s390/crc32-vx: use vector instructions to optimize CRC-32 computation

Use vector instructions to optimize the computation of CRC-32 checksums.
An optimized version is provided for CRC-32 (IEEE 802.3 Ethernet) in
normal and bitreflected domain, as well as, for bitreflected CRC-32C
(Castagnoli).

Signed-off-by: Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

arch/s390/crypto/crc32be-vx.S

@@ -0,0 +1,207 @@
+/*
+ * Hardware-accelerated CRC-32 variants for Linux on z Systems
+ *
+ * Use the z/Architecture Vector Extension Facility to accelerate the
+ * computing of CRC-32 checksums.
+ *
+ * This CRC-32 implementation algorithm processes the most-significant
+ * bit first (BE).
+ *
+ * Copyright IBM Corp. 2015
+ * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
+ */
+
+#include <linux/linkage.h>
+#include <asm/vx-insn.h>
+
+/* Vector register range containing CRC-32 constants */
+#define CONST_R1R2		%v9
+#define CONST_R3R4		%v10
+#define CONST_R5		%v11
+#define CONST_R6		%v12
+#define CONST_RU_POLY		%v13
+#define CONST_CRC_POLY		%v14
+
+.data
+.align 8
+
+/*
+ * The CRC-32 constant block contains reduction constants to fold and
+ * process particular chunks of the input data stream in parallel.
+ *
+ * For the CRC-32 variants, the constants are precomputed according to
+ * these defintions:
+ *
+ *	R1 = x4*128+64 mod P(x)
+ *	R2 = x4*128    mod P(x)
+ *	R3 = x128+64   mod P(x)
+ *	R4 = x128      mod P(x)
+ *	R5 = x96       mod P(x)
+ *	R6 = x64       mod P(x)
+ *
+ *	Barret reduction constant, u, is defined as floor(x**64 / P(x)).
+ *
+ *	where P(x) is the polynomial in the normal domain and the P'(x) is the
+ *	polynomial in the reversed (bitreflected) domain.
+ *
+ * Note that the constant definitions below are extended in order to compute
+ * intermediate results with a single VECTOR GALOIS FIELD MULTIPLY instruction.
+ * The righmost doubleword can be 0 to prevent contribution to the result or
+ * can be multiplied by 1 to perform an XOR without the need for a separate
+ * VECTOR EXCLUSIVE OR instruction.
+ *
+ * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials:
+ *
+ *	P(x)  = 0x04C11DB7
+ *	P'(x) = 0xEDB88320
+ */
+
+.Lconstants_CRC_32_BE:
+	.quad		0x08833794c, 0x0e6228b11	# R1, R2
+	.quad		0x0c5b9cd4c, 0x0e8a45605	# R3, R4
+	.quad		0x0f200aa66, 1 << 32		# R5, x32
+	.quad		0x0490d678d, 1			# R6, 1
+	.quad		0x104d101df, 0			# u
+	.quad		0x104C11DB7, 0			# P(x)
+
+.previous
+
+.text
+/*
+ * The CRC-32 function(s) use these calling conventions:
+ *
+ * Parameters:
+ *
+ *	%r2:	Initial CRC value, typically ~0; and final CRC (return) value.
+ *	%r3:	Input buffer pointer, performance might be improved if the
+ *		buffer is on a doubleword boundary.
+ *	%r4:	Length of the buffer, must be 64 bytes or greater.
+ *
+ * Register usage:
+ *
+ *	%r5:	CRC-32 constant pool base pointer.
+ *	V0:	Initial CRC value and intermediate constants and results.
+ *	V1..V4:	Data for CRC computation.
+ *	V5..V8:	Next data chunks that are fetched from the input buffer.
+ *
+ *	V9..V14: CRC-32 constants.
+ */
+ENTRY(crc32_be_vgfm_16)
+	/* Load CRC-32 constants */
+	larl	%r5,.Lconstants_CRC_32_BE
+	VLM	CONST_R1R2,CONST_CRC_POLY,0,%r5
+
+	/* Load the initial CRC value into the leftmost word of V0. */
+	VZERO	%v0
+	VLVGF	%v0,%r2,0
+
+	/* Load a 64-byte data chunk and XOR with CRC */
+	VLM	%v1,%v4,0,%r3		/* 64-bytes into V1..V4 */
+	VX	%v1,%v0,%v1		/* V1 ^= CRC */
