diff --git a/CMakeLists.txt b/CMakeLists.txt
index efd744a0e1..cae9d44192 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -217,6 +217,413 @@ check_c_source_compiles([[#include <gmp.h>
   #endif
   void main(){};]] LONG_LONG_LIMB)
 
+# longlong.h configuration
+set(CMAKE_REQUIRED_INCLUDES ${GMP_INCLUDE_DIRS})
+check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86_64.h.in"
+#if !defined(__GNUC__) || GMP_LIMB_BITS != 64 || !defined(__amd64__)
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_X86_64)
+
+if(LONGLONG_X86_64)
+    set(LONGLONG ON)
+    configure_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86_64.h.in
+        ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+        COPYONLY)
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm64.h.in"
+#if !defined(__GNUC__) || GMP_LIMB_BITS != 64 || !defined(__arm64__)
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_ARM64)
+
+    if(LONGLONG_ARM64)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm64.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86.h.in"
+#if !defined(__GNUC__) || GMP_LIMB_BITS != 32 || !(defined (__i386__) || defined (__i486__) || defined(__amd64__))
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_X86)
+
+    if(LONGLONG_X86)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm.h.in"
+#if !defined(__GNUC__) || GMP_LIMB_BITS != 32 || !defined(__arm__)
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_ARM)
+
+    if(LONGLONG_ARM)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-generic_gnu.h.in"
+#if !defined(__GNUC__)
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_GENERIC_GNU)
+
+    if(LONGLONG_GENERIC_GNU)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-generic_gnu.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86_win.h.in"
+#if !(defined(_WIN32) || defined(_WIN64)) || !(defined(_M_X64) || defined(_M_IX86))
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_X86_WIN)
+
+    if(LONGLONG_X86_WIN)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-x86_win.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm_win.h.in"
+#if !(defined(_WIN32) || defined(_WIN64)) || !(defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM_ARMV7VE) || defined(_M_ARM_FP) || defined(_M_ARM64EC))
+# error
+error
+#endif
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_ARM_WIN)
+
+    if(LONGLONG_ARM_WIN)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-arm_win.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    endif()
+endif()
+
+if(NOT LONGLONG)
+    check_c_source_compiles([[#include <gmp.h>
+typedef mp_limb_t ulong;
+#define FLINT_ASSERT(x)
+#define FLINT_DLL
+#define FLINT_BITS GMP_LIMB_BITS
+#include "${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-generic.h.in"
+int main(void)
+{
+    ulong s3, s2, s1, s0;
+    ulong a3, a2, a1, a0;
+    ulong b3, b2, b1, b0;
+
+    a0 = 19827;
+    a1 = 1872;
+    a2 = 338237;
+    a3 = 98080;
+    b0 = 1798291;
+    b1 = 719271;
+    b2 = 891;
+    b3 = 9112;
+
+    s0 = flint_clz(a0);
+    s1 = flint_ctz(a0);
+    add_ssaaaa(s1, s0, a1, a0, b1, b0);
+    add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+    sub_ddmmss(s1, s0, a1, a0, b1, b0);
+    sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+    umul_ppmm(s1, s0, a0, b0);
+    smul_ppmm(s1, s0, a0, b0);
+    udiv_qrnnd(s1, s0, a1, a0, b0);
+    sdiv_qrnnd(s1, s0, a1, a0, b0);
+    byte_swap(s2);
+    /* udiv_qrnnd_preinv is the same for every header */
+    return 0;
+}]] LONGLONG_GENERIC)
+
+    if(LONGLONG_GENERIC)
+        set(LONGLONG ON)
+        configure_file(
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong-generic.h.in
+            ${CMAKE_CURRENT_SOURCE_DIR}/src/longlong.h
+            COPYONLY)
+    else()
+        message(FATAL_ERROR "Could not find a suitable longlong.h that would compile.")
+    endif()
+endif()
+
+
 # Populate headers
 configure_file(
     ${CMAKE_CURRENT_SOURCE_DIR}/CMake/cmake_config.h.in
diff --git a/Makefile.in b/Makefile.in
index b962a10670..66f3da52c5 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -833,6 +833,9 @@ src/fmpz/fmpz.c: @FMPZ_C_IN@ config.status
 src/gmpcompat.h: src/@GMPCOMPAT_H_IN@ config.status
 	./config.status $@
 
