LOOP_REG, %MASK_GPR, CHAR_SIZE), %CHAR_REG
> jne L(zero_end)
> -# endif
> +# endif
> /* NB: Multiply sizeof char type (1 or 4) to get the number of
> bytes. */
> - leaq (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %rax
> + leaq (VEC_SIZE * 1)(%LOOP_REG, %MASK_GPR, CHAR_SIZE), %rax
> ret
> +# endif
>
> - /* Cold case for crossing page with first load. */
> - .p2align 4,, 8
> + /* Cold case for crossing page with first load. */
> + .p2align 4,, 10
> +# ifndef USE_AS_STRCHRNUL
> L(cross_page_boundary):
> - movq %rdi, %rdx
> +# endif
> +L(cross_page_boundary_real):
> /* Align rdi. */
> - andq $-VEC_SIZE, %rdi
> - VMOVA (%rdi), %YMM1
> - /* Leaves only CHARS matching esi as 0. */
> - vpxorq %YMM1, %YMM0, %YMM2
> - VPMINU %YMM2, %YMM1, %YMM2
> - /* Each bit in K0 represents a CHAR or a null byte in YMM1. */
> - VPTESTN %YMM2, %YMM2, %k0
> - kmovd %k0, %eax
> + xorq %rdi, %rax
> + VMOVA (PAGE_SIZE - VEC_SIZE)(%rax), %VMM(1)
> + /* Use high latency method of getting matches to save code size.
> + */
> +
> + /* K1 has 1s where VEC(1) does NOT match esi. */
> + VPCMP $4, %VMM(1), %VMATCH, %k1
> + /* K0 has ones where K1 is 1 (non-match with esi), and non-zero
> + (null). */
> + VPTEST %VMM(1), %VMM(1), %k0{%k1}
> + KMOV %k0, %VRAX
> /* Remove the leading bits. */
> # ifdef USE_AS_WCSCHR
> - movl %edx, %SHIFT_REG
> + movl %edi, %VGPR_SZ(SHIFT_REG, 32)
> /* NB: Divide shift count by 4 since each bit in K1 represent 4
> bytes. */
> - sarl $2, %SHIFT_REG
> - andl $(CHAR_PER_VEC - 1), %SHIFT_REG
> + sarl $2, %VGPR_SZ(SHIFT_REG, 32)
> + andl $(CHAR_PER_VEC - 1), %VGPR_SZ(SHIFT_REG, 32)
> +
> + /* if wcsrchr we need to reverse matches as we can't rely on
> + signed shift to bring in ones. There is not sarx for
> + gpr8/16. Also not we can't use inc here as the lower bits
> + represent matches out of range so we can't rely on overflow.
> + */
> + xorl $((1 << CHAR_PER_VEC)- 1), %eax
> +# endif
> + /* Use arithmatic shift so that leading 1s are filled in. */
> + sarx %VGPR(SHIFT_REG), %VRAX, %VRAX
> + /* If eax is all ones then no matches for esi or NULL. */
> +
> +# ifdef USE_AS_WCSCHR
> + test %VRAX, %VRAX
> +# else
> + inc %VRAX
> # endif
> - sarxl %SHIFT_REG, %eax, %eax
> - /* If eax is zero continue. */
> - testl %eax, %eax
> jz L(cross_page_continue)
> - bsfl %eax, %eax
>
> + .p2align 4,, 10
> +L(last_vec_x1_vec_size32):
> + bsf %VRAX, %VRAX
> # ifdef USE_AS_WCSCHR
> - /* NB: Multiply wchar_t count by 4 to get the number of
> - bytes. */
> - leaq (%rdx, %rax, CHAR_SIZE), %rax
> + /* NB: Multiply wchar_t count by 4 to get the number of bytes.
> + */
> + leaq (%rdi, %rax, CHAR_SIZE), %rax
> # else
> - addq %rdx, %rax
> + addq %rdi, %rax
> # endif
> # ifndef USE_AS_STRCHRNUL
> /* Check to see if match was CHAR or null. */
> cmp (%rax), %CHAR_REG
> - je L(cross_page_ret)
> -L(zero_end):
> - xorl %eax, %eax
> -L(cross_page_ret):
> + jne L(zero_end_0)
> # endif
> ret
> +# ifndef USE_AS_STRCHRNUL
> +L(zero_end_0):
> + xorl %eax, %eax
> + ret
> +# endif
>
> END (STRCHR)
> #endif
> --
> 2.34.1
>
LGTM.
