diff --git a/gentpl.py b/gentpl.py
index c986cecb7..d2aa03586 100644
--- a/gentpl.py
+++ b/gentpl.py
@@ -76,6 +76,12 @@ for i in GROUPS["terminfoinkernel"]: GROUPS["terminfomodule"].remove(i)
 # Flattened Device Trees (FDT)
 GROUPS["fdt"] = [ "arm64_efi", "arm_uboot", "arm_efi" ]
 
+# Needs software helpers for division
+# Must match GRUB_DIVISION_IN_SOFTWARE in misc.h
+GROUPS["softdiv"] = GROUPS["arm"] + ["ia64_efi"]
+GROUPS["no_softdiv"]   = GRUB_PLATFORMS[:]
+for i in GROUPS["softdiv"]: GROUPS["no_softdiv"].remove(i)
+
 # Miscelaneous groups schedulded to disappear in future
 GROUPS["i386_coreboot_multiboot_qemu"] = ["i386_coreboot", "i386_multiboot", "i386_qemu"]
 GROUPS["nopc"] = GRUB_PLATFORMS[:]; GROUPS["nopc"].remove("i386_pc")
diff --git a/grub-core/Makefile.core.def b/grub-core/Makefile.core.def
index 42443bc00..b539ccf9d 100644
--- a/grub-core/Makefile.core.def
+++ b/grub-core/Makefile.core.def
@@ -159,6 +159,8 @@ kernel = {
   terminfoinkernel = commands/extcmd.c;
   terminfoinkernel = lib/arg.c;
 
+  softdiv = lib/division.c;
+
   i386 = kern/i386/dl.c;
   i386_xen = kern/i386/dl.c;
 
@@ -1969,6 +1971,12 @@ module = {
   enable = xen;
 };
 
+module = {
+  name = div;
+  common = lib/division.c;
+  enable = no_softdiv;
+};
+
 module = {
   name = div_test;
   common = tests/div_test.c;
diff --git a/grub-core/kern/arm/misc.S b/grub-core/kern/arm/misc.S
index 16b6f8e60..645b42f50 100644
--- a/grub-core/kern/arm/misc.S
+++ b/grub-core/kern/arm/misc.S
@@ -58,15 +58,11 @@ FUNCTION(__aeabi_lmul)
         ldmfd   sp!, {r4, fp}
         bx      lr
 
-	.macro division parent
+	.macro division32 parent
 
         sub     sp, sp, #8      @ Allocate naturally aligned 64-bit space
         stmfd   sp!, {r3,lr}    @ Dummy r3 to maintain stack alignment
-        add     r3, sp, #8      @ Set r3 to address of 64-bit space
-        str     r3, [sp]        @ Stack parameter, pointer to 64-bit space
-        mov     r2, r1
-        mov     r1, #0
-        mov     r3, #0
+        add     r2, sp, #8      @ Set r2 to address of 64-bit space
         bl      \parent
         ldr     r1, [sp, #8]    @ Extract remainder
         ldmfd   sp!, {r3,lr}    @ Pop into an unused arg/scratch register
@@ -75,8 +71,9 @@ FUNCTION(__aeabi_lmul)
 	.endm
 
 FUNCTION(__aeabi_uidivmod)
-	division grub_divmod64
-
+	division32 grub_divmod32
+FUNCTION(__aeabi_idivmod)
+	division32 grub_divmod32s
 
 /*
  * Null divide-by-zero handler
diff --git a/grub-core/kern/misc.c b/grub-core/kern/misc.c
index 883d35327..b0347e7e2 100644
--- a/grub-core/kern/misc.c
+++ b/grub-core/kern/misc.c
@@ -594,10 +594,10 @@ grub_divmod64 (grub_uint64_t n, grub_uint64_t d, grub_uint64_t *r)
   grub_uint64_t m = 0;
 
   /* ARM and IA64 don't have a fast 32-bit division.
-     Using that code would just make us use libgcc routines, calling
-     them twice (once for modulo and once for quotient.
+     Using that code would just make us use software division routines, calling
+     ourselves indirectly and hence getting infinite recursion.
   */
-#if !defined (__arm__) && !defined (__ia64__)
+#if !GRUB_DIVISION_IN_SOFTWARE
   /* Skip the slow computation if 32-bit arithmetic is possible.  */
   if (n < 0xffffffff && d < 0xffffffff)
     {
@@ -633,7 +633,7 @@ grub_divmod64 (grub_uint64_t n, grub_uint64_t d, grub_uint64_t *r)
 