+src/longlong.h: @LONGLONG_H_IN@ config.status
+	./config.status $@
+
 ################################################################################
 # maintainer stuff
 ################################################################################
diff --git a/acinclude.m4 b/acinclude.m4
index 63cc07e6cd..6ea5dda83e 100644
--- a/acinclude.m4
+++ b/acinclude.m4
@@ -53,3 +53,51 @@ AS_VAR_IF(CACHEVAR,yes,
   [m4_default([$3], :)])
 AS_VAR_POPDEF([CACHEVAR])dnl
 ])dnl AX_CHECK_COMPILE_FLAGS
+
+# AX_CHECK_LONGLONG_HEADER(FLAG, [ACTION-SUCCESS], [ACTION-FAILURE], [EXTRA-PROLOGUE])
+AC_DEFUN([AX_CHECK_LONGLONG_HEADER],
+[AS_VAR_PUSHDEF([CACHEVAR],[ax_cv_check_$1])dnl
+AC_CACHE_CHECK([whether system can compile with $1], CACHEVAR, [
+AC_COMPILE_IFELSE(
+    [AC_LANG_PROGRAM(
+    [[#include <gmp.h>
+      typedef mp_limb_t ulong;
+      #define FLINT_ASSERT(x)
+      #define FLINT_DLL
+      #define FLINT_BITS GMP_LIMB_BITS
+      #include "$1"
+      $4]],
+    [[ulong s3, s2, s1, s0;
+      ulong a3, a2, a1, a0;
+      ulong b3, b2, b1, b0;
+
+      a0 = 19827;
+      a1 = 1872;
+      a2 = 338237;
+      a3 = 98080;
+      b0 = 1798291;
+      b1 = 719271;
+      b2 = 891;
+      b3 = 9112;
+
+      s0 = flint_clz(a0);
+      s1 = flint_ctz(a0);
+      add_ssaaaa(s1, s0, a1, a0, b1, b0);
+      add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+      add_ssssaaaaaaaa(s3, s2, s1, s0, a3, a2, a1, a0, b3, b2, b1, b0);
+      sub_ddmmss(s1, s0, a1, a0, b1, b0);
+      sub_dddmmmsss(s2, s1, s0, a2, a1, a0, b2, b1, b0);
+      umul_ppmm(s1, s0, a0, b0);
+      smul_ppmm(s1, s0, a0, b0);
+      udiv_qrnnd(s1, s0, a1, a0, b0);
+      sdiv_qrnnd(s1, s0, a1, a0, b0);
+      byte_swap(s2);
+      /* udiv_qrnnd_preinv is the same for every header */]]
+    )],
+    [AS_VAR_SET(CACHEVAR,[yes])],
+    [AS_VAR_SET(CACHEVAR,[no])])])
+AS_VAR_IF(CACHEVAR,yes,
+  [m4_default([$2], :)],
+  [m4_default([$3], :)])
+AS_VAR_POPDEF([CACHEVAR])dnl
+])dnl AX_CHECK_LONGLONG_HEADER
diff --git a/configure.ac b/configure.ac
index f1b58b7e18..2e23f720d2 100644
--- a/configure.ac
+++ b/configure.ac
@@ -919,7 +919,6 @@ then
     AX_CHECK_COMPILE_FLAG([-march=native],[DEFAULT_CFLAGS="-march=native $DEFAULT_CFLAGS"])
 fi
 AX_CHECK_COMPILE_FLAG([-std=c11],[DEFAULT_CFLAGS="-std=c11 $DEFAULT_CFLAGS"])
-AX_CHECK_COMPILE_FLAG([-pedantic],[DEFAULT_CFLAGS="-pedantic $DEFAULT_CFLAGS"])
 
 dnl NOTE: The following flags shouldn't be default. They are leftovers from
 dnl previous configure script
@@ -942,6 +941,90 @@ fi
 
