Added support for blacklisting via the assembler.

Details:
- Added logic to configure that attempts to assemble various small files
  containing select instructions designed to reveal whether binutils
  (specifically, the assembler) supports emitting those instruction sets.
  This information provides additional opportunities to blacklist sub-
  configurations that are unsupported by the environment. Thanks to Devin
  Matthews for pointing me towards a similar solution in TBLIS as an
  example.
- Various other cleanups in configure.
- Reorganized the detection code in the 'build' directory, bringing the
  "auto-detect" configuration detection, libmemkind detection, and new
  instruction set detection codes into a single new subdirectory named
  'detect'.

build/detect/iset/avx.s

@@ -0,0 +1,6 @@
+//
+// Test for AVX instruction set.
+//
+vzeroall
+vmovapd  %ymm0, %ymm1
+vmulpd   %ymm0, %ymm0, %ymm1

build/detect/iset/avx512dq.s

@@ -0,0 +1,6 @@
+//
+// Test for AVX-512dq instruction set.
+//
+vzeroall
+vpmullq %zmm0, %zmm0, %zmm1
+vpmullw %zmm0, %zmm0, %zmm1

build/detect/iset/avx512f.s

@@ -0,0 +1,7 @@
+//
+// Test for AVX-512f instruction set.
+//
+vzeroall
+vmovapd  %zmm0, %zmm1
+vmulpd %zmm0, %zmm0, %zmm1
+vfmadd213pd 0x400(%rax,%rsi,8) {1to8}, %zmm1, %zmm2

build/detect/iset/fma3.s

@@ -0,0 +1,5 @@
+//
+// Test for FMA3 instruction set.
+//
+vzeroall
+vfmadd213pd %ymm0, %ymm1, %ymm2

build/detect/iset/fma4.s

@@ -0,0 +1,5 @@
+//
+// Test for FMA4 instruction set (AMD Bulldozer only).
+//
+vzeroall
+vfmaddpd %ymm0, %ymm1, %ymm2, %ymm3

configure

@@ -825,27 +825,27 @@ select_cc()
 
 auto_detect()
 {
-	local rval autocc config_defines detected_config
+	local cc cflags config_defines detected_config rval
 
-	# The first and only argument: the compiler to use.
-	autocc="$1"
+	# Use the same compiler that was found earlier.
+	cc="${found_cc}"
 
 	# For debugging: reveal what compiler was chosen for auto-detection.
-	#touch "${autocc}.txt"
+	#touch "${cc}.txt"
 
 	# Tweak the flags we use based on the compiler. This is mostly just
 	# an opportunity to turn off annoying warnings that some compilers
 	# may throw off.
-	if [ "${autocc}" == "clang" ]; then
-		autoccflags="-Wno-tautological-compare"
+	if [ "${cc}" == "clang" ]; then
+		cflags="-Wno-tautological-compare"
 	else
-		autoccflags=
+		cflags=
 	fi
 
 	# Locate our source files.
 	bli_arch_c="bli_arch.c"
 	bli_cpuid_c="bli_cpuid.c"
-	main_c="auto_detect.c"
+	main_c="config_detect.c"
 
 	bli_arch_c_filepath=$(find ${dist_path}/frame -name "${bli_arch_c}")
 	bli_cpuid_c_filepath=$(find ${dist_path}/frame -name "${bli_cpuid_c}")
@@ -883,22 +883,21 @@ auto_detect()
 
 	# Compile the auto-detect program using source code inside the
 	# framework.
-	$autocc ${config_defines} \
-	        -DBLIS_CONFIGURETIME_CPUID \
-	        -I${bli_cpuid_h_path} \
-	        -I${bli_arch_h_path} \
-	        -I${bli_typed_h_path} \
-	        -I${bli_mutex_h_path} \
-	        -I${bli_malloc_h_path} \
-	        -std=c99 \
-	        ${autoccflags} \
-	        ${bli_arch_c_filepath} \
-	        ${bli_cpuid_c_filepath} \
-	        ${main_c_filepath} \
-	     -o ${autodetect_x}
+	${cc} ${config_defines} \
+	      -DBLIS_CONFIGURETIME_CPUID \
+	      -I${bli_cpuid_h_path} \
+	      -I${bli_arch_h_path} \
+	      -I${bli_typed_h_path} \
+	      -I${bli_mutex_h_path} \
+	      -I${bli_malloc_h_path} \
+	      -std=c99 \
+	      ${cflags} \
+	      ${bli_arch_c_filepath} \
+	      ${bli_cpuid_c_filepath} \
+	      ${main_c_filepath} \
+	   -o ${autodetect_x}
 