 #if !defined (GRUB_UTIL) && !defined (GRUB_MACHINE_EMU)
 
-#if defined (__arm__)
+#if GRUB_DIVISION_IN_SOFTWARE
 
 grub_uint32_t
 __udivsi3 (grub_uint32_t a, grub_uint32_t b)
@@ -641,6 +641,12 @@ __udivsi3 (grub_uint32_t a, grub_uint32_t b)
   return grub_divmod64 (a, b, 0);
 }
 
+grub_int32_t
+__divsi3 (grub_int32_t a, grub_int32_t b)
+{
+  return grub_divmod64s (a, b, 0);
+}
+
 grub_uint32_t
 __umodsi3 (grub_uint32_t a, grub_uint32_t b)
 {
@@ -649,6 +655,42 @@ __umodsi3 (grub_uint32_t a, grub_uint32_t b)
   return ret;
 }
 
+grub_int32_t
+__modsi3 (grub_int32_t a, grub_int32_t b)
+{
+  grub_int64_t ret;
+  grub_divmod64s (a, b, &ret);
+  return ret;
+}
+
+grub_uint64_t
+__udivdi3 (grub_uint64_t a, grub_uint64_t b)
+{
+  return grub_divmod64 (a, b, 0);
+}
+
+grub_uint64_t
+__umoddi3 (grub_uint64_t a, grub_uint64_t b)
+{
+  grub_uint64_t ret;
+  grub_divmod64 (a, b, &ret);
+  return ret;
+}
+
+grub_int64_t
+__divdi3 (grub_int64_t a, grub_int64_t b)
+{
+  return grub_divmod64s (a, b, 0);
+}
+
+grub_int64_t
+__moddi3 (grub_int64_t a, grub_int64_t b)
+{
+  grub_int64_t ret;
+  grub_divmod64s (a, b, &ret);
+  return ret;
+}
+
 #endif
 
 #ifdef NEED_CTZDI2
@@ -735,30 +777,15 @@ __ctzsi2 (grub_uint32_t x)
 grub_uint32_t
 __aeabi_uidiv (grub_uint32_t a, grub_uint32_t b)
   __attribute__ ((alias ("__udivsi3")));
+grub_int32_t
+__aeabi_idiv (grub_int32_t a, grub_int32_t b)
+  __attribute__ ((alias ("__divsi3")));
 void *__aeabi_memcpy (void *dest, const void *src, grub_size_t n)
   __attribute__ ((alias ("grub_memcpy")));
 void *__aeabi_memset (void *s, int c, grub_size_t n)
   __attribute__ ((alias ("grub_memset")));
 #endif
 
-#if defined (__ia64__)
-
-grub_uint64_t
-__udivdi3 (grub_uint64_t a, grub_uint64_t b)
-{
-  return grub_divmod64 (a, b, 0);
-}
-
-grub_uint64_t
-__umoddi3 (grub_uint64_t a, grub_uint64_t b)
-{
-  grub_uint64_t ret;
-  grub_divmod64 (a, b, &ret);
-  return ret;
-}
-
-#endif
-
 #endif /* GRUB_UTIL */
 
 /* Convert a long long value to a string. This function avoids 64-bit
diff --git a/grub-core/tests/div_test.c b/grub-core/tests/div_test.c
index 7bdc80b40..9abc6d5c4 100644
--- a/grub-core/tests/div_test.c
+++ b/grub-core/tests/div_test.c
@@ -97,6 +97,72 @@ test64 (grub_uint64_t a, grub_uint64_t b)
 #endif
 }
 
