summaryrefslogtreecommitdiffstats
path: root/arch/x86/crypto/crc32-pclmul_asm.S
diff options
context:
space:
mode:
Diffstat (limited to 'arch/x86/crypto/crc32-pclmul_asm.S')
-rw-r--r--arch/x86/crypto/crc32-pclmul_asm.S218
1 files changed, 218 insertions, 0 deletions
diff --git a/arch/x86/crypto/crc32-pclmul_asm.S b/arch/x86/crypto/crc32-pclmul_asm.S
new file mode 100644
index 0000000000..5d31137e2c
--- /dev/null
+++ b/arch/x86/crypto/crc32-pclmul_asm.S
@@ -0,0 +1,218 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright 2012 Xyratex Technology Limited
+ *
+ * Using hardware provided PCLMULQDQ instruction to accelerate the CRC32
+ * calculation.
+ * CRC32 polynomial:0x04c11db7(BE)/0xEDB88320(LE)
+ * PCLMULQDQ is a new instruction in Intel SSE4.2, the reference can be found
+ * at:
+ * http://www.intel.com/products/processor/manuals/
+ * Intel(R) 64 and IA-32 Architectures Software Developer's Manual
+ * Volume 2B: Instruction Set Reference, N-Z
+ *
+ * Authors: Gregory Prestas <Gregory_Prestas@us.xyratex.com>
+ * Alexander Boyko <Alexander_Boyko@xyratex.com>
+ */
+
+#include <linux/linkage.h>
+
+
+.section .rodata
+.align 16
+/*
+ * [x4*128+32 mod P(x) << 32)]' << 1 = 0x154442bd4
+ * #define CONSTANT_R1 0x154442bd4LL
+ *
+ * [(x4*128-32 mod P(x) << 32)]' << 1 = 0x1c6e41596
+ * #define CONSTANT_R2 0x1c6e41596LL
+ */
+.Lconstant_R2R1:
+ .octa 0x00000001c6e415960000000154442bd4
+/*
+ * [(x128+32 mod P(x) << 32)]' << 1 = 0x1751997d0
+ * #define CONSTANT_R3 0x1751997d0LL
+ *
+ * [(x128-32 mod P(x) << 32)]' << 1 = 0x0ccaa009e
+ * #define CONSTANT_R4 0x0ccaa009eLL
+ */
+.Lconstant_R4R3:
+ .octa 0x00000000ccaa009e00000001751997d0
+/*
+ * [(x64 mod P(x) << 32)]' << 1 = 0x163cd6124
+ * #define CONSTANT_R5 0x163cd6124LL
+ */
+.Lconstant_R5:
+ .octa 0x00000000000000000000000163cd6124
+.Lconstant_mask32:
+ .octa 0x000000000000000000000000FFFFFFFF
+/*
+ * #define CRCPOLY_TRUE_LE_FULL 0x1DB710641LL
+ *
+ * Barrett Reduction constant (u64`) = u` = (x**64 / P(x))` = 0x1F7011641LL
+ * #define CONSTANT_RU 0x1F7011641LL
+ */
+.Lconstant_RUpoly:
+ .octa 0x00000001F701164100000001DB710641
+
+#define CONSTANT %xmm0
+
+#ifdef __x86_64__
+#define BUF %rdi
+#define LEN %rsi
+#define CRC %edx
+#else
+#define BUF %eax
+#define LEN %edx
+#define CRC %ecx
+#endif
+
+
+
+.text
+/**
+ * Calculate crc32
+ * BUF - buffer (16 bytes aligned)
+ * LEN - sizeof buffer (16 bytes aligned), LEN should be grater than 63
+ * CRC - initial crc32
+ * return %eax crc32
+ * uint crc32_pclmul_le_16(unsigned char const *buffer,
+ * size_t len, uint crc32)
+ */
+
+SYM_FUNC_START(crc32_pclmul_le_16) /* buffer and buffer size are 16 bytes aligned */
+ movdqa (BUF), %xmm1
+ movdqa 0x10(BUF), %xmm2
+ movdqa 0x20(BUF), %xmm3
+ movdqa 0x30(BUF), %xmm4
+ movd CRC, CONSTANT
+ pxor CONSTANT, %xmm1
+ sub $0x40, LEN
+ add $0x40, BUF
+ cmp $0x40, LEN
+ jb .Lless_64
+
