diff --git a/.circleci/config.yml b/.circleci/config.yml
index 035177edc7..65bb3c2275 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -7,6 +7,8 @@ jobs:
     #       preinstalled to reduce installation time.
     docker:
       - image: fbopensource/zstd-circleci-primary:0.0.1
+    # TODO: Re-enable aarch64 build:
+    #       make aarch64build     && make clean
     steps:
       - checkout
       - run:
@@ -14,12 +16,11 @@ jobs:
           command: |
             ./tests/test-license.py
             cc -v; CFLAGS="-O0 -Werror -pedantic" make all && make clean
-            make c99build         ; make clean
-            make c11build         ; make clean
-            make aarch64build     ; make clean
-            make -j regressiontest; make clean
-            make shortest         ; make clean
-            make cxxtest          ; make clean
+            make c99build         && make clean
+            make c11build         && make clean
+            make -j regressiontest&& make clean
+            make shortest         && make clean
+            make cxxtest          && make clean
   # the second half of the jobs are in this test
   short-tests-1:
     docker:
diff --git a/.github/workflows/dev-short-tests.yml b/.github/workflows/dev-short-tests.yml
index cdc9d7fe3b..5d7e54b51a 100644
--- a/.github/workflows/dev-short-tests.yml
+++ b/.github/workflows/dev-short-tests.yml
@@ -200,6 +200,15 @@ jobs:
         make clean && make -j all MOREFLAGS="-Werror -DZSTD_NO_INLINE -DZSTD_STRIP_ERROR_STRINGS"
         make clean && make check MOREFLAGS="-Werror -DZSTD_NO_INLINE -DZSTD_STRIP_ERROR_STRINGS"
 
+  test-variants:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+    - name: make all variants & validate
+      run: |
+        make -j -C programs allVariants MOREFLAGS=-O0
+        ./tests/test-variants.sh
+
 
   qemu-consistency:
     name: QEMU ${{ matrix.name }}
@@ -263,7 +272,6 @@ jobs:
       run: |
         LDFLAGS="-static" CC=$XCC QEMU_SYS=$XEMU make clean check
 
-
 # This test currently fails on Github Actions specifically.
 # Possible reason : TTY emulation.
 # Note that the same test works fine locally and on travisCI.
diff --git a/Makefile b/Makefile
index b9265e7f85..ff49fc0d8b 100644
--- a/Makefile
+++ b/Makefile
@@ -217,7 +217,7 @@ armbuild: clean
 	CC=arm-linux-gnueabi-gcc CFLAGS="-Werror" $(MAKE) allzstd
 
 aarch64build: clean
-	CC=aarch64-linux-gnu-gcc CFLAGS="-Werror" $(MAKE) allzstd
+	CC=aarch64-linux-gnu-gcc CFLAGS="-Werror -O0" $(MAKE) allzstd
 
 ppcbuild: clean
 	CC=powerpc-linux-gnu-gcc CFLAGS="-m32 -Wno-attributes -Werror" $(MAKE) -j allzstd
@@ -381,23 +381,23 @@ cmakebuild:
 
 c89build: clean
 	$(CC) -v
-	CFLAGS="-std=c89 -Werror" $(MAKE) allmost  # will fail, due to missing support for `long long`
+	CFLAGS="-std=c89 -Werror -O0" $(MAKE) allmost  # will fail, due to missing support for `long long`
 
 gnu90build: clean
 	$(CC) -v
-	CFLAGS="-std=gnu90 -Werror" $(MAKE) allmost
+	CFLAGS="-std=gnu90 -Werror -O0" $(MAKE) allmost
 
 c99build: clean
 	$(CC) -v
-	CFLAGS="-std=c99 -Werror" $(MAKE) allmost
+	CFLAGS="-std=c99 -Werror -O0" $(MAKE) allmost
 
 gnu99build: clean
 	$(CC) -v
-	CFLAGS="-std=gnu99 -Werror" $(MAKE) allmost
+	CFLAGS="-std=gnu99 -Werror -O0" $(MAKE) allmost
 
 c11build: clean
 	$(CC) -v
-	CFLAGS="-std=c11 -Werror" $(MAKE) allmost
+	CFLAGS="-std=c11 -Werror -O0" $(MAKE) allmost
 
 bmix64build: clean
 	$(CC) -v
diff --git a/build/cmake/lib/CMakeLists.txt b/build/cmake/lib/CMakeLists.txt
index 5f756652e7..d91ebe1f06 100644
--- a/build/cmake/lib/CMakeLists.txt
+++ b/build/cmake/lib/CMakeLists.txt
@@ -7,7 +7,7 @@
 # in the COPYING file in the root directory of this source tree).
 # ################################################################
 
-project(libzstd C)
+project(libzstd C ASM)
 
 set(CMAKE_INCLUDE_CURRENT_DIR TRUE)
 option(ZSTD_BUILD_STATIC "BUILD STATIC LIBRARIES" ON)
@@ -22,7 +22,7 @@ include_directories(${LIBRARY_DIR} ${LIBRARY_DIR}/common)
 
 file(GLOB CommonSources ${LIBRARY_DIR}/common/*.c)
 file(GLOB CompressSources ${LIBRARY_DIR}/compress/*.c)
-file(GLOB DecompressSources ${LIBRARY_DIR}/decompress/*.c)
+file(GLOB DecompressSources ${LIBRARY_DIR}/decompress/*.c ${LIBRARY_DIR}/decompress/*.S)
 file(GLOB DictBuilderSources ${LIBRARY_DIR}/dictBuilder/*.c)
 
 set(Sources
diff --git a/contrib/freestanding_lib/freestanding.py b/contrib/freestanding_lib/freestanding.py
index a191f592fa..cd9d637749 100755
--- a/contrib/freestanding_lib/freestanding.py
+++ b/contrib/freestanding_lib/freestanding.py
@@ -460,7 +460,8 @@ def _log(self, *args, **kwargs):
         print(*args, **kwargs)
 
     def _copy_file(self, lib_path):
-        if not (lib_path.endswith(".c") or lib_path.endswith(".h")):
+        suffixes = [".c", ".h", ".S"]
+        if not any((lib_path.endswith(suffix) for suffix in suffixes)):
             return
         if lib_path in SKIPPED_FILES:
             self._log(f"\tSkipping file: {lib_path}")
diff --git a/contrib/linux-kernel/test/Makefile b/contrib/linux-kernel/test/Makefile
index 2908839fd5..dc76a5f401 100644
--- a/contrib/linux-kernel/test/Makefile
+++ b/contrib/linux-kernel/test/Makefile
@@ -18,9 +18,13 @@ CPPFLAGS += -DZSTD_ASAN_DONT_POISON_WORKSPACE
 LINUX_ZSTD_MODULE     := $(wildcard $(LINUX_ZSTDLIB)/*.c)
 LINUX_ZSTD_COMMON     := $(wildcard $(LINUX_ZSTDLIB)/common/*.c)
 LINUX_ZSTD_COMPRESS   := $(wildcard $(LINUX_ZSTDLIB)/compress/*.c)
-LINUX_ZSTD_DECOMPRESS := $(wildcard $(LINUX_ZSTDLIB)/decompress/*.c)
+LINUX_ZSTD_DECOMPRESS := $(wildcard $(LINUX_ZSTDLIB)/decompress/*.c $(LINUX_ZSTDLIB)/decompress/*.S)
 LINUX_ZSTD_FILES      := $(LINUX_ZSTD_MODULE) $(LINUX_ZSTD_COMMON) $(LINUX_ZSTD_COMPRESS) $(LINUX_ZSTD_DECOMPRESS)
-LINUX_ZSTD_OBJECTS    := $(LINUX_ZSTD_FILES:.c=.o)
+LINUX_ZSTD_OBJECTS0   := $(LINUX_ZSTD_FILES:.c=.o)
+LINUX_ZSTD_OBJECTS    := $(LINUX_ZSTD_OBJECTS0:.S=.o)
+
+%.o: %.S
+	$(CC) -c $(CPPFLAGS) $(CFLAGS) $^ -o $@
 
 liblinuxzstd.a: $(LINUX_ZSTD_OBJECTS)
 	$(AR) $(ARFLAGS) $@ $^
diff --git a/contrib/pzstd/Makefile b/contrib/pzstd/Makefile
index 25265e79ad..3d930cca93 100644
--- a/contrib/pzstd/Makefile
+++ b/contrib/pzstd/Makefile
@@ -57,19 +57,6 @@ LD_COMMAND  = $(CXX) $^          $(ALL_LDFLAGS) $(LIBS) -pthread -o $@
 CC_COMMAND  = $(CC)  $(DEPFLAGS) $(ALL_CFLAGS)   -c $<  -o $@
 CXX_COMMAND = $(CXX) $(DEPFLAGS) $(ALL_CXXFLAGS) -c $<  -o $@
 
