Integrate SIMD optimisations for zlib

third_party/zlib/fill_window_sse.c

Index: third_party/zlib/fill_window_sse.c
diff --git a/third_party/zlib/fill_window_sse.c b/third_party/zlib/fill_window_sse.c
new file mode 100644
index 0000000000000000000000000000000000000000..6570116aeafdd8144e7058d82e63d1c010914875
--- /dev/null
+++ b/third_party/zlib/fill_window_sse.c
@@ -0,0 +1,202 @@
+/*
+ * Fill Window with SSE2-optimized hash shifting
+ *
+ * Copyright (C) 2013 Intel Corporation
+ * Authors:
+ *  Arjan van de Ven    <arjan@linux.intel.com>
+ *  Jim Kukunas         <james.t.kukunas@linux.intel.com>
+ *
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include <immintrin.h>
+#include "deflate.h"
+
+#define UPDATE_HASH(s,h,i) \
+    {\
+        if (s->level < 6) { \
+            h = (3483 * (s->window[i]) +\
+                 23081* (s->window[i+1]) +\
+                 6954 * (s->window[i+2]) +\
+                 20947* (s->window[i+3])) & s->hash_mask;\
+        } else {\
+            h = (25881* (s->window[i]) +\
+                 24674* (s->window[i+1]) +\
+                 25811* (s->window[i+2])) & s->hash_mask;\
+        }\
+    }\
+
+extern int read_buf OF((z_streamp strm, Bytef *buf, unsigned size));
+
+void fill_window_sse(deflate_state *s)
+{
+    const __m128i xmm_wsize = _mm_set1_epi16(s->w_size);
+
+    register unsigned n;
+    register Posf *p;
+    unsigned more;    /* Amount of free space at the end of the window. */
+    uInt wsize = s->w_size;
+
+    Assert(s->lookahead < MIN_LOOKAHEAD, "already enough lookahead");
+
+    do {
+        more = (unsigned)(s->window_size -(ulg)s->lookahead -(ulg)s->strstart);
+
+        /* Deal with !@#$% 64K limit: */
+        if (sizeof(int) <= 2) {
+            if (more == 0 && s->strstart == 0 && s->lookahead == 0) {
+                more = wsize;
+
+            } else if (more == (unsigned)(-1)) {
+                /* Very unlikely, but possible on 16 bit machine if
+                 * strstart == 0 && lookahead == 1 (input done a byte at time)
+                 */
+                more--;
+            }
+        }
+
+        /* If the window is almost full and there is insufficient lookahead,
+         * move the upper half to the lower one to make room in the upper half.
+         */
+        if (s->strstart >= wsize+MAX_DIST(s)) {
+
+            zmemcpy(s->window, s->window+wsize, (unsigned)wsize);
+            s->match_start -= wsize;
+            s->strstart    -= wsize; /* we now have strstart >= MAX_DIST */
+            s->block_start -= (long) wsize;
+
+            /* Slide the hash table (could be avoided with 32 bit values
+               at the expense of memory usage). We slide even when level == 0
+               to keep the hash table consistent if we switch back to level > 0
+               later. (Using level 0 permanently is not an optimal usage of
+               zlib, so we don't care about this pathological case.)
+             */
+            n = s->hash_size;
+            p = &s->head[n];
+            p -= 8;
+            do {
+                __m128i value, result;
+
+                value = _mm_loadu_si128((__m128i *)p);
+                result = _mm_subs_epu16(value, xmm_wsize);
+                _mm_storeu_si128((__m128i *)p, result);
+
+                p -= 8;
+                n -= 8;
+            } while (n > 0);
+
+            n = wsize;
+#ifndef FASTEST
+            p = &s->prev[n];
+            p -= 8;
+            do {
+                __m128i value, result;
+
+                value = _mm_loadu_si128((__m128i *)p);
+                result = _mm_subs_epu16(value, xmm_wsize);
+                _mm_storeu_si128((__m128i *)p, result);
+
+                p -= 8;
+                n -= 8;
+            } while (n > 0);
+#endif
+            more += wsize;
+        }
+        if (s->strm->avail_in == 0) break;
+
+        /* If there was no sliding:
+         *    strstart <= WSIZE+MAX_DIST-1 && lookahead <= MIN_LOOKAHEAD - 1 &&
+         *    more == window_size - lookahead - strstart
+         * => more >= window_size - (MIN_LOOKAHEAD-1 + WSIZE + MAX_DIST-1)
+         * => more >= window_size - 2*WSIZE + 2
+         * In the BIG_MEM or MMAP case (not yet supported),
+         *   window_size == input_size + MIN_LOOKAHEAD  &&
+         *   strstart + s->lookahead <= input_size => more >= MIN_LOOKAHEAD.
+         * Otherwise, window_size == 2*WSIZE so more >= 2.
+         * If there was sliding, more >= WSIZE. So in all cases, more >= 2.
+         */
+        Assert(more >= 2, "more < 2");
+
+        n = read_buf(s->strm, s->window + s->strstart + s->lookahead, more);
+        s->lookahead += n;
+
+        /* Initialize the hash value now that we have some input: */
+        if (s->lookahead >= MIN_MATCH) {
+            uInt str = s->strstart;
+            s->ins_h = s->window[str];
+            if (str >= 1)
+                UPDATE_HASH(s, s->ins_h, str + 1 - (MIN_MATCH-1));
+#if MIN_MATCH != 3
+            Call UPDATE_HASH() MIN_MATCH-3 more times
+#endif
+        }
+        /* If the whole input has less than MIN_MATCH bytes, ins_h is garbage,
+         * but this is not important since only literal bytes will be emitted.
+         */
+
+    } while (s->lookahead < MIN_LOOKAHEAD && s->strm->avail_in != 0);
+
+    /* If the WIN_INIT bytes after the end of the current data have never been
+     * written, then zero those bytes in order to avoid memory check reports of
+     * the use of uninitialized (or uninitialised as Julian writes) bytes by
+     * the longest match routines.  Update the high water mark for the next
+     * time through here.  WIN_INIT is set to MAX_MATCH since the longest match
+     * routines allow scanning to strstart + MAX_MATCH, ignoring lookahead.
+     */
+    if (s->high_water < s->window_size) {
+        ulg curr = s->strstart + (ulg)(s->lookahead);
+        ulg init;
+
+        if (s->high_water < curr) {
+            /* Previous high water mark below current data -- zero WIN_INIT
+             * bytes or up to end of window, whichever is less.
+             */
+            init = s->window_size - curr;
+            if (init > WIN_INIT)
+                init = WIN_INIT;
+            zmemzero(s->window + curr, (unsigned)init);
+            s->high_water = curr + init;
+        }
+        else if (s->high_water < (ulg)curr + WIN_INIT) {
+            /* High water mark at or above current data, but below current data
+             * plus WIN_INIT -- zero out to current data plus WIN_INIT, or up
+             * to end of window, whichever is less.
+             */
+            init = (ulg)curr + WIN_INIT - s->high_water;
+            if (init > s->window_size - s->high_water)
+                init = s->window_size - s->high_water;
+            zmemzero(s->window + s->high_water, (unsigned)init);
+            s->high_water += init;
+        }
+    }
+
+    Assert((ulg)s->strstart <= s->window_size - MIN_LOOKAHEAD,
+           "not enough room for search");
+}
+
+ZLIB_INTERNAL Pos insert_string_sse(deflate_state *const s, const Pos str)
+{
+    Pos ret;
+    unsigned *ip, val, h = 0;
+
+    ip = (unsigned *)&s->window[str];
+    val = *ip;
+
+    if (s->level >= 6)
+        val &= 0xFFFFFF;
+
+#ifndef _MSC_VER
+    __asm__ __volatile__ (
+        "crc32 %1,%0\n\t"
+    : "+r" (h)
+    : "r" (val)
+    );
+#else
+    h = _mm_crc32_u32(h, val);

[hans, 2014/10/23 16:11:27]

I think this intrinsic is available in smmintrin.h on GCC and Clang as well. Can
we just include that and avoid the inline asm above?

(The reason I noticed this is because the code failed to build on Clang on
Windows because smmintrin.h isn't included.)

+#endif
+
+    ret = s->head[h & s->hash_mask];
+    s->head[h & s->hash_mask] = str;
+    s->prev[str & s->w_mask] = ret;
+    return ret;
+}