From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf1-x433.google.com (mail-pf1-x433.google.com [IPv6:2607:f8b0:4864:20::433]) by sourceware.org (Postfix) with ESMTPS id 28CB33857410 for ; Tue, 12 Jul 2022 21:17:02 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org 28CB33857410 Received: by mail-pf1-x433.google.com with SMTP id o12so8525463pfp.5 for ; Tue, 12 Jul 2022 14:17:02 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc; bh=9eVxiGvOh7RKdekXVKELafDEXZrXSGUn0bqOr/K55UU=; b=0vni93yXcCFS3c/+CaX3NbLR8rHEuizPlkvvsJfNCOyaXRlYs2NyFm12bsudc9f8pq CKMQurBsFdJ21YUhL7SmFvUWcKIv9TyaS7RWC8kS1abvVtcM5TkAMOUXv28GvNAVTbkA 1Bx6claE4jFE32GtDl4xmnDB1Qn4+kAPQyVuIZMNzC6CRHfGdhhxtNYW8/4SgjZ3cUVx pGPsLxBp7BAJllE/YKzQFiXUe/+R46oTDFDdNi+miWWM8rbdEZsI0wWNBUCOyShNJIVq vnKKMaJFF53xZY7pAMnFmOED3VbleXDhIKymH4g/t4+ipoz7F9xwqBmQsjplPL0YhtaI ZJQw== X-Gm-Message-State: AJIora/1WznV0/vgY55msSVdpP74XEzTCEv5mqUKxpcUspWFnSzItSD1 UbarTFRYQ5hrgL9T2zCkNQI/amEO+RBqMuw+wWQ= X-Google-Smtp-Source: AGRyM1umyhEFXB+J8AV7VPOYi6vIp/1TobRFfaVIUebRTCnmRoEoSsulsyOzg52eCUCCrtV1NfajhBOiisnIv2j246w= X-Received: by 2002:a05:6a00:14c7:b0:52a:d3ba:74c9 with SMTP id w7-20020a056a0014c700b0052ad3ba74c9mr11554115pfu.1.1657660620758; Tue, 12 Jul 2022 14:17:00 -0700 (PDT) MIME-Version: 1.0 References: <20220712192910.351121-1-goldstein.w.n@gmail.com> <20220712192910.351121-6-goldstein.w.n@gmail.com> In-Reply-To: <20220712192910.351121-6-goldstein.w.n@gmail.com> From: "H.J. Lu" Date: Tue, 12 Jul 2022 14:16:24 -0700 Message-ID: Subject: Re: [PATCH v1] x86: Move strcat SSE2 implementation to multiarch/strcat-sse2.S To: Noah Goldstein Cc: GNU C Library , "Carlos O'Donell" Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3024.7 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, FREEMAIL_FROM, GIT_PATCH_0, KAM_SHORT, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 12 Jul 2022 21:17:05 -0000 On Tue, Jul 12, 2022 at 12:29 PM Noah Goldstein wrote: > > This commit doesn't affect libc.so.6, its just housekeeping to prepare > for adding explicit ISA level support. > > Tested build on x86_64 and x86_32 with/without multiarch. > --- > sysdeps/x86_64/multiarch/strcat-sse2.S | 242 ++++++++++++++++++++++++- > sysdeps/x86_64/strcat.S | 239 +----------------------- > 2 files changed, 238 insertions(+), 243 deletions(-) > > diff --git a/sysdeps/x86_64/multiarch/strcat-sse2.S b/sysdeps/x86_64/multiarch/strcat-sse2.S > index 449e102438..244c4a6d74 100644 > --- a/sysdeps/x86_64/multiarch/strcat-sse2.S > +++ b/sysdeps/x86_64/multiarch/strcat-sse2.S > @@ -17,12 +17,242 @@ > . */ > > #if IS_IN (libc) > +# ifndef STRCAT > +# define STRCAT __strcat_sse2 > +# endif > +#endif > > -# include > -# define strcat __strcat_sse2 > +#include > + > + .text > +ENTRY (STRCAT) > + movq %rdi, %rcx /* Dest. register. */ > + andl $7, %ecx /* mask alignment bits */ > + movq %rdi, %rax /* Duplicate destination pointer. */ > + movq $0xfefefefefefefeff,%r8 > + > + /* First step: Find end of destination. */ > + jz 4f /* aligned => start loop */ > + > + neg %ecx /* We need to align to 8 bytes. */ > + addl $8,%ecx > + /* Search the first bytes directly. */ > +0: cmpb $0x0,(%rax) /* is byte NUL? */ > + je 2f /* yes => start copy */ > + incq %rax /* increment pointer */ > + decl %ecx > + jnz 0b > + > + > + > + /* Now the source is aligned. Scan for NUL byte. */ > + .p2align 4 > +4: > + /* First unroll. */ > + movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > + addq $8,%rax /* adjust pointer for next word */ > + movq %r8, %rdx /* magic value */ > + addq %rcx, %rdx /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 3f /* highest byte is NUL => return pointer */ > + xorq %rcx, %rdx /* (word+magic)^word */ > + orq %r8, %rdx /* set all non-carry bits */ > + incq %rdx /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + jnz 3f /* found NUL => return pointer */ > + > + /* Second unroll. */ > + movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > + addq $8,%rax /* adjust pointer for next word */ > + movq %r8, %rdx /* magic value */ > + addq %rcx, %rdx /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 3f /* highest byte is NUL => return pointer */ > + xorq %rcx, %rdx /* (word+magic)^word */ > + orq %r8, %rdx /* set all non-carry bits */ > + incq %rdx /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + jnz 3f /* found NUL => return pointer */ > + > + /* Third unroll. */ > + movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > + addq $8,%rax /* adjust pointer for next word */ > + movq %r8, %rdx /* magic value */ > + addq %rcx, %rdx /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 3f /* highest byte is NUL => return pointer */ > + xorq %rcx, %rdx /* (word+magic)^word */ > + orq %r8, %rdx /* set all non-carry bits */ > + incq %rdx /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + jnz 3f /* found NUL => return pointer */ > + > + /* Fourth unroll. */ > + movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > + addq $8,%rax /* adjust pointer for next word */ > + movq %r8, %rdx /* magic value */ > + addq %rcx, %rdx /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 3f /* highest byte is NUL => return pointer */ > + xorq %rcx, %rdx /* (word+magic)^word */ > + orq %r8, %rdx /* set all non-carry bits */ > + incq %rdx /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + jz 4b /* no NUL found => continue loop */ > + > + .p2align 4 /* Align, it's a jump target. */ > +3: subq $8,%rax /* correct pointer increment. */ > + > + testb %cl, %cl /* is first byte NUL? */ > + jz 2f /* yes => return */ > + incq %rax /* increment pointer */ > + > + testb %ch, %ch /* is second byte NUL? */ > + jz 2f /* yes => return */ > + incq %rax /* increment pointer */ > + > + testl $0x00ff0000, %ecx /* is third byte NUL? */ > + jz 2f /* yes => return pointer */ > + incq %rax /* increment pointer */ > + > + testl $0xff000000, %ecx /* is fourth byte NUL? */ > + jz 2f /* yes => return pointer */ > + incq %rax /* increment pointer */ > + > + shrq $32, %rcx /* look at other half. */ > + > + testb %cl, %cl /* is first byte NUL? */ > + jz 2f /* yes => return */ > + incq %rax /* increment pointer */ > + > + testb %ch, %ch /* is second byte NUL? */ > + jz 2f /* yes => return */ > + incq %rax /* increment pointer */ > + > + testl $0xff0000, %ecx /* is third byte NUL? */ > + jz 2f /* yes => return pointer */ > + incq %rax /* increment pointer */ > + > +2: > + /* Second step: Copy source to destination. */ > + > + movq %rsi, %rcx /* duplicate */ > + andl $7,%ecx /* mask alignment bits */ > + movq %rax, %rdx /* move around */ > + jz 22f /* aligned => start loop */ > + > + neg %ecx /* align to 8 bytes. */ > + addl $8, %ecx > + /* Align the source pointer. */ > +21: > + movb (%rsi), %al /* Fetch a byte */ > + testb %al, %al /* Is it NUL? */ > + movb %al, (%rdx) /* Store it */ > + jz 24f /* If it was NUL, done! */ > + incq %rsi > + incq %rdx > + decl %ecx > + jnz 21b > + > + /* Now the sources is aligned. Unfortunatly we cannot force > + to have both source and destination aligned, so ignore the > + alignment of the destination. */ > + .p2align 4 > +22: > + /* 1st unroll. */ > + movq (%rsi), %rax /* Read double word (8 bytes). */ > + addq $8, %rsi /* Adjust pointer for next word. */ > + movq %rax, %r9 /* Save a copy for NUL finding. */ > + addq %r8, %r9 /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 23f /* highest