-# Get a list of all zstd files so we rebuild the static library when we need to
-ZSTDCOMMON_FILES := $(wildcard $(ZSTDDIR)/common/*.c) \
-                    $(wildcard $(ZSTDDIR)/common/*.h)
-ZSTDCOMP_FILES   := $(wildcard $(ZSTDDIR)/compress/*.c) \
-                    $(wildcard $(ZSTDDIR)/compress/*.h)
-ZSTDDECOMP_FILES := $(wildcard $(ZSTDDIR)/decompress/*.c) \
-                    $(wildcard $(ZSTDDIR)/decompress/*.h)
-ZSTDPROG_FILES   := $(wildcard $(PROGDIR)/*.c) \
-                    $(wildcard $(PROGDIR)/*.h)
-ZSTD_FILES       := $(wildcard $(ZSTDDIR)/*.h) \
-                    $(ZSTDDECOMP_FILES) $(ZSTDCOMMON_FILES) $(ZSTDCOMP_FILES) \
-                    $(ZSTDPROG_FILES)
-
 # List all the pzstd source files so we can determine their dependencies
 PZSTD_SRCS  := $(wildcard *.cpp)
 PZSTD_TESTS := $(wildcard test/*.cpp)
@@ -189,7 +176,8 @@ roundtrip: test/RoundTripTest$(EXT)
 
 # Use the static library that zstd builds for simplicity and
 # so we get the compiler options correct
-$(ZSTDDIR)/libzstd.a: $(ZSTD_FILES)
+.PHONY: $(ZSTDDIR)/libzstd.a
+$(ZSTDDIR)/libzstd.a:
 	CFLAGS="$(ALL_CFLAGS)" LDFLAGS="$(ALL_LDFLAGS)" $(MAKE) -C $(ZSTDDIR) libzstd.a
 
 # Rules to build the tests
diff --git a/lib/Makefile b/lib/Makefile
index 376a6690cd..9b74e708cb 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -8,110 +8,13 @@
 # You may select, at your option, one of the above-listed licenses.
 # ################################################################
 
-# Note: by default, the static library is built single-threaded and dynamic library is built
-# multi-threaded. It is possible to force multi or single threaded builds by appending
-# -mt or -nomt to the build target (like lib-mt for multi-threaded, lib-nomt for single-threaded).
-.PHONY: default
-default: lib-release
-
-# define silent mode as default (verbose mode with V=1 or VERBOSE=1)
-$(V)$(VERBOSE).SILENT:
-
-# When cross-compiling from linux to windows,
-# one might need to specify TARGET_SYSTEM as "Windows."
-# Building from Fedora fails without it.
-# (but Ubuntu and Debian don't need to set anything)
-TARGET_SYSTEM ?= $(OS)
-
-# Version numbers
-LIBVER_MAJOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MAJOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < ./zstd.h`
-LIBVER_MINOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MINOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < ./zstd.h`
-LIBVER_PATCH_SCRIPT:=`sed -n '/define ZSTD_VERSION_RELEASE/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < ./zstd.h`
-LIBVER_SCRIPT:= $(LIBVER_MAJOR_SCRIPT).$(LIBVER_MINOR_SCRIPT).$(LIBVER_PATCH_SCRIPT)
-LIBVER_MAJOR := $(shell echo $(LIBVER_MAJOR_SCRIPT))
-LIBVER_MINOR := $(shell echo $(LIBVER_MINOR_SCRIPT))
-LIBVER_PATCH := $(shell echo $(LIBVER_PATCH_SCRIPT))
-LIBVER := $(shell echo $(LIBVER_SCRIPT))
-VERSION?= $(LIBVER)
-CCVER := $(shell $(CC) --version)
-
-# ZSTD_LIB_MINIFY is a helper variable that
-# configures a bunch of other variables to space-optimized defaults.
-ZSTD_LIB_MINIFY ?= 0
-ifneq ($(ZSTD_LIB_MINIFY), 0)
-  HAVE_CC_OZ ?= $(shell echo "" | $(CC) -Oz -x c -c - -o /dev/null 2> /dev/null && echo 1 || echo 0)
-  ZSTD_LEGACY_SUPPORT ?= 0
-  ZSTD_LIB_DEPRECATED ?= 0
-  HUF_FORCE_DECOMPRESS_X1 ?= 1
-  ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT ?= 1
-  ZSTD_NO_INLINE ?= 1
-  ZSTD_STRIP_ERROR_STRINGS ?= 1
-ifneq ($(HAVE_CC_OZ), 0)
-    # Some compilers (clang) support an even more space-optimized setting.
-    CFLAGS += -Oz
-else
-    CFLAGS += -Os
-endif
-  CFLAGS += -fno-stack-protector -fomit-frame-pointer -fno-ident \
-            -DDYNAMIC_BMI2=0 -DNDEBUG
-else
-  CFLAGS += -O3
-endif
-
-DEBUGLEVEL ?= 0
-CPPFLAGS += -DXXH_NAMESPACE=ZSTD_ -DDEBUGLEVEL=$(DEBUGLEVEL)
-ifeq ($(TARGET_SYSTEM),Windows_NT)   # MinGW assumed
-  CPPFLAGS += -D__USE_MINGW_ANSI_STDIO   # compatibility with %zu formatting
-endif
-DEBUGFLAGS= -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow \
-            -Wstrict-aliasing=1 -Wswitch-enum -Wdeclaration-after-statement \
-            -Wstrict-prototypes -Wundef -Wpointer-arith \
-            -Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings \
-            -Wredundant-decls -Wmissing-prototypes -Wc++-compat
-CFLAGS   += $(DEBUGFLAGS) $(MOREFLAGS)
-FLAGS     = $(CPPFLAGS) $(CFLAGS)
-
-CPPFLAGS_DYNLIB  = -DZSTD_MULTITHREAD # dynamic library build defaults to multi-threaded
-LDFLAGS_DYNLIB   = -pthread
-CPPFLAGS_STATLIB =                    # static library build defaults to single-threaded
-
-HAVE_COLORNEVER = $(shell echo a | grep --color=never a > /dev/null 2> /dev/null && echo 1 || echo 0)
-GREP_OPTIONS ?=
-ifeq ($HAVE_COLORNEVER, 1)
-  GREP_OPTIONS += --color=never
-endif
-GREP = grep $(GREP_OPTIONS)
-SED_ERE_OPT ?= -E
-
-ZSTDCOMMON_FILES := $(sort $(wildcard common/*.c))
-ZSTDCOMP_FILES := $(sort $(wildcard compress/*.c))
-ZSTDDECOMP_FILES := $(sort $(wildcard decompress/*.c))
-ZDICT_FILES := $(sort $(wildcard dictBuilder/*.c))
-ZDEPR_FILES := $(sort $(wildcard deprecated/*.c))
-ZSTD_FILES := $(ZSTDCOMMON_FILES)
-
-ifeq ($(findstring GCC,$(CCVER)),GCC)
-decompress/zstd_decompress_block.o : CFLAGS+=-fno-tree-vectorize
-endif
-
 # Modules
 ZSTD_LIB_COMPRESSION ?= 1
 ZSTD_LIB_DECOMPRESSION ?= 1
 ZSTD_LIB_DICTBUILDER ?= 1
 ZSTD_LIB_DEPRECATED ?= 0
 
-# Legacy support
-ZSTD_LEGACY_SUPPORT ?= 5
-ZSTD_LEGACY_MULTITHREADED_API ?= 0
-
-# Build size optimizations
-HUF_FORCE_DECOMPRESS_X1 ?= 0
-HUF_FORCE_DECOMPRESS_X2 ?= 0
-ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT ?= 0
-ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG ?= 0
-ZSTD_NO_INLINE ?= 0
-ZSTD_STRIP_ERROR_STRINGS ?= 0
-
+# Input variables for libzstd.mk
 ifeq ($(ZSTD_LIB_COMPRESSION), 0)
   ZSTD_LIB_DICTBUILDER = 0
   ZSTD_LIB_DEPRECATED = 0
@@ -122,86 +25,46 @@ ifeq ($(ZSTD_LIB_DECOMPRESSION), 0)
   ZSTD_LIB_DEPRECATED = 0
 endif
 
