[zlib] Increase inflate dynamic table sizes

First patch* based on Dougall Johnson zlib research**.

The basic idea is to use a bigger root table to avoid a branch
misprediction while performing a lookup when the searched
Huffman code is actually located in a nested sub-table.

The increase of memory use in the structure inflate_state will be
1920 bytes (i.e. 1332 - 852 = 480 x sizeof(code) = 480 x 4) per active inflate context which should be fine given that even lower spec boards generally have CPUs with a L1 cache size of 32KB.

Performance gains are dependent on input entropy features plus architecture, but are a net positive on both x86 (Haswell: 5.1%), Cortex big.LITTLE.Medium (X1, A55, A76) on 32bit (6.3%, 1.44%, 5.7%) and 64bit (8.3%, 2.4%, 6.7%) plus M1 (7.6%).

* https://github.com/dougallj/zlib-dougallj/commits/main
** https://dougallj.wordpress.com/2022/08/20/

Bug: 1354990
Change-Id: I1f784bfaf9a8108f657e3c9753030482e5026366
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/3888219


Reviewed-by: Chris Blume <cblume@chromium.org>
Commit-Queue: Adenilson Cavalcanti <cavalcantii@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1046251}
NOKEYCHECK=True
GitOrigin-RevId: b40b1f7a30b58aeb0d6e4906ae25c6e4734f3b1f

contrib/optimizations/inflate.c

@@ -1043,11 +1043,11 @@ int flush;
             }
 
             /* build code tables -- note: do not change the lenbits or distbits
-               values here (9 and 6) without reading the comments in inftrees.h
+               values here (10 and 9) without reading the comments in inftrees.h
                concerning the ENOUGH constants, which depend on those values */
             state->next = state->codes;
             state->lencode = (const code FAR *)(state->next);
-            state->lenbits = 9;
+            state->lenbits = 10;
             ret = inflate_table(LENS, state->lens, state->nlen, &(state->next),
                                 &(state->lenbits), state->work);
             if (ret) {
@@ -1056,7 +1056,7 @@ int flush;
                 break;
             }
             state->distcode = (const code FAR *)(state->next);
-            state->distbits = 6;
+            state->distbits = 9;
             ret = inflate_table(DISTS, state->lens + state->nlen, state->ndist,
                             &(state->next), &(state->distbits), state->work);
             if (ret) {

infback.c

@@ -455,11 +455,11 @@ void FAR *out_desc;
             }
 
             /* build code tables -- note: do not change the lenbits or distbits
-               values here (9 and 6) without reading the comments in inftrees.h
+               values here (10 and 9) without reading the comments in inftrees.h
                concerning the ENOUGH constants, which depend on those values */
             state->next = state->codes;
             state->lencode = (code const FAR *)(state->next);
-            state->lenbits = 9;
+            state->lenbits = 10;
             ret = inflate_table(LENS, state->lens, state->nlen, &(state->next),
                                 &(state->lenbits), state->work);
             if (ret) {
@@ -468,7 +468,7 @@ void FAR *out_desc;
                 break;
             }
             state->distcode = (code const FAR *)(state->next);
-            state->distbits = 6;
+            state->distbits = 9;
             ret = inflate_table(DISTS, state->lens + state->nlen, state->ndist,
                             &(state->next), &(state->distbits), state->work);
             if (ret) {

inflate.c

@@ -1032,11 +1032,11 @@ int flush;
             }
 
             /* build code tables -- note: do not change the lenbits or distbits
-               values here (9 and 6) without reading the comments in inftrees.h
+               values here (10 and 9) without reading the comments in inftrees.h
                concerning the ENOUGH constants, which depend on those values */
             state->next = state->codes;
             state->lencode = (const code FAR *)(state->next);
-            state->lenbits = 9;
+            state->lenbits = 10;
             ret = inflate_table(LENS, state->lens, state->nlen, &(state->next),
                                 &(state->lenbits), state->work);
             if (ret) {
@@ -1045,7 +1045,7 @@ int flush;
                 break;
             }
             state->distcode = (const code FAR *)(state->next);
-            state->distbits = 6;
+            state->distbits = 9;
             ret = inflate_table(DISTS, state->lens + state->nlen, state->ndist,
                             &(state->next), &(state->distbits), state->work);
             if (ret) {

inftrees.h

@@ -36,17 +36,17 @@ typedef struct {
  */
 
 /* Maximum size of the dynamic table.  The maximum number of code structures is
-   1444, which is the sum of 852 for literal/length codes and 592 for distance
+   1924, which is the sum of 1332 for literal/length codes and 592 for distance
    codes.  These values were found by exhaustive searches using the program
    examples/enough.c found in the zlib distribution.  The arguments to that
    program are the number of symbols, the initial root table size, and the
-   maximum bit length of a code.  "enough 286 9 15" for literal/length codes
-   returns returns 852, and "enough 30 6 15" for distance codes returns 592.
-   The initial root table size (9 or 6) is found in the fifth argument of the
+   maximum bit length of a code.  "enough 286 10 15" for literal/length codes
+   returns returns 1332, and "enough 30 9 15" for distance codes returns 592.
+   The initial root table size (10 or 9) is found in the fifth argument of the
    inflate_table() calls in inflate.c and infback.c.  If the root table size is
    changed, then these maximum sizes would be need to be recalculated and
    updated. */
-#define ENOUGH_LENS 852
+#define ENOUGH_LENS 1332
 #define ENOUGH_DISTS 592
 #define ENOUGH (ENOUGH_LENS+ENOUGH_DISTS)