 	# Run the auto-detect program.
-	#detected_config=`./${autodetect_x}`
 	detected_config=$(./${autodetect_x})
 
 	# Remove the executable file.
@@ -910,18 +909,23 @@ auto_detect()
 
 has_libmemkind()
 {
-	local found_cc LDFLAGS_mk libmemkind_bin rval
+	local found_cc main_c main_c_filepath LDFLAGS_mk binname rval
 
 	found_cc="$1"
 
-	# Temporarily add libmemkind to LDFLAGS.
+	# Path to libmemkind detection source file.
+	main_c="libmemkind_detect.c"
+	main_c_filepath=$(find ${dist_path}/build -name "${main_c}")
+
+	# Add libmemkind to LDFLAGS.
 	LDFLAGS_mk="${LDFLAGS} -lmemkind"
 
-	libmemkind_bin="libmemkind-detect.x"
+	# Binary executable filename.
+	binname="libmemkind-detect.x"
 
 	# Attempt to compile a simple main() program that contains a call
 	# to hbw_malloc() and that links to libmemkind.
-	${found_cc} -o ${libmemkind_bin} ${libmemkind_detect_c} ${LDFLAGS_mk} 2> /dev/null
+	${found_cc} -o ${binname} ${main_c_filepath} ${LDFLAGS_mk} 2> /dev/null
 
 	# Depending on the return code from the compile step above, we set
 	# enable_memkind accordingly.
@@ -932,7 +936,7 @@ has_libmemkind()
 	fi
 
 	# Remove the executable generated above.
-	rm -f ${libmemkind_bin}
+	rm -f ./${binname}
 
 	echo "${rval}"
 }
@@ -947,33 +951,69 @@ echowarn()
 	printf "${script_name}: warning: %s\n" "$*" #>&2;
 }
 
-blacklist_add()
+blacklistcc_add()
 {
 	echowarn "${cc_vendor} ${cc_version} does not support '$1'; adding to blacklist."
 	config_blist="${config_blist} $1"
 }
 
+blacklistbu_add()
+{
+	echowarn "binutils ${bu_version} does not support '$1'; adding to blacklist."
+	config_blist="${config_blist} $1"
+}
+
 blacklist_init()
 {
 	config_blist=""
 }
 
-echoerr_unsupported()
+blacklist_cleanup()
+{
+	# Remove duplicates and whitespace from the blacklist.
+	config_blist=$(rm_duplicate_words "${config_blist}")
+	config_blist=$(canonicalize_ws "${config_blist}")
+}
+
+echoerr_unsupportedcc()
 {
 	echoerr "*** Unsupported compiler version: ${cc_vendor} ${cc_version}."
 	exit 1
 }
 
-check_compiler_version()
+get_binutils_version()
 {
-	local found_cc vendor_string
+	binutil="objdump"
 
-	found_cc="$1"
+	# Query the full binutils version string output. This includes the
+	# version string along with (potentially) a bunch of other textual
+    # clutter.
+	bu_string="$(${binutil} --version 2>/dev/null)"
+
+	# Query the binutils version number.
+	# The last part ({ read first rest ; echo $first ; }) is a workaround
+	# to OS X's egrep only returning the first match.
+	bu_version=$(echo ${bu_string} | egrep -o '[0-9]+\.[0-9]+\.?[0-9]*' | { read first rest ; echo ${first} ; })
+
+	# Parse the version number into its major, minor, and revision
+	# components.
+	bu_major=$(echo "${bu_version}" | cut -d. -f1)
+	bu_minor=$(echo "${bu_version}" | cut -d. -f2)
+	bu_revision=$(echo "${bu_version}" | cut -d. -f3)
+
+	echo "${script_name}: found binutils version ${bu_version} (maj: ${bu_major}, min: ${bu_minor}, rev: ${bu_revision})."
+}
+
+get_compiler_version()
+{
+	local cc vendor_string
+
+	cc="${found_cc}"
 
 	# Query the full vendor version string output. This includes the
 	# version number along with (potentially) a bunch of other textual
 	# clutter.
-	vendor_string="$(${found_cc} --version 2>/dev/null)"
+	vendor_string="$(${cc} --version 2>/dev/null)"
 
 	# Query the compiler "vendor" (ie: the compiler's simple name) and
 	# isolate the version number.
@@ -988,7 +1028,14 @@ check_compiler_version()
 	cc_minor=$(echo "${cc_version}" | cut -d. -f2)
 	cc_revision=$(echo "${cc_version}" | cut -d. -f3)
 
-	echo "${script_name}: found ${found_cc} version ${cc_version} (maj: ${cc_major}, min: ${cc_minor}, rev: ${cc_revision})."
+	echo "${script_name}: found ${cc} version ${cc_version} (maj: ${cc_major}, min: ${cc_minor}, rev: ${cc_revision})."
+}
+
+check_compiler()
+{
+	local cc
+
+	cc="${found_cc}"
 
 	#
 	# Compiler requirements
@@ -1023,25 +1070,22 @@ check_compiler_version()
 	# [1] https://github.com/devinamatthews/tblis/
 	#
 
-	echo "${script_name}: checking for blacklisted configurations for ${found_cc} ${cc_version}."
-
-	# Initialize the blacklist to empty.
-	blacklist_init
+	echo "${script_name}: checking for blacklisted configurations for ${cc} ${cc_version}."
 