+include libzstd.mk
+
+ZSTD_FILES := $(ZSTD_COMMON_FILES) $(ZSTD_LEGACY_FILES)
+
 ifneq ($(ZSTD_LIB_COMPRESSION), 0)
-  ZSTD_FILES += $(ZSTDCOMP_FILES)
+  ZSTD_FILES += $(ZSTD_COMPRESS_FILES)
 endif
 
 ifneq ($(ZSTD_LIB_DECOMPRESSION), 0)
-  ZSTD_FILES += $(ZSTDDECOMP_FILES)
+  ZSTD_FILES += $(ZSTD_DECOMPRESS_FILES)
 endif
 
 ifneq ($(ZSTD_LIB_DEPRECATED), 0)
-  ZSTD_FILES += $(ZDEPR_FILES)
+  ZSTD_FILES += $(ZSTD_DEPRECATED_FILES)
 endif
 
 ifneq ($(ZSTD_LIB_DICTBUILDER), 0)
-  ZSTD_FILES += $(ZDICT_FILES)
-endif
-
-ifneq ($(HUF_FORCE_DECOMPRESS_X1), 0)
-  CFLAGS += -DHUF_FORCE_DECOMPRESS_X1
-endif
-
-ifneq ($(HUF_FORCE_DECOMPRESS_X2), 0)
-  CFLAGS += -DHUF_FORCE_DECOMPRESS_X2
-endif
-
-ifneq ($(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT), 0)
-  CFLAGS += -DZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
-endif
-
-ifneq ($(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG), 0)
-  CFLAGS += -DZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
-endif
-
-ifneq ($(ZSTD_NO_INLINE), 0)
-  CFLAGS += -DZSTD_NO_INLINE
+  ZSTD_FILES += $(ZSTD_DICTBUILDER_FILES)
 endif
 
-ifneq ($(ZSTD_STRIP_ERROR_STRINGS), 0)
-  CFLAGS += -DZSTD_STRIP_ERROR_STRINGS
-endif
-
-ifneq ($(ZSTD_LEGACY_MULTITHREADED_API), 0)
-  CFLAGS += -DZSTD_LEGACY_MULTITHREADED_API
-endif
-
-ifneq ($(ZSTD_LEGACY_SUPPORT), 0)
-ifeq ($(shell test $(ZSTD_LEGACY_SUPPORT) -lt 8; echo $$?), 0)
-  ZSTD_FILES += $(shell ls legacy/*.c | $(GREP) 'v0[$(ZSTD_LEGACY_SUPPORT)-7]')
-endif
-endif
-CPPFLAGS  += -DZSTD_LEGACY_SUPPORT=$(ZSTD_LEGACY_SUPPORT)
-
 ZSTD_LOCAL_SRC := $(notdir $(ZSTD_FILES))
-ZSTD_LOCAL_OBJ := $(ZSTD_LOCAL_SRC:.c=.o)
+ZSTD_LOCAL_OBJ0 := $(ZSTD_LOCAL_SRC:.c=.o)
+ZSTD_LOCAL_OBJ := $(ZSTD_LOCAL_OBJ0:.S=.o)
 
-ZSTD_SUBDIR := common compress decompress dictBuilder legacy deprecated
-vpath %.c $(ZSTD_SUBDIR)
+VERSION := $(ZSTD_VERSION)
 
-UNAME := $(shell uname)
+# Note: by default, the static library is built single-threaded and dynamic library is built
+# multi-threaded. It is possible to force multi or single threaded builds by appending
+# -mt or -nomt to the build target (like lib-mt for multi-threaded, lib-nomt for single-threaded).
+.PHONY: default
+default: lib-release
+
+CPPFLAGS_DYNLIB  = -DZSTD_MULTITHREAD # dynamic library build defaults to multi-threaded
+LDFLAGS_DYNLIB   = -pthread
+CPPFLAGS_STATLIB =                    # static library build defaults to single-threaded
 
-ifndef BUILD_DIR
-ifeq ($(UNAME), Darwin)
-  ifeq ($(shell md5 < /dev/null > /dev/null; echo $$?), 0)
-    HASH ?= md5
-  endif
-else ifeq ($(UNAME), FreeBSD)
-  HASH ?= gmd5sum
-else ifeq ($(UNAME), NetBSD)
-  HASH ?= md5 -n
-else ifeq ($(UNAME), OpenBSD)
-  HASH ?= md5
-endif
-HASH ?= md5sum
 
-HASH_DIR = conf_$(shell echo $(CC) $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) $(ZSTD_FILES) | $(HASH) | cut -f 1 -d " " )
-HAVE_HASH :=$(shell echo 1 | $(HASH) > /dev/null && echo 1 || echo 0)
-ifeq ($(HAVE_HASH),0)
-  $(info warning : could not find HASH ($(HASH)), needed to differentiate builds using different flags)
-  BUILD_DIR := obj/generic_noconf
+ifeq ($(findstring GCC,$(CCVER)),GCC)
+decompress/zstd_decompress_block.o : CFLAGS+=-fno-tree-vectorize
 endif
-endif # BUILD_DIR
 
 
 # macOS linker doesn't support -soname, and use different extension
@@ -218,13 +81,6 @@ else
   SHARED_EXT_VER = $(SHARED_EXT).$(LIBVER)
 endif
 
-SET_CACHE_DIRECTORY = \
-   +$(MAKE) --no-print-directory $@ \
-    BUILD_DIR=obj/$(HASH_DIR) \
-    CPPFLAGS="$(CPPFLAGS)" \
-    CFLAGS="$(CFLAGS)" \
-    LDFLAGS="$(LDFLAGS)"
-
 
 .PHONY: all
 all: lib
@@ -233,6 +89,13 @@ all: lib
 .PHONY: libzstd.a  # must be run every time
 libzstd.a: CPPFLAGS += $(CPPFLAGS_STATLIB)
 
+SET_CACHE_DIRECTORY = \
+   +$(MAKE) --no-print-directory $@ \
+    BUILD_DIR=obj/$(HASH_DIR) \
+    CPPFLAGS="$(CPPFLAGS)" \
+    CFLAGS="$(CFLAGS)" \
+    LDFLAGS="$(LDFLAGS)"
+
 ifndef BUILD_DIR
 # determine BUILD_DIR from compilation flags
 
@@ -343,6 +206,14 @@ $(ZSTD_STATLIB_DIR)/%.o : %.c $(ZSTD_STATLIB_DIR)/%.d | $(ZSTD_STATLIB_DIR)
 	@echo CC $@
 	$(COMPILE.c) $(DEPFLAGS) $(ZSTD_STATLIB_DIR)/$*.d $(OUTPUT_OPTION) $<
 
+$(ZSTD_DYNLIB_DIR)/%.o : %.S | $(ZSTD_DYNLIB_DIR)
+	@echo AS $@
+	$(COMPILE.c) $(OUTPUT_OPTION) $<
+
+$(ZSTD_STATLIB_DIR)/%.o : %.S | $(ZSTD_STATLIB_DIR)
+	@echo AS $@
+	$(COMPILE.c) $(OUTPUT_OPTION) $<
+
 MKDIR ?= mkdir
 $(BUILD_DIR) $(ZSTD_DYNLIB_DIR) $(ZSTD_STATLIB_DIR):
 	$(MKDIR) -p $@
diff --git a/lib/compress/huf_compress.c b/lib/compress/huf_compress.c
index db11ab3dcf..915778eb12 100644
--- a/lib/compress/huf_compress.c
+++ b/lib/compress/huf_compress.c
@@ -445,7 +445,7 @@ typedef struct {
 /* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing.
  * Strategy is to use as many buckets as possible for representing distinct
  * counts while using the remainder to represent all "large" counts.
- * 
+ *
  * To satisfy this requirement for 192 buckets, we can do the following:
  * Let buckets 0-166 represent distinct counts of [0, 166]
  * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
@@ -517,7 +517,7 @@ static int HUF_quickSortPartition(nodeElt arr[], int const low, int const high)
 }
 