Thanks.
--
H.J.
next prev parent reply other threads:[~2022-10-19 16:54 UTC|newest]
Thread overview: 41+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-10-18 2:48 [PATCH v1 1/7] x86: Optimize memchr-evex.S " Noah Goldstein
2022-10-18 2:48 ` [PATCH v1 2/7] x86: Shrink / minorly optimize strchr-evex " Noah Goldstein
2022-10-18 2:51 ` Noah Goldstein
2022-10-18 2:48 ` [PATCH v1 3/7] x86: Optimize strnlen-evex.S " Noah Goldstein
2022-10-18 2:51 ` Noah Goldstein
2022-10-18 2:48 ` [PATCH v1 4/7] x86: Optimize memrchr-evex.S Noah Goldstein
2022-10-18 2:51 ` Noah Goldstein
2022-10-18 2:48 ` [PATCH v1 5/7] x86: Optimize strrchr-evex.S and implement with VMM headers Noah Goldstein
2022-10-18 2:52 ` Noah Goldstein
2022-10-18 2:49 ` [PATCH v1 6/7] x86: Add support for VEC_SIZE == 64 in strcmp-evex.S impl Noah Goldstein
2022-10-20 2:15 ` [PATCH v4] " Noah Goldstein
2022-10-20 3:46 ` H.J. Lu
2022-10-18 2:49 ` [PATCH v1 7/7] Bench: Improve benchtests for memchr, strchr, strnlen, strrchr Noah Goldstein
2022-10-18 21:00 ` H.J. Lu
2022-10-18 21:05 ` Noah Goldstein
2022-10-18 21:53 ` H.J. Lu
2022-10-18 22:58 ` Noah Goldstein
2022-10-18 2:50 ` [PATCH v1 1/7] x86: Optimize memchr-evex.S and implement with VMM headers Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 " Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 2/7] x86: Shrink / minorly optimize strchr-evex " Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 3/7] x86: Optimize strnlen-evex.S " Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 4/7] x86: Optimize memrchr-evex.S Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 5/7] x86: Optimize strrchr-evex.S and implement with VMM headers Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 6/7] x86: Add support for VEC_SIZE == 64 in strcmp-evex.S impl Noah Goldstein
2022-10-18 23:19 ` [PATCH v2 7/7] Bench: Improve benchtests for memchr, strchr, strnlen, strrchr Noah Goldstein
2022-10-19 0:01 ` H.J. Lu
2022-10-19 0:44 ` Noah Goldstein
2022-10-19 0:44 ` [PATCH v3 1/7] x86: Optimize memchr-evex.S and implement with VMM headers Noah Goldstein
2022-10-19 0:44 ` [PATCH v3 2/7] x86: Shrink / minorly optimize strchr-evex " Noah Goldstein
2022-10-19 16:53 ` H.J. Lu [this message]
2022-10-19 0:44 ` [PATCH v3 3/7] x86: Optimize strnlen-evex.S " Noah Goldstein
2022-10-19 16:57 ` H.J. Lu
2022-10-19 0:44 ` [PATCH v3 4/7] x86: Optimize memrchr-evex.S Noah Goldstein
2022-10-19 16:58 ` H.J. Lu
2022-10-19 0:44 ` [PATCH v3 5/7] x86: Optimize strrchr-evex.S and implement with VMM headers Noah Goldstein
2022-10-19 16:58 ` H.J. Lu
2022-10-19 0:44 ` [PATCH v3 6/7] x86: Add support for VEC_SIZE == 64 in strcmp-evex.S impl Noah Goldstein
2022-10-19 16:59 ` H.J. Lu
2022-10-19 0:44 ` [PATCH v3 7/7] Bench: Improve benchtests for memchr, strchr, strnlen, strrchr Noah Goldstein
2022-10-19 17:00 ` H.J. Lu
2022-10-19 16:52 ` [PATCH v3 1/7] x86: Optimize memchr-evex.S and implement with VMM headers H.J. Lu
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=CAMe9rOp90hB_PuXr2nq0+Oo+9jdFDshhAKucxEPmWRmtmfV8gw@mail.gmail.com \
--to=hjl.tools@gmail.com \
--cc=carlos@systemhalted.org \
--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).