 	# gcc
 	if [ "x${cc_vendor}" = "xgcc" ]; then
 
 		if [ ${cc_major} -lt 4 ]; then
-			echoerr_unsupported
+			echoerr_unsupportedcc
 		fi
 		if [ ${cc_major} -eq 4 ]; then
-			blacklist_add "knl"
+			blacklistcc_add "knl"
 			if [ ${cc_minor} -lt 7 ]; then
-				echoerr_unsupported
+				echoerr_unsupportedcc
 			fi
 			if [ ${cc_minor} -lt 9 ]; then
-				blacklist_add "excavator"
-				blacklist_add "zen"
+				blacklistcc_add "excavator"
+				blacklistcc_add "zen"
 			fi
 		fi
 		if [ ${cc_major} -lt 6 ]; then
@@ -1049,9 +1093,9 @@ check_compiler_version()
 			# However, we have a workaround in place in the zen
 			# configuration's make_defs.mk file that starts with bdver4
 			# and disables the instructions that were removed in znver1.
-			# Thus, this "blacklist_add" statement has been moved above.
-			#blacklist_add "zen"
-			blacklist_add "skx"
+			# Thus, this "blacklistcc_add" statement has been moved above.
+			#blacklistcc_add "zen"
+			blacklistcc_add "skx"
 		fi
 	fi
 
@@ -1059,11 +1103,11 @@ check_compiler_version()
 	if [ "x${cc_vendor}" = "xicc" ]; then
 
 		if [ ${cc_major} -lt 15 ]; then
-			echoerr_unsupported
+			echoerr_unsupportedcc
 		fi
 		if [ ${cc_major} -eq 15 ]; then
 			if [ ${cc_revision} -lt 1 ]; then
-				blacklist_add "skx"
+				blacklistcc_add "skx"
 			fi
 		fi
 	fi
@@ -1072,29 +1116,139 @@ check_compiler_version()
 	if [ "x${cc_vendor}" = "xclang" ]; then
 
 		if [ ${cc_major} -lt 3 ]; then
-			echoerr_unsupported
+			echoerr_unsupportedcc
 		fi
 		if [ ${cc_major} -eq 3 ]; then
 			if [ ${cc_minor} -lt 3 ]; then
-				echoerr_unsupported
+				echoerr_unsupportedcc
 			fi
 			if [ ${cc_minor} -lt 5 ]; then
-				blacklist_add "excavator"
-				blacklist_add "zen"
-				blacklist_add "knl"
+				blacklistcc_add "excavator"
+				blacklistcc_add "zen"
+				blacklistcc_add "knl"
 			fi
 			if [ ${cc_minor} -lt 9 ]; then
-				blacklist_add "skx"
+				blacklistcc_add "skx"
 			fi
 		fi
 		if [ ${cc_major} -lt 4 ]; then
 			# See comment above regarding zen support.
-			#blacklist_add "zen"
+			#blacklistcc_add "zen"
 			: # explicit no-op since bash can't handle empty loop bodies.
 		fi
 	fi
+}
 