 /* Classic quicksort by descending with partially iterative calls
- * to reduce worst case callstack size. 
+ * to reduce worst case callstack size.
  */
 static void HUF_simpleQuickSort(nodeElt arr[], int low, int high) {
     int const kInsertionSortThreshold = 8;
@@ -813,6 +813,7 @@ FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int id
         assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
         /* We didn't overwrite any bits in the bit container. */
         assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+        (void)dirtyBits;
     }
 #endif
 }
diff --git a/lib/libzstd.mk b/lib/libzstd.mk
new file mode 100644
index 0000000000..45ee1ce629
--- /dev/null
+++ b/lib/libzstd.mk
@@ -0,0 +1,185 @@
+# ################################################################
+# Copyright (c) Yann Collet, Facebook, Inc.
+# All rights reserved.
+#
+# This source code is licensed under both the BSD-style license (found in the
+# LICENSE file in the root directory of this source tree) and the GPLv2 (found
+# in the COPYING file in the root directory of this source tree).
+# You may select, at your option, one of the above-listed licenses.
+# ################################################################
+
+##################################################################
+# Input Variables
+##################################################################
+
+# Zstd lib directory
+LIBZSTD ?= ./
+
+# Legacy support
+ZSTD_LEGACY_SUPPORT ?= 5
+ZSTD_LEGACY_MULTITHREADED_API ?= 0
+
+# Build size optimizations
+HUF_FORCE_DECOMPRESS_X1 ?= 0
+HUF_FORCE_DECOMPRESS_X2 ?= 0
+ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT ?= 0
+ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG ?= 0
+ZSTD_NO_INLINE ?= 0
+ZSTD_STRIP_ERROR_STRINGS ?= 0
+
+# Assembly support
+ZSTD_NO_ASM ?= 0
+
+##################################################################
+# libzstd helpers
+##################################################################
+
+# Make 4.3 doesn't support '\#' anymore (https://lwn.net/Articles/810071/)
+NUM_SYMBOL := \#
+
+# define silent mode as default (verbose mode with V=1 or VERBOSE=1)
+$(V)$(VERBOSE).SILENT:
+
+# When cross-compiling from linux to windows,
+# one might need to specify TARGET_SYSTEM as "Windows."
+# Building from Fedora fails without it.
+# (but Ubuntu and Debian don't need to set anything)
+TARGET_SYSTEM ?= $(OS)
+
+# Version numbers
+LIBVER_SRC := $(LIBZSTD)/zstd.h
+LIBVER_MAJOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MAJOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
+LIBVER_MINOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MINOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
+LIBVER_PATCH_SCRIPT:=`sed -n '/define ZSTD_VERSION_RELEASE/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
+LIBVER_SCRIPT:= $(LIBVER_MAJOR_SCRIPT).$(LIBVER_MINOR_SCRIPT).$(LIBVER_PATCH_SCRIPT)
+LIBVER_MAJOR := $(shell echo $(LIBVER_MAJOR_SCRIPT))
+LIBVER_MINOR := $(shell echo $(LIBVER_MINOR_SCRIPT))
+LIBVER_PATCH := $(shell echo $(LIBVER_PATCH_SCRIPT))
+LIBVER := $(shell echo $(LIBVER_SCRIPT))
+CCVER := $(shell $(CC) --version)
+ZSTD_VERSION?= $(LIBVER)
+
+# ZSTD_LIB_MINIFY is a helper variable that
+# configures a bunch of other variables to space-optimized defaults.
+ZSTD_LIB_MINIFY ?= 0
+ifneq ($(ZSTD_LIB_MINIFY), 0)
+  HAVE_CC_OZ ?= $(shell echo "" | $(CC) -Oz -x c -c - -o /dev/null 2> /dev/null && echo 1 || echo 0)
+  ZSTD_LEGACY_SUPPORT ?= 0
+  ZSTD_LIB_DEPRECATED ?= 0
+  HUF_FORCE_DECOMPRESS_X1 ?= 1
+  ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT ?= 1
+  ZSTD_NO_INLINE ?= 1
+  ZSTD_STRIP_ERROR_STRINGS ?= 1
+ifneq ($(HAVE_CC_OZ), 0)
+    # Some compilers (clang) support an even more space-optimized setting.
+    CFLAGS += -Oz
+else
+    CFLAGS += -Os
+endif
+  CFLAGS += -fno-stack-protector -fomit-frame-pointer -fno-ident \
+            -DDYNAMIC_BMI2=0 -DNDEBUG
+else
+  CFLAGS += -O3
+endif
+
+DEBUGLEVEL ?= 0
+CPPFLAGS += -DXXH_NAMESPACE=ZSTD_ -DDEBUGLEVEL=$(DEBUGLEVEL)
+ifeq ($(TARGET_SYSTEM),Windows_NT)   # MinGW assumed
+  CPPFLAGS += -D__USE_MINGW_ANSI_STDIO   # compatibility with %zu formatting
+endif
+DEBUGFLAGS= -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow \
+            -Wstrict-aliasing=1 -Wswitch-enum -Wdeclaration-after-statement \
+            -Wstrict-prototypes -Wundef -Wpointer-arith \
+            -Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings \
+            -Wredundant-decls -Wmissing-prototypes -Wc++-compat
+CFLAGS   += $(DEBUGFLAGS) $(MOREFLAGS)
+LDFLAGS  += $(MOREFLAGS)
+FLAGS     = $(CPPFLAGS) $(CFLAGS) $(LDFLAGS)
+
+HAVE_COLORNEVER = $(shell echo a | grep --color=never a > /dev/null 2> /dev/null && echo 1 || echo 0)
+GREP_OPTIONS ?=
+ifeq ($HAVE_COLORNEVER, 1)
+  GREP_OPTIONS += --color=never
+endif
+GREP = grep $(GREP_OPTIONS)
+SED_ERE_OPT ?= -E
+
+ZSTD_COMMON_FILES := $(sort $(wildcard $(LIBZSTD)/common/*.c))
+ZSTD_COMPRESS_FILES := $(sort $(wildcard $(LIBZSTD)/compress/*.c))
+ZSTD_DECOMPRESS_FILES := $(sort $(wildcard $(LIBZSTD)/decompress/*.c))
+ZSTD_DICTBUILDER_FILES := $(sort $(wildcard $(LIBZSTD)/dictBuilder/*.c))
+ZSTD_DEPRECATED_FILES := $(sort $(wildcard $(LIBZSTD)/deprecated/*.c))
+ZSTD_LEGACY_FILES :=
+
+ZSTD_DECOMPRESS_AMD64_ASM_FILES := $(sort $(wildcard $(LIBZSTD)/decompress/*_amd64.S))
+
+ifneq ($(ZSTD_NO_ASM), 0)
+  CPPFLAGS += -DHUF_DISABLE_ASM
+else
+  # Unconditionally add the ASM files they are disabled by
+  # macros in the .S file.
+  ZSTD_DECOMPRESS_FILES += $(ZSTD_DECOMPRESS_AMD64_ASM_FILES)
+endif
+
+ifneq ($(HUF_FORCE_DECOMPRESS_X1), 0)
+  CFLAGS += -DHUF_FORCE_DECOMPRESS_X1
+endif
+
+ifneq ($(HUF_FORCE_DECOMPRESS_X2), 0)
+  CFLAGS += -DHUF_FORCE_DECOMPRESS_X2
+endif
+
+ifneq ($(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT), 0)
+  CFLAGS += -DZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
+endif
+
+ifneq ($(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG), 0)
+  CFLAGS += -DZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
+endif
+
+ifneq ($(ZSTD_NO_INLINE), 0)
+  CFLAGS += -DZSTD_NO_INLINE
+endif
+
+ifneq ($(ZSTD_STRIP_ERROR_STRINGS), 0)
+  CFLAGS += -DZSTD_STRIP_ERROR_STRINGS
+endif
+
+ifneq ($(ZSTD_LEGACY_MULTITHREADED_API), 0)
+  CFLAGS += -DZSTD_LEGACY_MULTITHREADED_API
+endif
+
+ifneq ($(ZSTD_LEGACY_SUPPORT), 0)
+ifeq ($(shell test $(ZSTD_LEGACY_SUPPORT) -lt 8; echo $$?), 0)
+  ZSTD_LEGACY_FILES += $(shell ls $(LIBZSTD)/legacy/*.c | $(GREP) 'v0[$(ZSTD_LEGACY_SUPPORT)-7]')
+endif
+endif
+CPPFLAGS  += -DZSTD_LEGACY_SUPPORT=$(ZSTD_LEGACY_SUPPORT)
+
+UNAME := $(shell uname)
+
+ifndef BUILD_DIR
+ifeq ($(UNAME), Darwin)
+  ifeq ($(shell md5 < /dev/null > /dev/null; echo $$?), 0)
+    HASH ?= md5
+  endif
+else ifeq ($(UNAME), FreeBSD)
+  HASH ?= gmd5sum
+else ifeq ($(UNAME), NetBSD)
+  HASH ?= md5 -n
+else ifeq ($(UNAME), OpenBSD)
+  HASH ?= md5
+endif
+HASH ?= md5sum
+
+HASH_DIR = conf_$(shell echo $(CC) $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) $(ZSTD_FILES) | $(HASH) | cut -f 1 -d " " )
+HAVE_HASH :=$(shell echo 1 | $(HASH) > /dev/null && echo 1 || echo 0)
+ifeq ($(HAVE_HASH),0)
+  $(info warning : could not find HASH ($(HASH)), needed to differentiate builds using different flags)
+  BUILD_DIR := obj/generic_noconf
+endif
+endif # BUILD_DIR
+
+ZSTD_SUBDIR := $(LIBZSTD)/common $(LIBZSTD)/compress $(LIBZSTD)/decompress $(LIBZSTD)/dictBuilder $(LIBZSTD)/legacy $(LIBZSTD)/deprecated
+vpath %.c $(ZSTD_SUBDIR)
+vpath %.S $(ZSTD_SUBDIR)
diff --git a/programs/Makefile b/programs/Makefile
index ad05dbe1db..ef094fb743 100644
--- a/programs/Makefile
+++ b/programs/Makefile
@@ -9,7 +9,7 @@
 # ##########################################################################
 # zstd : Command Line Utility, supporting gzip-like arguments
 # zstd32 : Same as zstd, but forced to compile in 32-bits mode
-# zstd_nolegacy : zstd without support of decompression of legacy versions
+# zstd-nolegacy : zstd without support of decompression of legacy versions
 # zstd-small : minimal zstd without dictionary builder and benchmark
 # zstd-compress : compressor-only version of zstd
 # zstd-decompress : decompressor-only version of zstd
@@ -18,31 +18,9 @@
 .PHONY: default
 default: zstd-release
 
