diff --git a/Makefile b/Makefile
index 8341ffdb..b5b40a8f 100644
--- a/Makefile
+++ b/Makefile
@@ -86,11 +86,14 @@ else ifneq (,$(findstring $(ARCH), aarch64 arm64 s390x powerpc64 powerpc64le ppc
   POPCNT_CAPABILITY ?= 0
 endif
 
-ifdef USE_SAIS
+ifeq (1, $(if $(or $(USE_SAIS),$(USE_SAIS_OPENMP)),1))
   CPPFLAGS += -Ithird_party/libsais/include
-  CXXFLAGS += -fopenmp -DUSE_SAIS
-  LDFLAGS += -Lthird_party/libsais/lib
-  LDLIBS += -lsais
+ifdef USE_SAIS_OPENMP
+  BUILD_CXXFLAGS += -fopenmp
+endif
+  BUILD_CXXFLAGS += -DUSE_SAIS
+  BUILD_LDFLAGS += -Lthird_party/libsais/lib
+  BUILD_LDLIBS += -lsais
 endif
 
 # msys will always be 32 bit so look at the cpu arch instead.
@@ -322,6 +325,9 @@ else
 $(BOWTIE2_BIN_LIST_SAN): $(error "Compiler does not support...")
 endif
 
+bowtie2-build-%: CXXFLAGS += $(BUILD_CXXFLAGS)
+bowtie2-build-%: LDFLAGS += $(BUILD_LDFLAGS)
+bowtie2-build-%: LDLIBS += $(BUILD_LDLIBS)
 #
 # bowtie2-build targets
 #
@@ -576,6 +582,16 @@ static-libs:
 		rm -rf zstd-1.5.5* ; \
         fi
 
+.PHONY: libsais
+libsais:
+	if [ ! -d $(CURDIR)/third_party/libsais ]; then \
+		git submodule init && git submodule update --recursive ; \
+	fi ; \
+	if [ -n "$(USE_SAIS_OPENMP)" ]; then \
+		export CFLAGS="$(CFLAGS) -fopenmp -O2" ; \
+	fi ; \
+	cd $(CURDIR)/third_party/libsais && $(MAKE) -B CC=$(CC) && $(MAKE) install PREFIX=$$PWD ;
+
 .PHONY: sra-deps
 sra-deps:
 	DL=$$( ( which wget >/dev/null 2>&1 && echo "wget --no-check-certificate" ) || echo "curl -LOk") ; \
diff --git a/blockwise_sa.h b/blockwise_sa.h
index 90be5245..1e7574f1 100644
--- a/blockwise_sa.h
+++ b/blockwise_sa.h
@@ -204,11 +204,18 @@ class SAISBlockwiseSA : public InorderBlockwiseSA<TStr> {
 			__text.wbuf()[__text.length()] = (char)127; // $ is larger than any character in the suffix array
 			_suffixes.resize(__text.length() + 1);
 #ifdef BOWTIE_64BIT_INDEX
+#ifdef _OPENMP
 			libsais64_omp((const uint8_t *)__text.buf(), (int64_t *)(_suffixes.ptr()), (int64_t)_suffixes.size(), 0, NULL, _nthreads);
+#else
+			libsais64((const uint8_t *)__text.buf(), (int64_t *)(_suffixes.ptr()), (int64_t)_suffixes.size(), 0, NULL);
+#endif
 
 #else
+#ifdef _OPENMP
 			libsais_omp((const uint8_t *)__text.buf(), (int32_t *)(_suffixes.ptr()), (int32_t)_suffixes.size(), 0, NULL, _nthreads);
-
+#else
+			libsais((const uint8_t *)__text.buf(), (int32_t *)(_suffixes.ptr()), (int32_t)_suffixes.size(), 0, NULL);
+#endif
 #endif
 			_suffixes[__text.length()] = __text.length();
 		}
@@ -231,7 +238,7 @@ class SAISBlockwiseSA : public InorderBlockwiseSA<TStr> {
 	}
 
 	bool hasMoreBlocks() const {
-		return _i == _suffixes.size();
+		return _i < _suffixes.size();
 	}
 
 private:
diff --git a/bt2_idx.h b/bt2_idx.h
index d34439f5..e35dc298 100644
--- a/bt2_idx.h
+++ b/bt2_idx.h
@@ -1067,13 +1067,17 @@ class Ebwt {
 				bmax = (TIndexOffU)sqrt(s.length());
 				VMSG_NL("bmax defaulted to: " << bmax);
 			}
+#ifndef USE_SAIS
 			int iter = 0;
 			bool first = true;
 			streampos out1pos = out1.tellp();
 			streampos out2pos = out2.tellp();
+#endif
+
 			// Look for bmax/dcv parameters that work.
 			thread_pool pool(nthreads - 1);
 			while(true) {
+#ifndef USE_SAIS
 				if(!first && bmax < 40 && _passMemExc) {
 					cerr << "Could not find approrpiate bmax/dcv settings for building this index." << endl;
 					if(!isPacked()) {
@@ -1107,7 +1111,9 @@ class Ebwt {
 					VMSG_NL(" --dcv " << dcv);
 				}
 				iter++;
+#endif
 				try {
+#ifndef USE_SAIS
 					{
 						VMSG_NL("  Doing ahead-of-time memory usage test");
 						// Make a quick-and-dirty attempt to force a bad_alloc iff
@@ -1135,10 +1141,12 @@ class Ebwt {
 						}
 						VMSG_NL("");
 					}
-					VMSG_NL("Constructing suffix-array element generator");
+#endif
 #ifdef USE_SAIS
+					VMSG_NL("Using SAIS algorithm for constructing suffix array");
 					SAISBlockwiseSA<TStr> bsa(s, bmax, nthreads);
 #else
+					VMSG_NL("Constructing suffix-array element generator");
 					KarkkainenBlockwiseSA<TStr> bsa(s, bmax, nthreads, pool, dcv, seed, _sanity, _passMemExc, _verbose, outfile);
 #endif
 					assert(bsa.suffixItrIsReset());
@@ -1161,6 +1169,10 @@ class Ebwt {
 					}
 					break;
 				} catch(bad_alloc& e) {
+#ifdef USE_SAIS
+					cerr << "Out of memory while constructing suffix array.  Please try using a smaller" << endl;
+					throw 1;
+#else
 					if(_passMemExc) {
 						VMSG_NL("  Ran out of memory; automatically trying more memory-economical parameters.");
 					} else {
@@ -1168,8 +1180,12 @@ class Ebwt {
 						     << "number of blocks by specifying a smaller --bmax or a larger --bmaxdivn" << endl;
 						throw 1;
 					}
+#endif
 				}
+#ifndef USE_SAIS
 				first = false;
+#endif
+
 			}
 			assert(repOk());
 			// Now write reference sequence names on the end