-	config_blist=$(canonicalize_ws "${config_blist}")
+check_assembler()
+{
+	local cc asm_dir cflags asm_fp
+
+	cc="${found_cc}"
+
+	# The directory where the assembly files will be.
+	asm_dir="${dist_path}/build"
+
+	# Most of the time, we won't need any additional compiler flags.
+	cflags=""
+
+	echo "${script_name}: checking for blacklisted configurations for binutils ${bu_version}."
+
+	#
+	# Check support for FMA4 (amd: bulldozer).
+	#
+	asm_fp=$(find ${asm_dir} -name "fma4.s")
+	knows_fma4=$(try_assemble "${cc}" "${cflags}" "${asm_fp}")
+
+	if [ "x${knows_fma4}" == "xno" ]; then
+		blacklistbu_add "bulldozer"
+	fi
+
+	#
+	# Check support for AVX (intel: sandybridge+, amd: piledriver+).
+	#
+	asm_fp=$(find ${asm_dir} -name "avx.s")
+	knows_avx=$(try_assemble "${cc}" "${cflags}" "${asm_fp}")
+
+	if [ "x${knows_avx}" == "xno" ]; then
+		blacklistbu_add "sandybridge"
+	fi
+
+	#
+	# Check support for FMA3 (intel: haswell+, amd: piledriver+).
+	#
+	asm_fp=$(find ${asm_dir} -name "fma3.s")
+	knows_fma3=$(try_assemble "${cc}" "${cflags}" "${asm_fp}")
+
+	if [ "x${knows_fma3}" == "xno" ]; then
+		blacklistbu_add "haswell"
+		blacklistbu_add "piledriver"
+		blacklistbu_add "steamroller"
+		blacklistbu_add "excavator"
+		blacklistbu_add "skx"
+	fi
+
+	#
+	# Check support for AVX-512f (knl, skx).
+	#
+
+	# The assembler on OS X won't recognize AVX-512 without help.
+	if [ "$(uname -s)" == "Darwin" ]; then
+		cflags="-Wa,-march=knl"
+	fi
+
+	asm_fp=$(find ${asm_dir} -name "avx512f.s")
+	knows_avx512f=$(try_assemble "${cc}" "${cflags}" "${asm_fp}")
+
+	if [ "x${knows_avx512f}" == "xno" ]; then
+		blacklistbu_add "knl"
+		blacklistbu_add "skx"
+	fi
+
+	#
+	# Check support for AVX-512dq (skx).
+	#
+
+	# The assembler on OS X won't recognize AVX-512 without help.
+	if [ "$(uname -s)" == "Darwin" ]; then
+		cflags="-Wa,-march=skylake-avx512"
+	fi
+
+	asm_fp=$(find ${asm_dir} -name "avx512dq.s")
+	knows_avx512dq=$(try_assemble "${cc}" "${cflags}" "${asm_fp}")
+
+	if [ "x${knows_avx512dq}" == "xno" ]; then
+		blacklistbu_add "skx"
+	fi
+}
+
+try_assemble()
+{
+	local cc cflags asm_src asm_base asm_bin rval
+
+	cc="$1"
+	cflags="$2"
+	asm_src="$3"
+
+	# Construct the filename to the .o file corresponding to asm_src.
+	# (Strip the filepath, then the file extension, and then add ".o".)
+	asm_base=${asm_src##*/}
+	asm_base=${asm_base%.*}
+	asm_bin="${asm_base}.o"
+
+	# Try to assemble the file.
+	${cc} ${cflags} -c ${asm_src} -o ${asm_bin} > /dev/null 2>&1
+
+	if [ "$?" == 0 ]; then
+	    rval='yes'
+	else
+	    rval='no'
+	fi
+
+	# Remove the object file.
+	rm -f "${asm_bin}"
+
+	# Return the result.
+	echo "${rval}"
 }
 
 #
@@ -1157,9 +1311,6 @@ main()
 	gen_make_frags_dirpath="${build_dirpath}/gen-make-frags"
 	gen_make_frags_sh="${gen_make_frags_dirpath}/gen-make-frag.sh"
 
-	# Path to libmemkind-detect.c file.
-	libmemkind_detect_c="${build_dirpath}/memkind/libmemkind_detect.c"
-
 	# The name of the (top-level) configuration directory.
 	config_dir='config'
 	config_dirpath="${dist_path}/${config_dir}"
@@ -1389,10 +1540,24 @@ main()
 	found_cc=$(select_cc)
 	echo "${script_name}: using '${found_cc}' compiler."
 
-	# Check the found compiler's version. Certain versions of certain
-	# compilers will preclude building certain sub-configurations, which
-	# are added to a blacklist.
-	check_compiler_version "${found_cc}"
+	# Initialize the blacklist to empty.
+	blacklist_init
+
+	# Check the compiler's version. Certain versions of certain compilers
+	# will preclude building certain sub-configurations, which are added
+	# to a blacklist.
+	get_compiler_version
+	check_compiler
+
+	# Now check the assembler's ability to assemble code. Older versions
+	# of binutils may not be aware of certain instruction sets. Those
+	# sub-configurations employing kernels that use such instruction sets
+	# will also be blacklisted.
+	get_binutils_version
+	check_assembler
+
+	# Remove duplicates and whitespace from the blacklist.
+	blacklist_cleanup
 
 	if [ -n "${config_blist}" ]; then
 
@@ -1485,7 +1650,7 @@ main()
 
 		# Call the auto_detect() function and save the returned string in
 		# config_name.
-		config_name=$(auto_detect "${found_cc}")
+		config_name=$(auto_detect)
 
 		echo "${script_name}: hardware detection driver returned '${config_name}'."
 	else