 CPPFLAGS="-I./src $CPPFLAGS"
 
+################################################################################
+# check which longlong.h to use
+################################################################################
+
+save_CFLAGS="$CFLAGS"
+CFLAGS="-Wall"
+
+AX_CHECK_LONGLONG_HEADER([src/longlong-x86_64.h.in],
+                         [longlong_h_in="src/longlong-x86_64.h.in"],[],
+                         [#if !defined(__GNUC__) || GMP_LIMB_BITS != 64 || !defined(__amd64__)
+                          # error
+                          error
+                          #endif])
+if test ! -n "$longlong_h_in";
+then
+    AX_CHECK_LONGLONG_HEADER([src/longlong-arm64.h.in],
+                             [longlong_h_in="src/longlong-arm64.h.in"],[],
+                             [#if !defined(__GNUC__) || GMP_LIMB_BITS != 64 || !(defined(__aarch64__) || defined(__ARM_ARCH_ISA_A64))
+                              # error
+                              error
+                              #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-x86.h.in],
+                         [longlong_h_in="src/longlong-x86.h.in"],[],
+                         [#if !defined(__GNUC__) || GMP_LIMB_BITS != 32 || !(defined (__i386__) || defined (__i486__) || defined(__amd64__))
+                          # error
+                          error
+                          #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-arm.h.in],
+                         [longlong_h_in="src/longlong-arm.h.in"],[],
+                         [#if !defined(__GNUC__) || GMP_LIMB_BITS != 32 || !defined(__arm__)
+                          # error
+                          error
+                          #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-generic_gnu.h.in],
+                         [longlong_h_in="src/longlong-generic_gnu.h.in"],[],
+                         [#if !defined(__GNUC__)
+                          # error
+                          error
+                          #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-x86_win.h.in],
+                         [longlong_h_in="src/longlong-x86_win.h.in"],[],
+                         [#if !(defined(_WIN32) || defined(_WIN64)) || !(defined(_M_X64) || defined(_M_IX86))
+                          # error
+                          error
+                          #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-arm_win.h.in],
+                         [longlong_h_in="src/longlong-arm_win.h.in"],[],
+                         [#if !(defined(_WIN32) || defined(_WIN64)) || !(defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM_ARMV7VE) || defined(_M_ARM_FP) || defined(_M_ARM64EC))
+                          # error
+                          error
+                          #endif])
+fi
+if test ! -n "$longlong_h_in";
+then
+AX_CHECK_LONGLONG_HEADER([src/longlong-generic.h.in],
+                         [longlong_h_in="src/longlong-generic.h.in"])
+fi
+
+if test ! -n "$longlong_h_in";
+then
+    AC_MSG_ERROR([Could not find a suitable longlong.h. Please submit a bug report to
+<https://github.com/flintlib/flint/issues/> and specify your operating system.])
+fi
+
+AC_CONFIG_FILES([src/longlong.h:$longlong_h_in],[],[longlong_h_in="$longlong_h_in"])
+AC_SUBST(LONGLONG_H_IN, $longlong_h_in)
+
+CFLAGS="$save_CFLAGS"
+
 ################################################################################
 # fft_small module
 ################################################################################
diff --git a/src/longlong-generic_gnu.h.in b/src/longlong-generic_gnu.h.in
index 7251323d66..8ca48087a6 100644
--- a/src/longlong-generic_gnu.h.in
+++ b/src/longlong-generic_gnu.h.in
@@ -46,17 +46,17 @@ extern "C" {
 #  define _flint_adc __builtin_addcl
 #  define _flint_sbc __builtin_subcl
 # else
-#  define _flint_adc(a, b, carry_in, carry_out)             \