-# silent mode by default; verbose can be triggered by V=1 or VERBOSE=1
-$(V)$(VERBOSE).SILENT:
+LIBZSTD := ../lib
 
-
-ZSTDDIR := ../lib
-
-# Version numbers
-LIBVER_SRC := $(ZSTDDIR)/zstd.h
-LIBVER_MAJOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MAJOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
-LIBVER_MINOR_SCRIPT:=`sed -n '/define ZSTD_VERSION_MINOR/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
-LIBVER_PATCH_SCRIPT:=`sed -n '/define ZSTD_VERSION_RELEASE/s/.*[[:blank:]]\([0-9][0-9]*\).*/\1/p' < $(LIBVER_SRC)`
-LIBVER_SCRIPT:= $(LIBVER_MAJOR_SCRIPT).$(LIBVER_MINOR_SCRIPT).$(LIBVER_PATCH_SCRIPT)
-LIBVER_MAJOR := $(shell echo $(LIBVER_MAJOR_SCRIPT))
-LIBVER_MINOR := $(shell echo $(LIBVER_MINOR_SCRIPT))
-LIBVER_PATCH := $(shell echo $(LIBVER_PATCH_SCRIPT))
-LIBVER  := $(shell echo $(LIBVER_SCRIPT))
-
-ZSTD_VERSION = $(LIBVER)
-
-HAVE_COLORNEVER = $(shell echo a | grep --color=never a > /dev/null 2> /dev/null && echo 1 || echo 0)
-GREP_OPTIONS ?=
-ifeq ($HAVE_COLORNEVER, 1)
-  GREP_OPTIONS += --color=never
-endif
-GREP = grep $(GREP_OPTIONS)
+include $(LIBZSTD)/libzstd.mk
 
 ifeq ($(shell $(CC) -v 2>&1 | $(GREP) -c "gcc version "), 1)
   ALIGN_LOOP = -falign-loops=32
@@ -50,78 +28,25 @@ else
   ALIGN_LOOP =
 endif
 
-DEBUGLEVEL ?= 0
-CPPFLAGS += -DXXH_NAMESPACE=ZSTD_ -DDEBUGLEVEL=$(DEBUGLEVEL)
-ifeq ($(OS),Windows_NT)   # MinGW assumed
-  CPPFLAGS += -D__USE_MINGW_ANSI_STDIO   # compatibility with %zu formatting
-endif
-CFLAGS   ?= -O3
-DEBUGFLAGS+=-Wall -Wextra -Wcast-qual -Wcast-align -Wshadow \
-            -Wstrict-aliasing=1 -Wswitch-enum -Wdeclaration-after-statement \
-            -Wstrict-prototypes -Wundef -Wpointer-arith \
-            -Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings \
-            -Wredundant-decls -Wmissing-prototypes -Wc++-compat
-CFLAGS   += $(DEBUGFLAGS)
-CPPFLAGS += $(MOREFLAGS)
-LDFLAGS  += $(MOREFLAGS)
-FLAGS     = $(CFLAGS) $(CPPFLAGS) $(LDFLAGS)
-
-ZSTDLIB_COMMON := $(ZSTDDIR)/common
-ZSTDLIB_COMPRESS := $(ZSTDDIR)/compress
-ZSTDLIB_DECOMPRESS := $(ZSTDDIR)/decompress
-ZDICT_DIR := $(ZSTDDIR)/dictBuilder
-ZSTDLEGACY_DIR := $(ZSTDDIR)/legacy
-
-vpath %.c $(ZSTDLIB_COMMON) $(ZSTDLIB_COMPRESS) $(ZSTDLIB_DECOMPRESS) $(ZDICT_DIR) $(ZSTDLEGACY_DIR)
-
-ZSTDLIB_COMMON_C := $(wildcard $(ZSTDLIB_COMMON)/*.c)
-ZSTDLIB_COMPRESS_C := $(wildcard $(ZSTDLIB_COMPRESS)/*.c)
-ZSTDLIB_DECOMPRESS_C := $(wildcard $(ZSTDLIB_DECOMPRESS)/*.c)
-ZSTDLIB_CORE_SRC := $(ZSTDLIB_DECOMPRESS_C) $(ZSTDLIB_COMMON_C) $(ZSTDLIB_COMPRESS_C)
-ZDICT_SRC := $(wildcard $(ZDICT_DIR)/*.c)
-
-ZSTD_LEGACY_SUPPORT ?= 5
-ZSTDLEGACY_SRC :=
-ifneq ($(ZSTD_LEGACY_SUPPORT), 0)
-ifeq ($(shell test $(ZSTD_LEGACY_SUPPORT) -lt 8; echo $$?), 0)
-  ZSTDLEGACY_SRC += $(shell ls $(ZSTDLEGACY_DIR)/*.c | $(GREP) 'v0[$(ZSTD_LEGACY_SUPPORT)-7]')
-endif
-endif
+ZSTDLIB_COMMON_SRC := $(sort $(ZSTD_COMMON_FILES))
+ZSTDLIB_COMPRESS_SRC := $(sort $(ZSTD_COMPRESS_FILES))
+ZSTDLIB_DECOMPRESS_SRC := $(sort $(ZSTD_DECOMPRESS_FILES))
+ZSTDLIB_CORE_SRC := $(sort $(ZSTD_DECOMPRESS_FILES) $(ZSTD_COMMON_FILES) $(ZSTD_COMPRESS_FILES))
+ZDICT_SRC := $(sort $(ZSTD_DICTBUILDER_FILES))
+ZSTDLEGACY_SRC := $(sort $(ZSTD_LEGACY_FILES))
 
 # Sort files in alphabetical order for reproducible builds
 ZSTDLIB_FULL_SRC = $(sort $(ZSTDLIB_CORE_SRC) $(ZSTDLEGACY_SRC) $(ZDICT_SRC))
 ZSTDLIB_LOCAL_SRC = $(notdir $(ZSTDLIB_FULL_SRC))
-ZSTDLIB_LOCAL_OBJ := $(ZSTDLIB_LOCAL_SRC:.c=.o)
+ZSTDLIB_LOCAL_OBJ0 := $(ZSTDLIB_LOCAL_SRC:.c=.o)
+ZSTDLIB_LOCAL_OBJ := $(ZSTDLIB_LOCAL_OBJ0:.S=.o)
 
 ZSTD_CLI_SRC := $(wildcard *.c)
 ZSTD_CLI_OBJ := $(ZSTD_CLI_SRC:.c=.o)
 
