Make zlib go faster

libc/runtime/memtrack.internal.h

@@ -108,12 +108,6 @@ forceinline pureconst bool IsArenaFrame(int x) {
   return 0x5004 <= x && x <= 0x7ffb;
 }
 
-forceinline pureconst bool IsKernelFrame(int x) {
-  intptr_t stack = GetStaticStackAddr(0);
-  return (int)(stack >> 16) <= x &&
-         x <= (int)((stack + (GetStackSize() - FRAMESIZE)) >> 16);
-}
-
 forceinline pureconst bool IsStaticStackFrame(int x) {
   intptr_t stack = GetStaticStackAddr(0);
   return (int)(stack >> 16) <= x &&

third_party/zlib/README.cosmo

@@ -26,4 +26,5 @@ ORIGIN
 LOCAL CHANGES
 
   - Changed Trace() calls to use kprintf()
+  - Use X86_HAVE() for runtime dispatching
   - We use our own crc32() implementation from LIBC_STR

third_party/zlib/adler32.c

@@ -2,6 +2,7 @@
  * Copyright (C) 1995-2011, 2016 Mark Adler
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
+#include "libc/nexgen32e/x86feature.h"
 #include "third_party/zlib/macros.internal.h"
 #include "third_party/zlib/zconf.h"
 #include "third_party/zlib/zutil.internal.h"
@@ -62,10 +63,7 @@ local uLong adler32_combine_ OF((uLong adler1, uLong adler2, z_off64_t len2));
 #  define MOD63(a) a %= BASE
 #endif
 
-//# include "cpu_features.h"
-#if defined(ADLER32_SIMD_SSSE3) || defined(ADLER32_SIMD_NEON)
-# include "adler32_simd.h"
-#endif
+uint32_t ZLIB_INTERNAL adler32_simd_(uint32_t, const unsigned char *, z_size_t);
 
 /* ========================================================================= */
 uLong ZEXPORT adler32_z(adler, buf, len)
@@ -77,7 +75,7 @@ uLong ZEXPORT adler32_z(adler, buf, len)
     unsigned n;
 
 #if defined(ADLER32_SIMD_SSSE3)
-    if (buf != Z_NULL && len >= 64 && x86_cpu_enable_ssse3)
+    if (buf != Z_NULL && len >= 64 && X86_HAVE(SSSE3))
         return adler32_simd_(adler, buf, len);
 #elif defined(ADLER32_SIMD_NEON)
     if (buf != Z_NULL && len >= 64)
@@ -99,24 +97,9 @@ uLong ZEXPORT adler32_z(adler, buf, len)
         return adler | (sum2 << 16);
     }
 
-#if defined(ADLER32_SIMD_SSSE3)
-    /*
-     * Use SSSE3 to compute the adler32. Since this routine can be
-     * freely used, check CPU features here. zlib convention is to
-     * call adler32(0, NULL, 0), before making calls to adler32().
-     * So this is a good early (and infrequent) place to cache CPU
-     * features for those later, more interesting adler32() calls.
-     */
-    if (buf == Z_NULL) {
-        if (!len) /* Assume user is calling adler32(0, NULL, 0); */
-            cpu_check_features();
-        return 1L;
-    }
-#else
     /* initial Adler-32 value (deferred check for len == 1 speed) */
     if (buf == Z_NULL)
         return 1L;
-#endif
 
     /* in case short lengths are provided, keep it somewhat fast */
     if (len < 16) {

third_party/zlib/deflate.c

@@ -243,14 +243,6 @@ int ZEXPORT deflateInit2(strm, level, method, windowBits, memLevel, strategy)
     deflate_state *s;
     int wrap = 1;
 