-  ({ ulong __s;                                             \
-   ulong c1 = __builtin_uaddl_overflow(a, b, &__s);         \
-   ulong c2 = __builtin_uaddl_overflow(__s, carry_in, &__s);\
-   *(carry_out) = c1 | c2;                                  \
+#  define _flint_adc(a, b, carry_in, carry_out)               \
+  ({ ulong __s;                                               \
+   ulong __c1 = __builtin_uaddl_overflow(a, b, &__s);         \
+   ulong __c2 = __builtin_uaddl_overflow(__s, carry_in, &__s);\
+   *(carry_out) = __c1 | __c2;                                \
    __s; })
-#  define _flint_sbc(a, b, carry_in, carry_out)             \
-  ({ ulong __s;                                             \
-   ulong c1 = __builtin_usubl_overflow(a, b, &__s);         \
-   ulong c2 = __builtin_usubl_overflow(__s, carry_in, &__s);\
-   *(carry_out) = c1 | c2;                                  \
+#  define _flint_sbc(a, b, carry_in, carry_out)               \
+  ({ ulong __s;                                               \
+   ulong __c1 = __builtin_usubl_overflow(a, b, &__s);         \
+   ulong __c2 = __builtin_usubl_overflow(__s, carry_in, &__s);\
+   *(carry_out) = __c1 | __c2;                                \
    __s; })
 # endif
 #else
@@ -68,17 +68,17 @@ extern "C" {
 #  define _flint_adc __builtin_addcll
 #  define _flint_sbc __builtin_subcll
 # else
-#  define _flint_adc(a, b, carry_in, carry_out)               \
-  ({ ulong __s;                                               \
-   ulong c1 = __builtin_uaddll_overflow(a, b, &__s);          \
-   ulong c2 = __builtin_uaddll_overflow(__s, carry_in, &__s); \
-   *(carry_out) = c1 | c2;                                    \
+#  define _flint_adc(a, b, carry_in, carry_out)                 \
+  ({ ulong __s;                                                 \
+   ulong __c1 = __builtin_uaddll_overflow(a, b, &__s);          \
+   ulong __c2 = __builtin_uaddll_overflow(__s, carry_in, &__s); \
+   *(carry_out) = __c1 | __c2;                                  \
    __s; })
-#  define _flint_sbc(a, b, carry_in, carry_out)               \
-  ({ ulong __s;                                               \
-   ulong c1 = __builtin_usubll_overflow(a, b, &__s);          \
-   ulong c2 = __builtin_usubll_overflow(__s, carry_in, &__s); \
-   *(carry_out) = c1 | c2;                                    \
+#  define _flint_sbc(a, b, carry_in, carry_out)                 \
+  ({ ulong __s;                                                 \
+   ulong __c1 = __builtin_usubll_overflow(a, b, &__s);          \
+   ulong __c2 = __builtin_usubll_overflow(__s, carry_in, &__s); \
+   *(carry_out) = __c1 | __c2;                                  \
    __s; })
 # endif
 #endif
@@ -125,76 +125,64 @@ extern "C" {
 #if GMP_LIMB_BITS == 32
 # define umul_ppmm(w1, w0, u, v)      \
   do {                                \
-    unsigned long long int res, a, b; \
-    a = (unsigned long long int) (u); \
-    b = (unsigned long long int) (v); \
-    res = a * b;                      \
-    (w1) = res >> 32;                 \
-    (w0) = res;                       \
+    unsigned long long int __res = (unsigned long long int) (u) * (unsigned long long int) (v); \
+    (w1) = __res >> 32;               \
+    (w0) = __res;                     \
   } while (0)
 
 # define smul_ppmm(w1, w0, u, v)  \
   do {                            \
-    long long int res, a, b;      \
-    a = (long long int) (u);      \
-    b = (long long int) (v);      \
-    res = a * b;                  \
-    (w1) = res >> 32;             \
-    (w0) = res;                   \
+    long long int __res = (long long int) (u) * (long long int) (v); \
+    (w1) = __res >> 32;           \
+    (w0) = __res;                 \
   } while (0)
 