+	aghi	%r3,64			/* BUF = BUF + 64 */
+	aghi	%r4,-64			/* LEN = LEN - 64 */
+
+	/* Check remaining buffer size and jump to proper folding method */
+	cghi	%r4,64
+	jl	.Lless_than_64bytes
+
+.Lfold_64bytes_loop:
+	/* Load the next 64-byte data chunk into V5 to V8 */
+	VLM	%v5,%v8,0,%r3
+
+	/*
+	 * Perform a GF(2) multiplication of the doublewords in V1 with
+	 * the reduction constants in V0.  The intermediate result is
+	 * then folded (accumulated) with the next data chunk in V5 and
+	 * stored in V1.  Repeat this step for the register contents
+	 * in V2, V3, and V4 respectively.
+	 */
+	VGFMAG	%v1,CONST_R1R2,%v1,%v5
+	VGFMAG	%v2,CONST_R1R2,%v2,%v6
+	VGFMAG	%v3,CONST_R1R2,%v3,%v7
+	VGFMAG	%v4,CONST_R1R2,%v4,%v8
+
+	/* Adjust buffer pointer and length for next loop */
+	aghi	%r3,64			/* BUF = BUF + 64 */
+	aghi	%r4,-64			/* LEN = LEN - 64 */
+
+	cghi	%r4,64
+	jnl	.Lfold_64bytes_loop
+
+.Lless_than_64bytes:
+	/* Fold V1 to V4 into a single 128-bit value in V1 */
+	VGFMAG	%v1,CONST_R3R4,%v1,%v2
+	VGFMAG	%v1,CONST_R3R4,%v1,%v3
+	VGFMAG	%v1,CONST_R3R4,%v1,%v4
+
+	/* Check whether to continue with 64-bit folding */
+	cghi	%r4,16
+	jl	.Lfinal_fold
+
+.Lfold_16bytes_loop:
+
+	VL	%v2,0,,%r3		/* Load next data chunk */
+	VGFMAG	%v1,CONST_R3R4,%v1,%v2	/* Fold next data chunk */
+
+	/* Adjust buffer pointer and size for folding next data chunk */
+	aghi	%r3,16
+	aghi	%r4,-16
+
+	/* Process remaining data chunks */
+	cghi	%r4,16
+	jnl	.Lfold_16bytes_loop
+
+.Lfinal_fold:
+	/*
+	 * The R5 constant is used to fold a 128-bit value into an 96-bit value
+	 * that is XORed with the next 96-bit input data chunk.  To use a single
+	 * VGFMG instruction, multiply the rightmost 64-bit with x^32 (1<<32) to
+	 * form an intermediate 96-bit value (with appended zeros) which is then
+	 * XORed with the intermediate reduction result.
+	 */
+	VGFMG	%v1,CONST_R5,%v1
+
+	/*
+	 * Further reduce the remaining 96-bit value to a 64-bit value using a
+	 * single VGFMG, the rightmost doubleword is multiplied with 0x1. The
+	 * intermediate result is then XORed with the product of the leftmost
+	 * doubleword with R6.	The result is a 64-bit value and is subject to
+	 * the Barret reduction.
+	 */
+	VGFMG	%v1,CONST_R6,%v1
+
+	/*
+	 * The input values to the Barret reduction are the degree-63 polynomial
+	 * in V1 (R(x)), degree-32 generator polynomial, and the reduction
+	 * constant u.	The Barret reduction result is the CRC value of R(x) mod
+	 * P(x).
+	 *
+	 * The Barret reduction algorithm is defined as:
+	 *
+	 *    1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
+	 *    2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
+	 *    3. C(x)  = R(x) XOR T2(x) mod x^32
+	 *
+	 * Note: To compensate the division by x^32, use the vector unpack
+	 * instruction to move the leftmost word into the leftmost doubleword
+	 * of the vector register.  The rightmost doubleword is multiplied
+	 * with zero to not contribute to the intermedate results.
+	 */
+
+	/* T1(x) = floor( R(x) / x^32 ) GF2MUL u */
+	VUPLLF	%v2,%v1
+	VGFMG	%v2,CONST_RU_POLY,%v2
+
+	/*
+	 * Compute the GF(2) product of the CRC polynomial in VO with T1(x) in
+	 * V2 and XOR the intermediate result, T2(x),  with the value in V1.
+	 * The final result is in the rightmost word of V2.
+	 */
+	VUPLLF	%v2,%v2
+	VGFMAG	%v2,CONST_CRC_POLY,%v2,%v1
+
+.Ldone:
+	VLGVF	%r2,%v2,3
+	br	%r14
+
+.previous