 ZSTD_ALL_SRC = $(ZSTDLIB_LOCAL_SRC) $(ZSTD_CLI_SRC)
-ZSTD_ALL_OBJ := $(ZSTD_ALL_SRC:.c=.o)
-
-UNAME := $(shell uname)
-
-ifndef BUILD_DIR
-ifeq ($(UNAME), Darwin)
-  ifeq ($(shell md5 < /dev/null > /dev/null; echo $$?), 0)
-    HASH ?= md5
-  endif
-else ifeq ($(UNAME), FreeBSD)
-  HASH ?= gmd5sum
-else ifeq ($(UNAME), NetBSD)
-  HASH ?= md5 -n
-else ifeq ($(UNAME), OpenBSD)
-  HASH ?= md5
-endif
-HASH ?= md5sum
-HAVE_HASH :=$(shell echo 1 | $(HASH) > /dev/null && echo 1 || echo 0)
-
-HASH_DIR = conf_$(shell echo $(CC) $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) $(LDLIBS) $(ZSTD_FILES) | $(HASH) | cut -f 1 -d " ")
-ifeq ($(HAVE_HASH),0)
-  $(info warning : could not find HASH ($(HASH)), needed to differentiate builds using different flags)
-  BUILD_DIR := obj/generic_noconf
-endif
-endif # BUILD_DIR
+ZSTD_ALL_OBJ0 := $(ZSTD_ALL_SRC:.c=.o)
+ZSTD_ALL_OBJ := $(ZSTD_ALL_OBJ0:.S=.o)
 
 # Define *.exe as extension for Windows systems
 ifneq (,$(filter Windows%,$(OS)))
@@ -139,9 +64,6 @@ endif
 
 VOID = /dev/null
 
-# Make 4.3 doesn't support '\#' anymore (https://lwn.net/Articles/810071/)
-NUM_SYMBOL := \#
-
 # thread detection
 NO_THREAD_MSG := ==> no threads, building without multithreading support
 HAVE_PTHREAD := $(shell printf '$(NUM_SYMBOL)include <pthread.h>\nint main(void) { return 0; }' > have_pthread.c && $(CC) $(FLAGS) -o have_pthread$(EXT) have_pthread.c -pthread 2> $(VOID) && rm have_pthread$(EXT) && echo 1 || echo 0; rm have_pthread.c)
@@ -212,7 +134,7 @@ SET_CACHE_DIRECTORY = \
 all: zstd
 
 .PHONY: allVariants
-allVariants: zstd zstd-compress zstd-decompress zstd-small zstd-nolegacy zstd-dictBuilder
+allVariants: zstd zstd-compress zstd-decompress zstd-small zstd-frugal zstd-nolegacy zstd-dictBuilder
 
 .PHONY: zstd  # must always be run
 zstd : CPPFLAGS += $(THREAD_CPP) $(ZLIBCPP) $(LZMACPP) $(LZ4CPP)
@@ -276,6 +198,7 @@ zstd32 : $(ZSTDLIB_FULL_SRC) $(ZSTD_CLI_SRC)
 
 ## zstd-nolegacy: same scope as zstd, with just support of legacy formats removed
 zstd-nolegacy : LDFLAGS += $(THREAD_LD) $(ZLIBLD) $(LZMALD) $(LZ4LD) $(DEBUGFLAGS_LD)
+zstd-nolegacy : CPPFLAGS += -UZSTD_LEGACY_SUPPORT -DZSTD_LEGACY_SUPPORT=0
 zstd-nolegacy : $(ZSTDLIB_CORE_SRC) $(ZDICT_SRC) $(ZSTD_CLI_OBJ)
 	$(CC) $(FLAGS) $^ -o $@$(EXT) $(LDFLAGS)
 
@@ -299,7 +222,7 @@ zstd-noxz : zstd
 
 ## zstd-dll: zstd executable linked to dynamic library libzstd (must have same version)
 .PHONY: zstd-dll
-zstd-dll : LDFLAGS+= -L$(ZSTDDIR)
+zstd-dll : LDFLAGS+= -L$(LIBZSTD)
 zstd-dll : LDLIBS += -lzstd
 zstd-dll : ZSTDLIB_LOCAL_SRC = xxhash.c
 zstd-dll : zstd
@@ -323,18 +246,17 @@ zstd-pgo :
 ## zstd-small: minimal target, supporting only zstd compression and decompression. no bench. no legacy. no other format.
 zstd-small: CFLAGS = -Os -s
 zstd-frugal zstd-small: $(ZSTDLIB_CORE_SRC) zstdcli.c util.c timefn.c fileio.c
-	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NOTRACE $^ -o $@$(EXT)
+	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NOTRACE -UZSTD_LEGACY_SUPPORT -DZSTD_LEGACY_SUPPORT=0 $^ -o $@$(EXT)
 
-zstd-decompress: $(ZSTDLIB_COMMON_C) $(ZSTDLIB_DECOMPRESS_C) zstdcli.c util.c timefn.c fileio.c
-	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NOCOMPRESS -DZSTD_NOTRACE $^ -o $@$(EXT)
+zstd-decompress: $(ZSTDLIB_COMMON_SRC) $(ZSTDLIB_DECOMPRESS_SRC) zstdcli.c util.c timefn.c fileio.c
+	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NOCOMPRESS -DZSTD_NOTRACE -UZSTD_LEGACY_SUPPORT -DZSTD_LEGACY_SUPPORT=0 $^ -o $@$(EXT)
 
-zstd-compress: $(ZSTDLIB_COMMON_C) $(ZSTDLIB_COMPRESS_C) zstdcli.c util.c timefn.c fileio.c
-	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NODECOMPRESS -DZSTD_NOTRACE $^ -o $@$(EXT)
+zstd-compress: $(ZSTDLIB_COMMON_SRC) $(ZSTDLIB_COMPRESS_SRC) zstdcli.c util.c timefn.c fileio.c
+	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODICT -DZSTD_NODECOMPRESS -DZSTD_NOTRACE -UZSTD_LEGACY_SUPPORT -DZSTD_LEGACY_SUPPORT=0 $^ -o $@$(EXT)
 
 ## zstd-dictBuilder: executable supporting dictionary creation and compression (only)
-zstd-dictBuilder: CPPFLAGS += -DZSTD_NOBENCH -DZSTD_NODECOMPRESS -DZSTD_NOTRACE
-zstd-dictBuilder: $(ZSTDLIB_COMMON_C) $(ZSTDLIB_COMPRESS_C) $(ZDICT_SRC) zstdcli.c util.c timefn.c fileio.c dibio.c
-	$(CC) $(FLAGS) $^ -o $@$(EXT)
+zstd-dictBuilder: $(ZSTDLIB_COMMON_SRC) $(ZSTDLIB_COMPRESS_SRC) $(ZDICT_SRC) zstdcli.c util.c timefn.c fileio.c dibio.c
+	$(CC) $(FLAGS) -DZSTD_NOBENCH -DZSTD_NODECOMPRESS -DZSTD_NOTRACE $^ -o $@$(EXT)
 
 zstdmt: zstd
 	ln -sf zstd zstdmt
@@ -398,6 +320,10 @@ $(BUILD_DIR)/%.o : %.c $(BUILD_DIR)/%.d | $(BUILD_DIR)
 	@echo CC $@
 	$(COMPILE.c) $(DEPFLAGS) $(BUILD_DIR)/$*.d $(OUTPUT_OPTION) $<
 
+$(BUILD_DIR)/%.o : %.S | $(BUILD_DIR)
+	@echo AS $@
+	$(COMPILE.c) $(OUTPUT_OPTION) $<
+
 MKDIR ?= mkdir
 $(BUILD_DIR): ; $(MKDIR) -p $@
 
diff --git a/tests/Makefile b/tests/Makefile
index 85553007d9..56c4483404 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -1,4 +1,5 @@
-# ################################################################
+
+ # ################################################################
 # Copyright (c) Yann Collet, Facebook, Inc.
 # All rights reserved.
 #
@@ -19,46 +20,43 @@
 # zstreamtest32: Same as zstreamtest, but forced to compile in 32-bits mode
 # ##########################################################################
 
-ZSTDDIR = ../lib
+LIBZSTD = ../lib
+
+ZSTD_LEGACY_SUPPORT ?= 0
+
+DEBUGLEVEL ?= 2
+export DEBUGLEVEL  # transmit value to sub-makefiles
+
+include $(LIBZSTD)/libzstd.mk
+
+ZSTDDIR = $(LIBZSTD)
 PRGDIR  = ../programs
 PYTHON ?= python3
 TESTARTEFACT := versionsTest
 