+static grub_int64_t
+abs64(grub_int64_t a)
+{
+  return a > 0 ? a : -a;
+}
+
+static void
+test32s (grub_int32_t a, grub_int32_t b)
+{
+  grub_int64_t q, r;
+  if (b == 0)
+    return;
+
+  q = grub_divmod64s (a, b, &r);
+  grub_test_assert (a > 0 ? r >= 0 : r <= 0, "remainder sign mismatch: %lld %% %lld = %lld",
+		    (long long) a, (long long) b, (long long) r);
+  grub_test_assert (((a > 0) == (b > 0)) ? q >= 0 : q <= 0, "quotient sign mismatch: %lld / %lld = %lld",
+		    (long long) a, (long long) b, (long long) q);
+  grub_test_assert (abs64(r) < abs64(b), "remainder is larger than dividend: %lld %% %lld = %lld",
+		    (long long) a, (long long) b, (long long) r);
+  grub_test_assert (q * b + r == a, "division doesn't satisfy base property: %lld * %lld + %lld != %lld", (long long) q, (long long) b, (long long) r,
+		    (long long) a);
+  if (0) {  grub_test_assert (q == (a / b),
+		    "C compiler division failure in 0x%llx, 0x%llx", (long long) a, (long long) b);
+  grub_test_assert (r == (a % b),
+		    "C compiler modulo failure in 0x%llx, 0x%llx", (long long) a, (long long) b);
+  }
+}
+
+static void
+test64s (grub_int64_t a, grub_int64_t b)
+{
+  grub_int64_t q, r;
+  q = grub_divmod64s (a, b, &r);
+
+  grub_test_assert (a > 0 ? r >= 0 : r <= 0, "remainder sign mismatch: %lld %% %lld = %lld",
+		    (long long) a, (long long) b, (long long) r);
+  grub_test_assert (((a > 0) == (b > 0)) ? q >= 0 : q <= 0, "quotient sign mismatch: %lld / %lld = %lld",
+		    (long long) a, (long long) b, (long long) q);
+  grub_test_assert (abs64(r) < abs64(b), "remainder is larger than dividend: %lld %% %lld = %lld",
+		    (long long) a, (long long) b, (long long) r);
+  grub_test_assert (q * b + r == a, "division doesn't satisfy base property: 0x%llx * 0x%llx + 0x%llx != 0x%llx", (long long) q, (long long) b, (long long) r,
+		    (long long) a);
+#if GRUB_TARGET_SIZEOF_VOID_P == 8
+  grub_test_assert (q == (a / b),
+		    "C compiler division failure in 0x%llx, 0x%llx", (long long) a, (long long) b);
+  grub_test_assert (r == (a % b),
+		    "C compiler modulo failure in 0x%llx, 0x%llx", (long long) a, (long long) b);
+#endif
+}
+
+static void
+test_all(grub_uint64_t a, grub_uint64_t b)
+{
+  test64 (a, b);
+  test32 (a, b);
+  test64s (a, b);
+  test32s (a, b);
+  test64s (a, -b);
+  test32s (a, -b);
+  test64s (-a, b);
+  test32s (-a, b);
+  test64s (-a, -b);
+  test32s (-a, -b);
+}
+
 static void
 div_test (void)
 {
@@ -105,8 +171,7 @@ div_test (void)
 
   for (i = 0; i < ARRAY_SIZE (vectors); i++)
     {
-      test64 (vectors[i][0], vectors[i][1]);
-      test32 (vectors[i][0], vectors[i][1]);
+      test_all (vectors[i][0], vectors[i][1]);
     }
   for (i = 0; i < 40000; i++)
     {
@@ -116,9 +181,7 @@ div_test (void)
 	b = 1;
       if (a == 0)
 	a = 1;
-      test64 (a, b);
-      test32 (a, b);
-
+      test_all (a, b);
     }
 }
 
diff --git a/include/grub/misc.h b/include/grub/misc.h
index e35934cbd..a05b3fd39 100644
--- a/include/grub/misc.h
+++ b/include/grub/misc.h
@@ -358,6 +358,64 @@ grub_uint64_t EXPORT_FUNC(grub_divmod64) (grub_uint64_t n,
 					  grub_uint64_t d,
 					  grub_uint64_t *r);
 