-    // Needed to activate optimized insert_string() that helps compression
-    // for all wrapper formats (e.g. RAW, ZLIB, GZIP).
-    // Feature detection is not triggered while using RAW mode (i.e. we never
-    // call crc32() with a NULL buffer).
-#if defined(CRC32_ARMV8_CRC32) || defined(CRC32_SIMD_SSE42_PCLMUL)
-    cpu_check_features();
-#endif
-
     if (strm == Z_NULL) return Z_STREAM_ERROR;
 
     strm->msg = Z_NULL;
@@ -306,7 +298,7 @@ int ZEXPORT deflateInit2(strm, level, method, windowBits, memLevel, strategy)
     s->chromium_zlib_hash = 0;
 #if !defined(USE_ZLIB_RABIN_KARP_ROLLING_HASH)
   #if defined(TARGET_CPU_WITH_CRC) && defined(CRC32_SIMD_SSE42_PCLMUL)
-    if (x86_cpu_enable_simd)
+    if (X86_HAVE(SSE4_2))
       s->chromium_zlib_hash = 1;
   #elif defined(TARGET_CPU_WITH_CRC) && defined(CRC32_ARMV8_CRC32)
     if (arm_cpu_enable_crc32)

third_party/zlib/insert_string.internal.h

@@ -1,5 +1,6 @@
 #ifndef COSMOPOLITAN_THIRD_PARTY_ZLIB_INSERT_STRING_H_
 #define COSMOPOLITAN_THIRD_PARTY_ZLIB_INSERT_STRING_H_
+#include "libc/nexgen32e/x86feature.h"
 #include "third_party/zlib/deflate.internal.h"
 #include "third_party/zlib/zutil.internal.h"
 #if !(__ASSEMBLER__ + __LINKER__ + 0)
@@ -16,7 +17,7 @@ COSMOPOLITAN_C_START_
 
 // clang-format off
 #if defined(CRC32_SIMD_SSE42_PCLMUL)
-  #include <smmintrin.h>  /* Required to make MSVC bot build pass. */
+//  #include <smmintrin.h>  /* Required to make MSVC bot build pass. */
 
   #if defined(__clang__) || defined(__GNUC__)
     #define TARGET_CPU_WITH_CRC __attribute__((target("sse4.2")))
@@ -50,7 +51,6 @@ COSMOPOLITAN_C_START_
 
 #if defined(TARGET_CPU_WITH_CRC)
 
-TARGET_CPU_WITH_CRC
 local INLINE Pos insert_string_simd(deflate_state* const s, const Pos str) {
   Pos ret;
   unsigned *ip, val, h = 0;
@@ -61,7 +61,7 @@ local INLINE Pos insert_string_simd(deflate_state* const s, const Pos str) {
   if (s->level >= 6) val &= 0xFFFFFF;
 
   /* Compute hash from the CRC32C of |val|. */
-  h = _cpu_crc32c_hash_u32(h, val);
+  asm("crc32\t%1,%0" : "+r"(h) : "rm"(val));
 
   ret = s->head[h & s->hash_mask];
   s->head[h & s->hash_mask] = str;
@@ -129,7 +129,7 @@ local INLINE Pos insert_string(deflate_state* const s, const Pos str) {
  * the Rabin-Karp hash instead.
  */ /* FALLTHROUGH Rabin-Karp */
 #elif defined(TARGET_CPU_WITH_CRC) && defined(CRC32_SIMD_SSE42_PCLMUL)
-  if (x86_cpu_enable_simd) return insert_string_simd(s, str);
+  if (X86_HAVE(SSE4_2)) return insert_string_simd(s, str);
 #elif defined(TARGET_CPU_WITH_CRC) && defined(CRC32_ARMV8_CRC32)
   if (arm_cpu_enable_crc32) return insert_string_simd(s, str);
 #endif

third_party/zlib/zlib.mk

@@ -45,10 +45,13 @@ o/$(MODE)/third_party/zlib/adler32simd.o: private	\
 		OVERRIDE_CFLAGS +=			\
 			-mssse3
 
-o/$(MODE)/third_party/zlib/crcfold.o: private		\
-		OVERRIDE_CFLAGS +=			\
-			-mpclmul			\
-			-mssse3
+o/$(MODE)/third_party/zlib/adler32.o: private		\
+		OVERRIDE_CPPFLAGS +=			\
+			-DADLER32_SIMD_SSSE3
+
+o/$(MODE)/third_party/zlib/deflate.o: private		\
+		OVERRIDE_CPPFLAGS +=			\
+			-DCRC32_SIMD_SSE42_PCLMUL
 
 $(THIRD_PARTY_ZLIB_A_OBJS): private			\
 		OVERRIDE_CFLAGS +=			\