-DEBUGLEVEL ?= 2
-export DEBUGLEVEL  # transmit value to sub-makefiles
-DEBUGFLAGS  = -g -DDEBUGLEVEL=$(DEBUGLEVEL)
+DEBUGFLAGS += -g -Wno-c++-compat
 CPPFLAGS   += -I$(ZSTDDIR) -I$(ZSTDDIR)/common -I$(ZSTDDIR)/compress \
               -I$(ZSTDDIR)/dictBuilder -I$(ZSTDDIR)/deprecated -I$(PRGDIR) \
 			  -DZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY=1
-ifeq ($(OS),Windows_NT)   # MinGW assumed
-CPPFLAGS   += -D__USE_MINGW_ANSI_STDIO   # compatibility with %zu formatting
-endif
-CFLAGS     ?= -O3
-CFLAGS     += -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow                 \
-              -Wstrict-aliasing=1 -Wswitch-enum -Wdeclaration-after-statement \
-              -Wstrict-prototypes -Wundef                                     \
-              -Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings      \
-              -Wredundant-decls -Wmissing-prototypes -Wno-deprecated-declarations
-CFLAGS     += $(DEBUGFLAGS)
-CPPFLAGS   += $(MOREFLAGS)
-
-
-ZSTDCOMMON_FILES := $(ZSTDDIR)/common/*.c
-ZSTDCOMP_FILES   := $(ZSTDDIR)/compress/*.c
-ZSTDDECOMP_FILES := $(ZSTDDIR)/decompress/*.c
+
+ZSTDCOMMON_FILES := $(sort $(ZSTD_COMMON_FILES))
+ZSTDCOMP_FILES   := $(sort $(ZSTD_COMPRESS_FILES))
+ZSTDDECOMP_FILES := $(sort $(ZSTD_DECOMPRESS_FILES))
 ZSTD_FILES  := $(ZSTDDECOMP_FILES) $(ZSTDCOMMON_FILES) $(ZSTDCOMP_FILES)
-ZDICT_FILES := $(ZSTDDIR)/dictBuilder/*.c
+ZDICT_FILES := $(sort $(ZSTD_DICTBUILDER_FILES))
 
 ZSTD_F1 := $(wildcard $(ZSTD_FILES))
 ZSTD_OBJ1 := $(subst $(ZSTDDIR)/common/,zstdm_,$(ZSTD_F1))
 ZSTD_OBJ2 := $(subst $(ZSTDDIR)/compress/,zstdc_,$(ZSTD_OBJ1))
 ZSTD_OBJ3 := $(subst $(ZSTDDIR)/decompress/,zstdd_,$(ZSTD_OBJ2))
-ZSTD_OBJECTS := $(ZSTD_OBJ3:.c=.o)
+ZSTD_OBJ4 := $(ZSTD_OBJ3:.c=.o)
+ZSTD_OBJECTS := $(ZSTD_OBJ4:.S=.o)
 
 ZSTDMT_OBJ1 := $(subst $(ZSTDDIR)/common/,zstdmt_m_,$(ZSTD_F1))
 ZSTDMT_OBJ2 := $(subst $(ZSTDDIR)/compress/,zstdmt_c_,$(ZSTDMT_OBJ1))
 ZSTDMT_OBJ3 := $(subst $(ZSTDDIR)/decompress/,zstdmt_d_,$(ZSTDMT_OBJ2))
-ZSTDMT_OBJECTS := $(ZSTDMT_OBJ3:.c=.o)
+ZSTDMT_OBJ4 := $(ZSTDMT_OBJ3:.c=.o)
+ZSTDMT_OBJECTS := $(ZSTDMT_OBJ4:.S=.o)
 
 # Define *.exe as extension for Windows systems
 ifneq (,$(filter Windows%,$(OS)))
@@ -119,6 +117,9 @@ zstdc_%.o : $(ZSTDDIR)/compress/%.c
 zstdd_%.o : $(ZSTDDIR)/decompress/%.c
 	$(CC) -c $(CPPFLAGS) $(CFLAGS) $< -o $@
 
+zstdd_%.o : $(ZSTDDIR)/decompress/%.S
+	$(CC) -c $(CPPFLAGS) $(CFLAGS) $< -o $@
+
 zstdmt%.o : CPPFLAGS += $(MULTITHREAD_CPP)
 
 zstdmt_m_%.o : $(ZSTDDIR)/common/%.c
@@ -130,6 +131,9 @@ zstdmt_c_%.o : $(ZSTDDIR)/compress/%.c
 zstdmt_d_%.o : $(ZSTDDIR)/decompress/%.c
 	$(CC) -c $(CPPFLAGS) $(CFLAGS) $< -o $@
 
+zstdmt_d_%.o : $(ZSTDDIR)/decompress/%.S
+	$(CC) -c $(CPPFLAGS) $(CFLAGS) $< -o $@
+
 fullbench32: CPPFLAGS += -m32
 fullbench fullbench32 : CPPFLAGS += $(MULTITHREAD_CPP) -Wno-deprecated-declarations
 fullbench fullbench32 : LDFLAGS += $(MULTITHREAD_LD)
@@ -201,7 +205,7 @@ bigdict: $(ZSTDMT_OBJECTS) $(PRGDIR)/datagen.c bigdict.c
 
 invalidDictionaries : $(ZSTD_OBJECTS) invalidDictionaries.c
 
-legacy : CPPFLAGS += -I$(ZSTDDIR)/legacy -DZSTD_LEGACY_SUPPORT=4
+legacy : CPPFLAGS += -I$(ZSTDDIR)/legacy -UZSTD_LEGACY_SUPPORT -DZSTD_LEGACY_SUPPORT=4
 legacy : $(ZSTD_FILES) $(wildcard $(ZSTDDIR)/legacy/*.c) legacy.c
 
 decodecorpus : LDLIBS += -lm
diff --git a/tests/fuzz/Makefile b/tests/fuzz/Makefile
index 5c54ccd77d..72dbccb577 100644
--- a/tests/fuzz/Makefile
+++ b/tests/fuzz/Makefile
@@ -23,6 +23,12 @@ else
 endif
 CORPORA_URL_PREFIX:=https://github.com/facebook/zstd/releases/download/fuzz-corpora/
 
+LIBZSTD = ../../lib
+DEBUGLEVEL ?= 2
+ZSTD_LEGACY_SUPPORT ?= 1
+
+include $(LIBZSTD)/libzstd.mk
+
 ZSTDDIR = ../../lib
 PRGDIR = ../../programs
 CONTRIBDIR = ../../contrib
@@ -34,7 +40,7 @@ FUZZ_EXTRA_FLAGS := -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow \
 	-Wstrict-aliasing=1 -Wswitch-enum -Wdeclaration-after-statement \
 	-Wstrict-prototypes -Wundef \
 	-Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings \
-	-Wredundant-decls \
+	-Wredundant-decls -Wno-deprecated-declarations \
 	-g -fno-omit-frame-pointer
 FUZZ_CFLAGS := $(FUZZ_EXTRA_FLAGS) $(CFLAGS)
 FUZZ_CXXFLAGS := $(FUZZ_EXTRA_FLAGS) -std=c++11 $(CXXFLAGS)
@@ -50,11 +56,11 @@ FUZZ_SRC := $(PRGDIR)/util.c ./fuzz_helpers.c ./zstd_helpers.c ./fuzz_data_produ
 SEEKABLE_HEADERS = $(CONTRIBDIR)/seekable_format/zstd_seekable.h
 SEEKABLE_OBJS = $(CONTRIBDIR)/seekable_format/zstdseek_compress.c $(CONTRIBDIR)/seekable_format/zstdseek_decompress.c
 
-ZSTDCOMMON_SRC := $(ZSTDDIR)/common/*.c
-ZSTDCOMP_SRC   := $(ZSTDDIR)/compress/*.c
-ZSTDDECOMP_SRC := $(ZSTDDIR)/decompress/*.c
-ZSTDDICT_SRC := $(ZSTDDIR)/dictBuilder/*.c
-ZSTDLEGACY_SRC := $(ZSTDDIR)/legacy/*.c
+ZSTDCOMMON_SRC := $(ZSTD_COMMON_FILES)
+ZSTDCOMP_SRC   := $(ZSTD_COMPRESS_FILES)
+ZSTDDECOMP_SRC := $(ZSTD_DECOMPRESS_FILES)
+ZSTDDICT_SRC := $(ZSTD_DICTBUILDER_FILES)
+ZSTDLEGACY_SRC := $(ZSTD_LEGACY_FILES)
 FUZZ_SRC       := \
 	$(FUZZ_SRC) \
 	$(ZSTDDECOMP_SRC) \
@@ -71,7 +77,8 @@ FUZZ_D_OBJ4 := $(subst $(ZSTDDIR)/dictBuilder/,d_lib_dictBuilder_,$(FUZZ_D_OBJ3)
 FUZZ_D_OBJ5 := $(subst $(ZSTDDIR)/legacy/,d_lib_legacy_,$(FUZZ_D_OBJ4))
 FUZZ_D_OBJ6 := $(subst $(PRGDIR)/,d_prg_,$(FUZZ_D_OBJ5))
 FUZZ_D_OBJ7 := $(subst $\./,d_fuzz_,$(FUZZ_D_OBJ6))
-FUZZ_DECOMPRESS_OBJ := $(FUZZ_D_OBJ7:.c=.o)
+FUZZ_D_OBJ8 := $(FUZZ_D_OBJ7:.c=.o)
+FUZZ_DECOMPRESS_OBJ := $(FUZZ_D_OBJ8:.S=.o)
 