+/* Must match softdiv group in gentpl.py.  */
+#if !defined(GRUB_MACHINE_EMU) && (defined(__arm__) || defined(__ia64__))
+#define GRUB_DIVISION_IN_SOFTWARE 1
+#else
+#define GRUB_DIVISION_IN_SOFTWARE 0
+#endif
+
+/* Some division functions need to be in kernel if compiler generates calls
+   to them. Otherwise we still need them for consistent tests but they go
+   into a separate module.  */
+#if GRUB_DIVISION_IN_SOFTWARE
+#define EXPORT_FUNC_IF_SOFTDIV EXPORT_FUNC
+#else
+#define EXPORT_FUNC_IF_SOFTDIV(x) x
+#endif
+
+grub_int64_t
+EXPORT_FUNC_IF_SOFTDIV(grub_divmod64s) (grub_int64_t n,
+					grub_int64_t d,
+					grub_int64_t *r);
+grub_uint32_t
+EXPORT_FUNC_IF_SOFTDIV (grub_divmod32) (grub_uint32_t n,
+					grub_uint32_t d,
+					grub_uint32_t *r);
+
+grub_int32_t
+EXPORT_FUNC_IF_SOFTDIV (grub_divmod32s) (grub_int32_t n,
+					 grub_int32_t d,
+					 grub_int32_t *r);
+
+#if defined(GRUB_DIVISION_IN_SOFTWARE) && GRUB_DIVISION_IN_SOFTWARE
+
+grub_uint32_t
+EXPORT_FUNC (__udivsi3) (grub_uint32_t a, grub_uint32_t b);
+
+grub_uint32_t
+EXPORT_FUNC (__umodsi3) (grub_uint32_t a, grub_uint32_t b);
+
+grub_int32_t
+EXPORT_FUNC (__divsi3) (grub_int32_t a, grub_int32_t b);
+
+grub_int32_t
+EXPORT_FUNC (__modsi3) (grub_int32_t a, grub_int32_t b);
+
+grub_int64_t
+EXPORT_FUNC (__divdi3) (grub_int64_t a, grub_int64_t b);
+
+grub_int64_t
+EXPORT_FUNC (__moddi3) (grub_int64_t a, grub_int64_t b);
+
+grub_uint64_t
+EXPORT_FUNC (__udivdi3) (grub_uint64_t a, grub_uint64_t b);
+
+grub_uint64_t
+EXPORT_FUNC (__umoddi3) (grub_uint64_t a, grub_uint64_t b);
+
+#endif
+
 #if (defined (__MINGW32__) || defined (__CYGWIN__)) && !defined(GRUB_UTIL)
 void EXPORT_FUNC (__register_frame_info) (void);
 void EXPORT_FUNC (__deregister_frame_info) (void);
@@ -437,12 +495,6 @@ grub_error_load (const struct grub_error_saved *save)
 
 #if defined (__arm__)
 
-grub_uint32_t
-EXPORT_FUNC (__udivsi3) (grub_uint32_t a, grub_uint32_t b);
-
-grub_uint32_t
-EXPORT_FUNC (__umodsi3) (grub_uint32_t a, grub_uint32_t b);
-
 #endif
 
 #if defined (__sparc__) || defined (__powerpc__)
@@ -460,8 +512,12 @@ EXPORT_FUNC (__ctzsi2) (grub_uint32_t x);
 #ifdef __arm__
 grub_uint32_t
 EXPORT_FUNC (__aeabi_uidiv) (grub_uint32_t a, grub_uint32_t b);
+grub_int32_t
+EXPORT_FUNC (__aeabi_idiv) (grub_int32_t a, grub_int32_t b);
 grub_uint32_t
 EXPORT_FUNC (__aeabi_uidivmod) (grub_uint32_t a, grub_uint32_t b);
+grub_int32_t
+EXPORT_FUNC (__aeabi_idivmod) (grub_int32_t a, grub_int32_t b);
 
 /* Needed for allowing modules to be compiled as thumb.  */
 grub_uint64_t
@@ -476,16 +532,6 @@ EXPORT_FUNC(__aeabi_memset) (void *s, int c, grub_size_t n);
 
 #endif
 
-#if defined (__ia64__)
-
-grub_uint64_t
-EXPORT_FUNC (__udivdi3) (grub_uint64_t a, grub_uint64_t b);
-
-grub_uint64_t
-EXPORT_FUNC (__umoddi3) (grub_uint64_t a, grub_uint64_t b);
-
-#endif
-
 #endif /* GRUB_UTIL */