byte is NUL => return pointer */ > + xorq %rax, %r9 /* (word+magic)^word */ > + orq %r8, %r9 /* set all non-carry bits */ > + incq %r9 /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + > + jnz 23f /* found NUL => return pointer */ > + > + movq %rax, (%rdx) /* Write value to destination. */ > + addq $8, %rdx /* Adjust pointer. */ > + > + /* 2nd unroll. */ > + movq (%rsi), %rax /* Read double word (8 bytes). */ > + addq $8, %rsi /* Adjust pointer for next word. */ > + movq %rax, %r9 /* Save a copy for NUL finding. */ > + addq %r8, %r9 /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 23f /* highest byte is NUL => return pointer */ > + xorq %rax, %r9 /* (word+magic)^word */ > + orq %r8, %r9 /* set all non-carry bits */ > + incq %r9 /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + > + jnz 23f /* found NUL => return pointer */ > + > + movq %rax, (%rdx) /* Write value to destination. */ > + addq $8, %rdx /* Adjust pointer. */ > + > + /* 3rd unroll. */ > + movq (%rsi), %rax /* Read double word (8 bytes). */ > + addq $8, %rsi /* Adjust pointer for next word. */ > + movq %rax, %r9 /* Save a copy for NUL finding. */ > + addq %r8, %r9 /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 23f /* highest byte is NUL => return pointer */ > + xorq %rax, %r9 /* (word+magic)^word */ > + orq %r8, %r9 /* set all non-carry bits */ > + incq %r9 /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + > + jnz 23f /* found NUL => return pointer */ > + > + movq %rax, (%rdx) /* Write value to destination. */ > + addq $8, %rdx /* Adjust pointer. */ > + > + /* 4th unroll. */ > + movq (%rsi), %rax /* Read double word (8 bytes). */ > + addq $8, %rsi /* Adjust pointer for next word. */ > + movq %rax, %r9 /* Save a copy for NUL finding. */ > + addq %r8, %r9 /* add the magic value to the word. We get > + carry bits reported for each byte which > + is *not* 0 */ > + jnc 23f /* highest byte is NUL => return pointer */ > + xorq %rax, %r9 /* (word+magic)^word */ > + orq %r8, %r9 /* set all non-carry bits */ > + incq %r9 /* add 1: if one carry bit was *not* set > + the addition will not result in 0. */ > + > + jnz 23f /* found NUL => return pointer */ > + > + movq %rax, (%rdx) /* Write value to destination. */ > + addq $8, %rdx /* Adjust pointer. */ > + jmp 22b /* Next iteration. */ > + > + /* Do the last few bytes. %rax contains the value to write. > + The loop is unrolled twice. */ > + .p2align 4 > +23: > + movb %al, (%rdx) /* 1st byte. */ > + testb %al, %al /* Is it NUL. */ > + jz 24f /* yes, finish. */ > + incq %rdx /* Increment destination. */ > + movb %ah, (%rdx) /* 2nd byte. */ > + testb %ah, %ah /* Is it NUL?. */ > + jz 24f /* yes, finish. */ > + incq %rdx /* Increment destination. */ > + shrq $16, %rax /* Shift... */ > + jmp 23b /* and look at next two bytes in %rax. */ > > -# undef libc_hidden_builtin_def > -# define libc_hidden_builtin_def(strcat) > -#endif > > -#include > +24: > + movq %rdi, %rax /* Source is return value. */ > + retq > +END (STRCAT) > diff --git a/sysdeps/x86_64/strcat.S b/sysdeps/x86_64/strcat.S > index 565a9c785a..fc3e8a9bcf 100644 > --- a/sysdeps/x86_64/strcat.S > +++ b/sysdeps/x86_64/strcat.S > @@ -17,241 +17,6 @@ > License along with the GNU C Library; if not, see > . */ > > -#include > -#include "asm-syntax.h" > - > -/* Will be removed when new strcpy implementation gets merged. */ > - > - .text > -ENTRY (strcat) > - movq %rdi, %rcx /* Dest. register. */ > - andl $7, %ecx /* mask alignment bits */ > - movq %rdi, %rax /* Duplicate destination pointer. */ > - movq $0xfefefefefefefeff,%r8 > - > - /* First step: Find end of destination. */ > - jz 4f /* aligned => start loop */ > - > - neg %ecx /* We