-# define udiv_qrnnd(q, r, n1, n0, d)        \
-  do {                                      \
-    unsigned long long int n;               \
-    n = ((unsigned long long int) (n1)) << 32 + (unsigned long long int) (n0); \
-    (q) = n / (unsigned long long int) (d); \
-    (r) = n % (unsigned long long int) (d); \
+# define udiv_qrnnd(q, r, n1, n0, d)          \
+  do {                                        \
+    unsigned long long int __n;               \
+    __n = ((unsigned long long int) (n1)) << 32 + (unsigned long long int) (n0); \
+    (q) = __n / (unsigned long long int) (d); \
+    (r) = __n % (unsigned long long int) (d); \
   } while (0)
 
-# define sdiv_qrnnd(q, r, n1, n0, d)  \
-  do {                                \
-    long long int n;                  \
-    n = ((long long int) (n1)) << 32 + (long long int) (n0); \
-    (q) = n / (long long int) (d);    \
-    (r) = n % (long long int) (d);    \
+# define sdiv_qrnnd(q, r, n1, n0, d)    \
+  do {                                  \
+    long long int __n;                  \
+    __n = ((long long int) (n1)) << 32 + (long long int) (n0); \
+    (q) = __n / (long long int) (d);    \
+    (r) = __n % (long long int) (d);    \
   } while (0)
 
 # define byte_swap(n) do { n = __builtin_bswap32(n); } while (0)
 #else
-# define umul_ppmm(w1, w0, u, v)  \
-  do {                            \
-    unsigned __int128 res, a, b;  \
-    a = (unsigned __int128) (u);  \
-    b = (unsigned __int128) (v);  \
-    res = a * b;                  \
-    (w1) = res >> 64;             \
-    (w0) = res;                   \
+# define umul_ppmm(w1, w0, u, v)\
+  do {                          \
+    unsigned __int128 __res = (unsigned __int128) (u) * (unsigned __int128) (v); \
+    (w1) = __res >> 64;         \
+    (w0) = __res;               \
   } while (0)
 
 # define smul_ppmm(w1, w0, u, v)\
   do {                          \
-    __int128 res, a, b;         \
-    a = (__int128) (u);         \
-    b = (__int128) (v);         \
-    res = a * b;                \
-    (w1) = res >> 64;           \
-    (w0) = res;                 \
+    __int128 __res = (unsigned __int128) (u) * (unsigned __int128) (v); \
+    (w1) = __res >> 64;         \
+    (w0) = __res;               \
   } while (0)
 
 # define udiv_qrnnd(q, r, n1, n0, d)    \
   do {                                  \
-    unsigned __int128 n;                \
-    n = (((unsigned __int128) (n1)) << 64) + (unsigned __int128) (n0); \
-    (q) = n / (unsigned __int128) (d);  \
-    (r) = n % (unsigned __int128) (d);  \
+    unsigned __int128 __n;              \
+    __n = (((unsigned __int128) (n1)) << 64) + (unsigned __int128) (n0); \
+    (q) = __n / (unsigned __int128) (d);\
+    (r) = __n % (unsigned __int128) (d);\
   } while (0)
 
-# define sdiv_qrnnd(q, r, n1, n0, d)\
-  do {                              \
-    __int128 n;                     \
-    n = (((__int128) (n1)) << 64) + (__int128) (n0); \
-    (q) = n / (__int128) (d);       \
-    (r) = n % (__int128) (d);       \
+# define sdiv_qrnnd(q, r, n1, n0, d)  \
+  do {                                \
+    __int128 __n;                     \
+    __n = (((__int128) (n1)) << 64) + (__int128) (n0); \
+    (q) = __n / (__int128) (d);       \
+    (r) = __n % (__int128) (d);       \
   } while (0)
 
 # define byte_swap(n) do { n = __builtin_bswap64(n); } while (0)