Selective lazy evaluation with targetLength-adaptive gating

build/cmake/CMakeModules/ZstdOptions.cmake

@@ -66,3 +66,10 @@ endif()
 
 # Set global definitions
 add_definitions(-DXXH_NAMESPACE=ZSTD_)
+
+# Selective lazy evaluation optimization
+option(ZSTD_SELECTIVE_LAZY "Enable selective lazy evaluation optimization" OFF)
+if(ZSTD_SELECTIVE_LAZY)
+    message(STATUS "ZSTD_SELECTIVE_LAZY enabled")
+    add_definitions(-DZSTD_SELECTIVE_LAZY)
+endif()

build/meson/meson.build

@@ -108,6 +108,11 @@ use_lz4 = lz4_dep.found()
 
 add_project_arguments('-DXXH_NAMESPACE=ZSTD_', language: ['c'])
 
+use_selective_lazy = get_option('selective_lazy')
+if use_selective_lazy
+  add_project_arguments('-DZSTD_SELECTIVE_LAZY', language: ['c'])
+endif
+
 pzstd_warning_flags = []
 if [compiler_gcc, compiler_clang].contains(cc_id)
   common_warning_flags = [ '-Wundef', '-Wshadow', '-Wcast-align', '-Wcast-qual' ]

build/meson/meson_options.txt

@@ -34,3 +34,5 @@ option('lzma', type: 'feature', value: 'auto',
   description: 'Enable lzma support')
 option('lz4', type: 'feature', value: 'auto',
   description: 'Enable lz4 support')
+option('selective_lazy', type: 'boolean', value: false,
+  description: 'Enable selective lazy evaluation optimization')

lib/Makefile

@@ -59,6 +59,13 @@ VERSION := $(ZSTD_VERSION)
 # -mt or -nomt to the build target (like lib-mt for multi-threaded, lib-nomt for single-threaded).
 
 
+# Selective lazy optimization: reduces lazy evaluation overhead for
+# structured binary data workloads with predictable match patterns.
+# Enable with: make ZSTD_SELECTIVE_LAZY=1
+ifdef ZSTD_SELECTIVE_LAZY
+CPPFLAGS += -DZSTD_SELECTIVE_LAZY
+endif
+
 CPPFLAGS_DYNLIB  += -DZSTD_MULTITHREAD # dynamic library build defaults to multi-threaded
 LDFLAGS_DYNLIB   += -pthread
 CPPFLAGS_STATICLIB +=                  # static library build defaults to single-threaded

lib/compress/zstd_lazy.c

@@ -1669,7 +1669,16 @@ size_t ZSTD_compressBlock_lazy_generic(
         }
 
         /* let's try to find a better solution */
+#ifdef ZSTD_SELECTIVE_LAZY
+        /* Selective lazy: depth-adaptive threshold skips lazy evaluation
+         * for strong matches or nearby offsets. Repcode early exit avoids
+         * expensive searchMax when a repcode match is accepted. */
+        if (depth>=1
+          && matchLength < (ms->cParams.targetLength <= 8 ? 12 : 16)
+          && OFFBASE_IS_OFFSET(offBase) && OFFBASE_TO_OFFSET(offBase) > 256)
+#else
         if (depth>=1)
+#endif
         while (ip<ilimit) {
             DEBUGLOG(7, "search depth 1");
             ip ++;
@@ -1696,6 +1705,13 @@ size_t ZSTD_compressBlock_lazy_generic(
                         matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
             }
+#ifdef ZSTD_SELECTIVE_LAZY
+            /* Repcode early exit: if a repcode match was accepted above,
+             * skip the expensive searchMax. A regular match would need to be
+             * 4+ bytes longer to beat a repcode (zero offset encoding cost),
+             * making searchMax very unlikely to improve the result. */
+            if (!OFFBASE_IS_OFFSET(offBase)) break;
+#endif
             {   size_t ofbCandidate=999999999;
                 size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
                 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
@@ -1732,6 +1748,10 @@ size_t ZSTD_compressBlock_lazy_generic(
                             matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                     }
                 }
+#ifdef ZSTD_SELECTIVE_LAZY
+                /* Repcode early exit at depth 2 */
+                if (!OFFBASE_IS_OFFSET(offBase)) break;
+#endif
                 {   size_t ofbCandidate=999999999;
                     size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
                     int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */

programs/Makefile

@@ -21,6 +21,10 @@ zstd-release:
 LIBZSTD_MK_DIR = ../lib
 include $(LIBZSTD_MK_DIR)/libzstd.mk
 
+ifdef ZSTD_SELECTIVE_LAZY
+CPPFLAGS += -DZSTD_SELECTIVE_LAZY
+endif
+
 ifeq ($(shell $(CC) -v 2>&1 | $(GREP) -c "gcc version "), 1)
   ALIGN_LOOP = -falign-loops=32
 else