need to align to 8 bytes. */ > - addl $8,%ecx > - /* Search the first bytes directly. */ > -0: cmpb $0x0,(%rax) /* is byte NUL? */ > - je 2f /* yes => start copy */ > - incq %rax /* increment pointer */ > - decl %ecx > - jnz 0b > - > - > - > - /* Now the source is aligned. Scan for NUL byte. */ > - .p2align 4 > -4: > - /* First unroll. */ > - movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > - addq $8,%rax /* adjust pointer for next word */ > - movq %r8, %rdx /* magic value */ > - addq %rcx, %rdx /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 3f /* highest byte is NUL => return pointer */ > - xorq %rcx, %rdx /* (word+magic)^word */ > - orq %r8, %rdx /* set all non-carry bits */ > - incq %rdx /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - jnz 3f /* found NUL => return pointer */ > - > - /* Second unroll. */ > - movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > - addq $8,%rax /* adjust pointer for next word */ > - movq %r8, %rdx /* magic value */ > - addq %rcx, %rdx /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 3f /* highest byte is NUL => return pointer */ > - xorq %rcx, %rdx /* (word+magic)^word */ > - orq %r8, %rdx /* set all non-carry bits */ > - incq %rdx /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - jnz 3f /* found NUL => return pointer */ > - > - /* Third unroll. */ > - movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > - addq $8,%rax /* adjust pointer for next word */ > - movq %r8, %rdx /* magic value */ > - addq %rcx, %rdx /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 3f /* highest byte is NUL => return pointer */ > - xorq %rcx, %rdx /* (word+magic)^word */ > - orq %r8, %rdx /* set all non-carry bits */ > - incq %rdx /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - jnz 3f /* found NUL => return pointer */ > - > - /* Fourth unroll. */ > - movq (%rax), %rcx /* get double word (= 8 bytes) in question */ > - addq $8,%rax /* adjust pointer for next word */ > - movq %r8, %rdx /* magic value */ > - addq %rcx, %rdx /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 3f /* highest byte is NUL => return pointer */ > - xorq %rcx, %rdx /* (word+magic)^word */ > - orq %r8, %rdx /* set all non-carry bits */ > - incq %rdx /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - jz 4b /* no NUL found => continue loop */ > - > - .p2align 4 /* Align, it's a jump target. */ > -3: subq $8,%rax /* correct pointer increment. */ > - > - testb %cl, %cl /* is first byte NUL? */ > - jz 2f /* yes => return */ > - incq %rax /* increment pointer */ > - > - testb %ch, %ch /* is second byte NUL? */ > - jz 2f /* yes => return */ > - incq %rax /* increment pointer */ > - > - testl $0x00ff0000, %ecx /* is third byte NUL? */ > - jz 2f /* yes => return pointer */ > - incq %rax /* increment pointer */ > - > - testl $0xff000000, %ecx /* is fourth byte NUL? */ > - jz 2f /* yes => return pointer */ > - incq %rax /* increment pointer */ > - > - shrq $32, %rcx /* look at other half. */ > - > - testb %cl, %cl /* is first byte NUL? */ > - jz 2f /* yes => return */ > - incq %rax /* increment pointer */ > - > - testb %ch, %ch /* is second byte NUL? */ > - jz 2f /* yes => return */ > - incq %rax /* increment pointer */ > - > - testl $0xff0000, %ecx /* is third byte NUL? */ > - jz 2f /* yes => return pointer */ > - incq %rax /* increment pointer */ > - > -2: > - /* Second step: Copy source to destination. */ > - > - movq %rsi, %rcx /* duplicate */ > - andl $7,%ecx /* mask alignment bits */ > - movq %rax, %rdx /* move around */ > - jz 22f /* aligned => start loop */ > - > - neg %ecx /* align to 8 bytes. */ > - addl $8, %ecx > - /* Align the source pointer. */ > -21: > - movb (%rsi), %al /* Fetch a byte */ > - testb %al, %al /* Is it NUL? */ > - movb %al, (%rdx) /* Store it */ > - jz 24f /* If it was NUL, done! */ > - incq %rsi > - incq %rdx > - decl %ecx > - jnz 21b > - > - /* Now the sources is aligned. Unfortunatly we cannot force > - to have both source and destination aligned, so ignore the > - alignment of the destination. */ > - .p2align 4 > -22: > - /* 1st unroll. */ > - movq (%rsi), %rax /* Read double word (8 bytes). */ > - addq $8, %rsi /* Adjust pointer for next word. */ > - movq %rax, %r9 /* Save a copy for NUL finding. */ > - addq %r8, %r9 /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 23f /* highest byte is NUL => return pointer */ > - xorq %rax, %r9 /* (word+magic)^word */ > - orq %r8, %r9 /* set all non-carry bits */ > - incq %r9 /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - > - jnz 23f /* found NUL => return pointer */ > - > - movq %rax, (%rdx) /* Write value to destination. */ > - addq $8, %rdx /* Adjust pointer. */ > - > - /* 2nd unroll. */ > - movq (%rsi), %rax /* Read double word (8 bytes). */ > - addq $8, %rsi /* Adjust pointer for next word. */ > - movq %rax, %r9 /* Save a copy for NUL finding. */ > - addq %r8, %r9 /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 23f /* highest byte is NUL => return pointer */ > - xorq %rax, %r9 /* (word+magic)^word */ > - orq %r8, %r9 /* set all non-carry bits */ > - incq %r9 /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - > - jnz 23f /* found NUL => return pointer */ > - > - movq %rax, (%rdx) /* Write value to destination. */ > - addq $8, %rdx /* Adjust pointer. */ > - > - /* 3rd unroll. */ > - movq (%rsi), %rax /* Read double word (8 bytes). */ > - addq $8, %rsi /* Adjust pointer for next word. */ > - movq %rax, %r9 /* Save a copy for NUL finding. */ > - addq %r8, %r9 /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 23f /* highest byte is NUL => return pointer */ > - xorq %rax, %r9 /* (word+magic)^word */ > - orq %r8, %r9 /* set all non-carry bits */ > - incq %r9 /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - > - jnz 23f /* found NUL => return pointer */ > - > - movq %rax, (%rdx) /* Write value to destination. */ > - addq $8, %rdx /* Adjust pointer. */ > - > - /* 4th unroll. */ > - movq (%rsi), %rax /* Read double word (8 bytes). */ > - addq $8, %rsi /* Adjust pointer for next word. */ > - movq %rax, %r9 /* Save a copy for NUL finding. */ > - addq %r8, %r9 /* add the magic value to the word. We get > - carry bits reported for each byte which > - is *not* 0 */ > - jnc 23f /* highest byte is NUL => return pointer */ > - xorq %rax, %r9 /* (word+magic)^word */ > - orq %r8, %r9 /* set all non-carry bits */ > - incq %r9 /* add 1: if one carry bit was *not* set > - the addition will not result in 0. */ > - > - jnz 23f /* found NUL => return pointer */ > - > - movq %rax, (%rdx) /* Write value to destination. */ > - addq $8, %rdx /* Adjust pointer. */ > - jmp 22b /* Next iteration. */ > - > - /* Do the last few bytes. %rax contains the value to write. > - The loop is unrolled twice. */ > - .p2align 4 > -23: > - movb %al, (%rdx) /* 1st byte. */ > - testb %al, %al /* Is it NUL. */ > - jz 24f /* yes, finish. */ > - incq %rdx /* Increment destination. */ > - movb %ah, (%rdx) /* 2nd byte. */ > - testb %ah, %ah /* Is it NUL?. */ > - jz 24f /* yes, finish. */ > - incq %rdx /* Increment destination. */ > - shrq $16, %rax /* Shift... */ > - jmp 23b /* and look at next two bytes in %rax. */ > - > - > -24: > - movq %rdi, %rax /* Source is return value. */ > - retq > -END (strcat) > +#define STRCAT strcat > +#include "multiarch/strcat-sse2.S" > libc_hidden_builtin_def (strcat) > -- > 2.34.1 > LGTM. Thanks. -- H.J.