/* Placeholder function, not used by any processor at the moment. Copyright (C) 2022-2023 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 2.1 of the License, or (at your option) any later version. The GNU C Library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ /* UNUSED. Exists purely as reference implementation. */ #include <isa-level.h> #if ISA_SHOULD_BUILD (4) # include <sysdep.h> # ifdef USE_AS_WCSRCHR # define CHAR_SIZE 4 # define VPBROADCAST vpbroadcastd # define VPCMPEQ vpcmpeqd # define VPMINU vpminud # define VPTESTN vptestnmd # else # define CHAR_SIZE 1 # define VPBROADCAST vpbroadcastb # define VPCMPEQ vpcmpeqb # define VPMINU vpminub # define VPTESTN vptestnmb # endif # define PAGE_SIZE 4096 # define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE) .section SECTION(.text), "ax", @progbits /* Aligning entry point to 64 byte, provides better performance for one vector length string. */ ENTRY_P2ALIGN (STRRCHR, 6) /* Broadcast CHAR to VMM(0). */ VPBROADCAST %esi, %VMM(0) movl %edi, %eax sall $20, %eax cmpl $((PAGE_SIZE - VEC_SIZE) << 20), %eax ja L(page_cross) L(page_cross_continue): /* Compare [w]char for null, mask bit will be set for match. */ VMOVU (%rdi), %VMM(1) VPTESTN %VMM(1), %VMM(1), %k1 KMOV %k1, %VRCX test %VRCX, %VRCX jz L(align_more) VPCMPEQ %VMM(1), %VMM(0), %k0 KMOV %k0, %VRAX BLSMSK %VRCX, %VRCX and %VRCX, %VRAX jz L(ret) BSR %VRAX, %VRAX # ifdef USE_AS_WCSRCHR leaq (%rdi, %rax, CHAR_SIZE), %rax # else add %rdi, %rax # endif L(ret): ret L(vector_x2_end): VPCMPEQ %VMM(2), %VMM(0), %k2 KMOV %k2, %VRAX BLSMSK %VRCX, %VRCX and %VRCX, %VRAX jz L(vector_x1_ret) BSR %VRAX, %VRAX leaq (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %rax ret /* Check the first vector at very last to look for match. */ L(vector_x1_ret): VPCMPEQ %VMM(1), %VMM(0), %k2 KMOV %k2, %VRAX test %VRAX, %VRAX jz L(ret) BSR %VRAX, %VRAX # ifdef USE_AS_WCSRCHR leaq (%rsi, %rax, CHAR_SIZE), %rax # else add %rsi, %rax # endif ret L(align_more): /* Zero r8 to store match result. */ xorl %r8d, %r8d /* Save pointer of first vector, in case if no match found. */ movq %rdi, %rsi /* Align pointer to vector size. */ andq $-VEC_SIZE, %rdi /* Loop unroll for 2 vector loop. */ VMOVA (VEC_SIZE)(%rdi), %VMM(2) VPTESTN %VMM(2), %VMM(2), %k0 KMOV %k0, %VRCX test %VRCX, %VRCX jnz L(vector_x2_end) /* Save pointer of second vector, in case if no match found. */ movq %rdi, %r9 /* Align address to VEC_SIZE * 2 for loop. */ andq $-(VEC_SIZE * 2), %rdi .p2align 4,,11 L(loop): /* 2 vector loop, as it provide better performance as compared to 4 vector loop. */ VMOVA (VEC_SIZE * 2)(%rdi), %VMM(3) VMOVA (VEC_SIZE * 3)(%rdi), %VMM(4) VPCMPEQ %VMM(3), %VMM(0), %k1 VPCMPEQ %VMM(4), %VMM(0), %k2 VPMINU %VMM(3), %VMM(4), %VMM(5) VPTESTN %VMM(5), %VMM(5), %k0 KOR %k1, %k2, %k3 subq $-(VEC_SIZE * 2), %rdi /* If k0 and k3 zero, match and end of string not found. */ KORTEST %k0, %k3 jz L(loop) /* If k0 is non zero, end of string found. */ KORTEST %k0, %k0 jnz L(endloop) lea VEC_SIZE(%rdi), %r8 /* A match found, it need to be stored in r8 before loop continue. */ /* Check second vector first. */ KMOV %k2, %VRDX test %VRDX, %VRDX jnz L(loop_vec_x2_match) KMOV %k1, %VRDX /* Match is in first vector, rdi offset need to be subtracted by VEC_SIZE. */ sub $VEC_SIZE, %r8 /* If second vector doesn't have match, first vector must have match. */ L(loop_vec_x2_match): BSR %VRDX, %VRDX # ifdef USE_AS_WCSRCHR sal $2, %rdx # endif add %rdx, %r8 jmp L(loop) L(endloop): /* Check if string end in first loop vector. */ VPTESTN %VMM(3), %VMM(3), %k0 KMOV %k0, %VRCX test %VRCX, %VRCX jnz L(loop_vector_x1_end) /* Check if it has match in first loop vector. */ KMOV %k1, %VRAX test %VRAX, %VRAX jz L(loop_vector_x2_end) BSR %VRAX, %VRAX leaq (%rdi, %rax, CHAR_SIZE), %r8 /* String must end in second loop vector. */ L(loop_vector_x2_end): VPTESTN %VMM(4), %VMM(4), %k0 KMOV %k0, %VRCX KMOV %k2, %VRAX BLSMSK %VRCX, %VRCX /* Check if it has match in second loop vector. */ and %VRCX, %VRAX jz L(check_last_match) BSR %VRAX, %VRAX leaq (VEC_SIZE)(%rdi, %rax, CHAR_SIZE), %rax ret /* String end in first loop vector. */ L(loop_vector_x1_end): KMOV %k1, %VRAX BLSMSK %VRCX, %VRCX /* Check if it has match in second loop vector. */ and %VRCX, %VRAX jz L(check_last_match) BSR %VRAX, %VRAX leaq (%rdi, %rax, CHAR_SIZE), %rax ret /* No match in first and second loop vector. */ L(check_last_match): /* Check if any match recorded in r8. */ test %r8, %r8 jz L(vector_x2_ret) movq %r8, %rax ret /* No match recorded in r8. Check the second saved vector in beginning. */ L(vector_x2_ret): VPCMPEQ %VMM(2), %VMM(0), %k2 KMOV %k2, %VRAX test %VRAX, %VRAX jz L(vector_x1_ret) /* Match found in the second saved vector. */ BSR %VRAX, %VRAX leaq (VEC_SIZE)(%r9, %rax, CHAR_SIZE), %rax ret L(page_cross): mov %rdi, %rax movl %edi, %ecx # ifdef USE_AS_WCSRCHR /* Calculate number of compare result bits to be skipped for wide string alignment adjustment. */ andl $(VEC_SIZE - 1), %ecx sarl $2, %ecx # endif /* ecx contains number of w[char] to be skipped as a result of address alignment. */ andq $-VEC_SIZE, %rax VMOVA (%rax), %VMM(1) VPTESTN %VMM(1), %VMM(1), %k1 KMOV %k1, %VRAX SHR %cl, %VRAX jz L(page_cross_continue) VPCMPEQ %VMM(1), %VMM(0), %k0 KMOV %k0, %VRDX SHR %cl, %VRDX BLSMSK %VRAX, %VRAX and %VRDX, %VRAX jz L(ret) BSR %VRAX, %VRAX # ifdef USE_AS_WCSRCHR leaq (%rdi, %rax, CHAR_SIZE), %rax # else add %rdi, %rax # endif ret END (STRRCHR) #endif