Add use of the ARMv8 crc32 instructions when requested.

Define the macro Z_ARM_CRC32 at compile time to use the ARMv8
(aarch64) crc32x and crc32b instructions. This code does not check
for the presence of the crc32 instructions. Those instructions are
optional for ARMv8.0, though mandatory for ARMv8.1 and later. The
use of the crc32 instructions is about ten times as fast as the
software braided calculation of the CRC-32. This can noticeably
speed up the decompression of gzip streams.

crc32.c

@@ -617,6 +617,118 @@ const z_crc_t FAR * ZEXPORT get_crc_table()
     return (const z_crc_t FAR *)crc_table;
 }
 
+/* =========================================================================
+ * Use ARM machine instructions if requested. This will compute the CRC about
+ * ten times faster than the braided calculation. This code does not check for
+ * the presence of the CRC instruction. Compile with care.
+ */
+#if defined(Z_ARM_CRC32) && defined(__aarch64__) && W == 8
+
+/*
+   Constants empirically determined to maximize speed. These values are from
+   measurements on a Cortex-A57. Your mileage may vary.
+ */
+#define Z_BATCH 3990                /* number of words in a batch */
+#define Z_BATCH_ZEROS 0xa10d3d0c    /* computed from Z_BATCH = 3990 */
+#define Z_BATCH_MIN 800             /* fewest words in a final batch */
+
+unsigned long ZEXPORT crc32_z(crc, buf, len)
+    unsigned long crc;
+    const unsigned char FAR *buf;
+    z_size_t len;
+{
+    z_crc_t val;
+    z_word_t crc1, crc2;
+    const z_word_t *word;
+    z_word_t val0, val1, val2;
+    z_size_t last, last2, i;
+    z_size_t num;
+
+    /* Return initial CRC, if requested. */
+    if (buf == Z_NULL) return 0;
+
+#ifdef DYNAMIC_CRC_TABLE
+    once(&made, make_crc_table);
+#endif /* DYNAMIC_CRC_TABLE */
+
+    /* Pre-condition the CRC */
+    crc ^= 0xffffffff;
+
+    /* Compute the CRC up to a word boundary. */
+    while (len && ((z_size_t)buf & 7) != 0) {
+        len--;
+        val = *buf++;
+        __asm__ volatile("crc32b %w0, %w0, %w1" : "+r"(crc) : "r"(val));
+    }
+
+    /* Prepare to compute the CRC on full 64-bit words word[0..num-1]. */
+    word = (z_word_t const *)buf;
+    num = len >> 3;
+    len &= 7;
+
+    /* Do three interleaved CRCs to realize the throughput of one crc32x
+       instruction per cycle. Each CRC is calcuated on Z_BATCH words. The three
+       CRCs are combined into a single CRC after each set of batches. */
+    while (num >= 3 * Z_BATCH) {
+        crc1 = 0;
+        crc2 = 0;
+        for (i = 0; i < Z_BATCH; i++) {
+            val0 = word[i];
+            val1 = word[i + Z_BATCH];
+            val2 = word[i + 2 * Z_BATCH];
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc) : "r"(val0));
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc1) : "r"(val1));
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc2) : "r"(val2));
+        }
+        word += 3 * Z_BATCH;
+        num -= 3 * Z_BATCH;
+        crc = multmodp(Z_BATCH_ZEROS, crc) ^ crc1;
+        crc = multmodp(Z_BATCH_ZEROS, crc) ^ crc2;
+    }
+
+    /* Do one last smaller batch with the remaining words, if there are enough
+       to pay for the combination of CRCs. */
+    last = num / 3;
+    if (last >= Z_BATCH_MIN) {
+        last2 = last << 1;
+        crc1 = 0;
+        crc2 = 0;
+        for (i = 0; i < last; i++) {
+            val0 = word[i];
+            val1 = word[i + last];
+            val2 = word[i + last2];
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc) : "r"(val0));
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc1) : "r"(val1));
+            __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc2) : "r"(val2));
+        }
+        word += 3 * last;
+        num -= 3 * last;
+        val = x2nmodp(last, 6);
+        crc = multmodp(val, crc) ^ crc1;
+        crc = multmodp(val, crc) ^ crc2;
+    }
+
+    /* Compute the CRC on any remaining words. */
+    for (i = 0; i < num; i++) {
+        val0 = word[i];
+        __asm__ volatile("crc32x %w0, %w0, %x1" : "+r"(crc) : "r"(val0));
+    }
+    word += num;
+
+    /* Complete the CRC on any remaining bytes. */
+    buf = (const unsigned char FAR *)word;
+    while (len) {
+        len--;
+        val = *buf++;
+        __asm__ volatile("crc32b %w0, %w0, %w1" : "+r"(crc) : "r"(val));
+    }
+
+    /* Return the CRC, post-conditioned. */
+    return crc ^ 0xffffffff;
+}
+
+#else
+
 /* ========================================================================= */
 unsigned long ZEXPORT crc32_z(crc, buf, len)
     unsigned long crc;
@@ -939,6 +1051,8 @@ unsigned long ZEXPORT crc32_z(crc, buf, len)
     return crc ^ 0xffffffff;
 }
 
+#endif
+
 /* ========================================================================= */
 unsigned long ZEXPORT crc32(crc, buf, len)
     unsigned long crc;