+#ifdef __x86_64__
+ movdqa .Lconstant_R2R1(%rip), CONSTANT
+#else
+ movdqa .Lconstant_R2R1, CONSTANT
+#endif
+
+.Lloop_64:/* 64 bytes Full cache line folding */
+ prefetchnta 0x40(BUF)
+ movdqa %xmm1, %xmm5
+ movdqa %xmm2, %xmm6
+ movdqa %xmm3, %xmm7
+#ifdef __x86_64__
+ movdqa %xmm4, %xmm8
+#endif
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pclmulqdq $0x00, CONSTANT, %xmm2
+ pclmulqdq $0x00, CONSTANT, %xmm3
+#ifdef __x86_64__
+ pclmulqdq $0x00, CONSTANT, %xmm4
+#endif
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pclmulqdq $0x11, CONSTANT, %xmm6
+ pclmulqdq $0x11, CONSTANT, %xmm7
+#ifdef __x86_64__
+ pclmulqdq $0x11, CONSTANT, %xmm8
+#endif
+ pxor %xmm5, %xmm1
+ pxor %xmm6, %xmm2
+ pxor %xmm7, %xmm3
+#ifdef __x86_64__
+ pxor %xmm8, %xmm4
+#else
+ /* xmm8 unsupported for x32 */
+ movdqa %xmm4, %xmm5
+ pclmulqdq $0x00, CONSTANT, %xmm4
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pxor %xmm5, %xmm4
+#endif
+
+ pxor (BUF), %xmm1
+ pxor 0x10(BUF), %xmm2
+ pxor 0x20(BUF), %xmm3
+ pxor 0x30(BUF), %xmm4
+
+ sub $0x40, LEN
+ add $0x40, BUF
+ cmp $0x40, LEN
+ jge .Lloop_64
+.Lless_64:/* Folding cache line into 128bit */
+#ifdef __x86_64__
+ movdqa .Lconstant_R4R3(%rip), CONSTANT
+#else
+ movdqa .Lconstant_R4R3, CONSTANT
+#endif
+ prefetchnta (BUF)
+
+ movdqa %xmm1, %xmm5
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pxor %xmm5, %xmm1
+ pxor %xmm2, %xmm1
+
+ movdqa %xmm1, %xmm5
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pxor %xmm5, %xmm1
+ pxor %xmm3, %xmm1
+
+ movdqa %xmm1, %xmm5
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pxor %xmm5, %xmm1
+ pxor %xmm4, %xmm1
+
+ cmp $0x10, LEN
+ jb .Lfold_64
+.Lloop_16:/* Folding rest buffer into 128bit */
+ movdqa %xmm1, %xmm5
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pclmulqdq $0x11, CONSTANT, %xmm5
+ pxor %xmm5, %xmm1
+ pxor (BUF), %xmm1
+ sub $0x10, LEN
+ add $0x10, BUF
+ cmp $0x10, LEN
+ jge .Lloop_16
+
+.Lfold_64:
+ /* perform the last 64 bit fold, also adds 32 zeroes
+ * to the input stream */
+ pclmulqdq $0x01, %xmm1, CONSTANT /* R4 * xmm1.low */
+ psrldq $0x08, %xmm1
+ pxor CONSTANT, %xmm1
+
+ /* final 32-bit fold */
+ movdqa %xmm1, %xmm2
+#ifdef __x86_64__
+ movdqa .Lconstant_R5(%rip), CONSTANT
+ movdqa .Lconstant_mask32(%rip), %xmm3
+#else
+ movdqa .Lconstant_R5, CONSTANT
+ movdqa .Lconstant_mask32, %xmm3
+#endif
+ psrldq $0x04, %xmm2
+ pand %xmm3, %xmm1
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pxor %xmm2, %xmm1
+
+ /* Finish up with the bit-reversed barrett reduction 64 ==> 32 bits */
+#ifdef __x86_64__
+ movdqa .Lconstant_RUpoly(%rip), CONSTANT
+#else
+ movdqa .Lconstant_RUpoly, CONSTANT
+#endif
+ movdqa %xmm1, %xmm2
+ pand %xmm3, %xmm1
+ pclmulqdq $0x10, CONSTANT, %xmm1
+ pand %xmm3, %xmm1
+ pclmulqdq $0x00, CONSTANT, %xmm1
+ pxor %xmm2, %xmm1
+ pextrd $0x01, %xmm1, %eax
+
+ RET
+SYM_FUNC_END(crc32_pclmul_le_16)