 FUZZ_RT_OBJ1 := $(subst $(ZSTDDIR)/common/,rt_lib_common_,$(FUZZ_SRC))
 FUZZ_RT_OBJ2 := $(subst $(ZSTDDIR)/compress/,rt_lib_compress_,$(FUZZ_RT_OBJ1))
@@ -80,7 +87,8 @@ FUZZ_RT_OBJ4 := $(subst $(ZSTDDIR)/dictBuilder/,rt_lib_dictBuilder_,$(FUZZ_RT_OB
 FUZZ_RT_OBJ5 := $(subst $(ZSTDDIR)/legacy/,rt_lib_legacy_,$(FUZZ_RT_OBJ4))
 FUZZ_RT_OBJ6 := $(subst $(PRGDIR)/,rt_prg_,$(FUZZ_RT_OBJ5))
 FUZZ_RT_OBJ7 := $(subst $\./,rt_fuzz_,$(FUZZ_RT_OBJ6))
-FUZZ_ROUND_TRIP_OBJ := $(FUZZ_RT_OBJ7:.c=.o)
+FUZZ_RT_OBJ8 := $(FUZZ_RT_OBJ7:.c=.o)
+FUZZ_ROUND_TRIP_OBJ := $(FUZZ_RT_OBJ8:.S=.o)
 
 .PHONY: default all clean cleanall
 
@@ -117,6 +125,9 @@ rt_lib_compress_%.o: $(ZSTDDIR)/compress/%.c
 rt_lib_decompress_%.o: $(ZSTDDIR)/decompress/%.c
 	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $(FUZZ_ROUND_TRIP_FLAGS) $< -c -o $@
 
+rt_lib_decompress_%.o: $(ZSTDDIR)/decompress/%.S
+	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $(FUZZ_ROUND_TRIP_FLAGS) $< -c -o $@
+
 rt_lib_dictBuilder_%.o: $(ZSTDDIR)/dictBuilder/%.c
 	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $(FUZZ_ROUND_TRIP_FLAGS) $< -c -o $@
 
@@ -138,6 +149,9 @@ d_lib_compress_%.o: $(ZSTDDIR)/compress/%.c
 d_lib_decompress_%.o: $(ZSTDDIR)/decompress/%.c
 	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $< -c -o $@
 
+d_lib_decompress_%.o: $(ZSTDDIR)/decompress/%.S
+	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $< -c -o $@
+
 d_lib_dictBuilder_%.o: $(ZSTDDIR)/dictBuilder/%.c
 	$(CC) $(FUZZ_CPPFLAGS) $(FUZZ_CFLAGS) $< -c -o $@
 
diff --git a/tests/fuzzer.c b/tests/fuzzer.c
index ce79f53acf..c4de54247b 100644
--- a/tests/fuzzer.c
+++ b/tests/fuzzer.c
@@ -1528,7 +1528,7 @@ static int basicUnitTests(U32 const seed, double compressibility)
 
             DISPLAYLEVEL(3, "test%3i : resize context to full CCtx size : ", testNb++);
             staticCCtx = ZSTD_initStaticCStream(staticCCtxBuffer, staticCCtxSize);
-            DISPLAYLEVEL(4, "staticCCtxBuffer = %p,  staticCCtx = %p , ", staticCCtxBuffer, staticCCtx);
+            DISPLAYLEVEL(4, "staticCCtxBuffer = %p,  staticCCtx = %p , ", staticCCtxBuffer, (void*)staticCCtx);
             if (staticCCtx == NULL) goto _output_error;
             DISPLAYLEVEL(3, "OK \n");
 
diff --git a/tests/regression/Makefile b/tests/regression/Makefile
index 3a5a64e5b0..a440c6c942 100644
--- a/tests/regression/Makefile
+++ b/tests/regression/Makefile
@@ -46,6 +46,7 @@ result.o: result.c result.h
 test.o: test.c data.h config.h method.h
 	$(CC) $(REGRESSION_CFLAGS) $(REGRESSION_CPPFLAGS) $< -c -o $@
 
+.PHONY: libzstd.a
 libzstd.a:
 	$(MAKE) -C $(LIBDIR) libzstd.a-mt
 	cp $(LIBDIR)/libzstd.a .
diff --git a/tests/test-variants.sh b/tests/test-variants.sh
new file mode 100755
index 0000000000..f3a9e065bd
--- /dev/null
+++ b/tests/test-variants.sh
@@ -0,0 +1,115 @@
+#!/bin/sh
+set -e
+set -u
+set -x
+
+
+SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
+PROG_DIR="$SCRIPT_DIR/../programs"
+
+ZSTD="$PROG_DIR/zstd"
+ZSTD_COMPRESS="$PROG_DIR/zstd-compress"
+ZSTD_DECOMPRESS="$PROG_DIR/zstd-decompress"
+ZSTD_NOLEGACY="$PROG_DIR/zstd-nolegacy"
+ZSTD_DICTBUILDER="$PROG_DIR/zstd-dictBuilder"
+ZSTD_FRUGAL="$PROG_DIR/zstd-frugal"
+ZSTD_NOMT="$PROG_DIR/zstd-nomt"
+
+println() {
+    printf '%b\n' "${*}"
+}
+
+die() {
+    println "$@" 1>&2
+    exit 1
+}
+
+symbol_present() {
+	(nm $1 || echo "symbol_present $@ failed") | grep $2
+}
+
+symbol_not_present() {
+	symbol_present $@ && die "Binary '$1' mistakenly contains symbol '$2'" ||:
+}
+
+compress_not_present() {
+	symbol_not_present "$1" ZSTD_compress
+}
+
+decompress_not_present() {
+	symbol_not_present "$1" ZSTD_decompress
+}
+
+dict_not_present() {
+	symbol_not_present "$1" ZDICT_
+	symbol_not_present "$1" COVER_
+}
+
+cliextra_not_present() {
+	symbol_not_present "$1" TRACE_
+	symbol_not_present "$1" BMK_
+}
+
+legacy_not_present() {
+	symbol_not_present "$1" ZSTDv0
+}
+
+test_help() {
+	"$1" --help | grep -- "$2"
+}
+
+test_no_help() {
+	test_help $@ && die "'$1' supports '$2' when it shouldn't" ||:
+}
+
+extras_not_present() {
+	dict_not_present $@
+	legacy_not_present $@
+	cliextra_not_present $@
+	test_no_help $@ "--train"
+	test_no_help $@ "-b#"
+}
+
+test_compress() {
+	echo "hello" | "$1" | "$ZSTD" -t
+}
+
+test_decompress() {
+	echo "hello" | "$ZSTD" | "$1" -t
+}
+
+test_zstd() {
+	test_compress $@
+	test_decompress $@
+}
+
+extras_not_present "$ZSTD_FRUGAL"
+extras_not_present "$ZSTD_COMPRESS"
+extras_not_present "$ZSTD_DECOMPRESS"
+
+compress_not_present "$ZSTD_DECOMPRESS"
+
+decompress_not_present "$ZSTD_COMPRESS"
+decompress_not_present "$ZSTD_DICTBUILDER"
+
+cliextra_not_present "$ZSTD_DICTBUILDER"
+
+legacy_not_present "$ZSTD_DICTBUILDER"
+legacy_not_present "$ZSTD_NOLEGACY"
+
+symbol_not_present "$ZSTD" ZSTDv01
+symbol_not_present "$ZSTD" ZSTDv02
+symbol_not_present "$ZSTD" ZSTDv03
+symbol_not_present "$ZSTD" ZSTDv04
+
+test_compress "$ZSTD_COMPRESS"
+test_decompress "$ZSTD_DECOMPRESS"
+
+test_zstd "$ZSTD_FRUGAL"
+test_zstd "$ZSTD_NOLEGACY"
+
+test_help "$ZSTD" '-b#'
+test_help "$ZSTD" --train
+test_help "$ZSTD_DICTBUILDER" --train
+
+println "Success!"