diff --git a/mach/mips/libem/lar4.s b/mach/mips/libem/lar4.s
index 870140255..218496644 100644
--- a/mach/mips/libem/lar4.s
+++ b/mach/mips/libem/lar4.s
@@ -14,7 +14,7 @@
 	jal .aar4
 	nop
 
-	/* pass r2 = size from .aar4 to .los4
+	/* pass r2 = size from .aar4 to .los4 */
 
 	jal .los4
 	nop
diff --git a/mach/mips/libem/los4.s b/mach/mips/libem/los4.s
index be9434a34..bebb7b616 100644
--- a/mach/mips/libem/los4.s
+++ b/mach/mips/libem/los4.s
@@ -11,7 +11,7 @@
 .sect .text
 .define .los4
 .los4:
-	lw r4, 0(sp)            ! r4 = address
+	lw r4, 0(sp)            ! r4 = address of source block
 
 	! Sizes 1 and 2 are handled specially.
 
@@ -26,14 +26,14 @@
 	! Else the size must be a multiple of 4.
 
 	srl r5, r2, 2           ! r5 = number of words
-	addiu sp, sp, 4         ! adjust to end of block
-	subu sp, sp, r4         ! sp = start of block
-	mov r6, sp              ! r6 = start of block
+	addiu sp, sp, 4         ! retract over address
+	subu sp, sp, r2         ! allocate space for destination block
+	mov r6, sp              ! r6 = start of destination block
 
 1:
-	lw at, 0(r2)
+	lw at, 0(r4)
 	sw at, 0(r6)
-	addiu r2, r2, 4
+	addiu r4, r4, 4
 	addiu r6, r6, 4
 	addiu r5, r5, -1
 	bne r5, zero, 1b
@@ -43,13 +43,13 @@
 	nop
 
 byte_sized:
-	lb at, 0(r4)
+	lbu at, 0(r4)
 	sw at, 0(sp)
 	jr ra
 	nop
 
 word_sized:
-	lh at, 0(r4)
+	lhu at, 0(r4)
 	sw at, 0(sp)
 	jr ra
 	nop
diff --git a/mach/mips/libem/sar4.s b/mach/mips/libem/sar4.s
index a1ddea33c..8c70bf553 100644
--- a/mach/mips/libem/sar4.s
+++ b/mach/mips/libem/sar4.s
@@ -14,7 +14,7 @@
 	jal .aar4
 	nop
 
-	/* pass r2 = size from .aar4 to .sts4
+	/* pass r2 = size from .aar4 to .sts4 */
 
 	jal .sts4
 	nop
diff --git a/tests/plat/core/aar_e.e b/tests/plat/core/aar_e.e
index bd8230d7a..0b53e5b50 100644
--- a/tests/plat/core/aar_e.e
+++ b/tests/plat/core/aar_e.e
@@ -71,6 +71,5 @@ descriptor
 3
 
     cal $finished
-
     end
 
diff --git a/tests/plat/core/lar_e.e b/tests/plat/core/lar_e.e
new file mode 100644
index 000000000..650865b04
--- /dev/null
+++ b/tests/plat/core/lar_e.e
@@ -0,0 +1,73 @@
+#
+    mes 2, EM_WSIZE, EM_PSIZE
+
+/*
+ * Does basic testing of lar. Unfortunately, lar throws ERANGE on
+ * error, which we can't catch (or at least, the platforms I've looked at
+ * don't allow it to be caught, those platforms which actually throw it on
+ * error). So we just test the non-throwing cases, not the negative ones.
+ */
+
+#define ARRAY_SIZE 3*EM_WSIZE
+
+array
+	con 1
+	con 2
+	con 3
+
+descriptor
+	con -1			; lower bound
+	con 3			; range
+	con EM_WSIZE	; size of element
+
+	exp $_m_a_i_n
+	pro $_m_a_i_n, 0
+
+	/* Read element -1 */
+
+	lae array
+	loc -1
+	lae descriptor
+	lar EM_WSIZE
+
+	loc 1
+	beq *1
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+1
+
+	/* Read element 0 */
+
+	lae array
+	loc 0
+	lae descriptor
+	lar EM_WSIZE
+
+	loc 2
+	beq *2
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+2
+
+	/* Read element 1 */
+
+	lae array
+	loc 1
+	lae descriptor
+	lar EM_WSIZE
+
+	loc 3
+	beq *3
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+3
+
+    cal $finished
+    end
+
diff --git a/tests/plat/core/sar_e.e b/tests/plat/core/sar_e.e
new file mode 100644
index 000000000..7c6da941b
--- /dev/null
+++ b/tests/plat/core/sar_e.e
@@ -0,0 +1,125 @@
+#
+    mes 2, EM_WSIZE, EM_PSIZE
+
+/*
+ * Does basic testing of sar. Unfortunately, sar throws ERANGE on
+ * error, which we can't catch (or at least, the platforms I've looked at
+ * don't allow it to be caught, those platforms which actually throw it on
+ * error). So we just test the non-throwing cases, not the negative ones.
+ */
+
+#define ARRAY_SIZE 3*EM_WSIZE
+
+array
+	bss ARRAY_SIZE, 0, 0
+
+descriptor
+	con -1			; lower bound
+	con 3			; range
+	con EM_WSIZE	; size of element
+
+element0
+	con 1
+	con 0
+	con 0
+
+element1
+	con 0
+	con 1
+	con 0
+
+element2
+	con 0
+	con 0
+	con 1
+
+	exp $_m_a_i_n
+	pro $_m_a_i_n, 0
+
+	/* Write element -1 */
+
+	cal $clear_array
+	loc 1
+	lae array
+	loc -1
+	lae descriptor
+	sar EM_WSIZE
+
+	lae array
+	loi ARRAY_SIZE
+	lae element0
+	loi ARRAY_SIZE
+	loc ARRAY_SIZE
+	cms
+	zeq *1
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+1
+
+	/* Write element 0 */
+
+	cal $clear_array
+	loc 1
+	lae array
+	loc 0
+	lae descriptor
+	sar EM_WSIZE
+
+	lae array
+	loi ARRAY_SIZE
+	lae element1
+	loi ARRAY_SIZE
+	loc ARRAY_SIZE
+	cms
+	zeq *2
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+2
+
+	/* Write element 1 */
+
+	cal $clear_array
+	loc 1
+	lae array
+	loc 1
+	lae descriptor
+	sar EM_WSIZE
+
+	lae array
+	loi ARRAY_SIZE
+	lae element2
+	loi ARRAY_SIZE
+	loc ARRAY_SIZE
+	cms
+	zeq *3
+
+	loc __LINE__
+	cal $fail
+	ass EM_WSIZE
+3
+
+    cal $finished
+    end
+
+	exp $clear_array
+	pro $clear_array, 0
+
+	loc 0
+	lae array + EM_WSIZE*0
+	sti EM_WSIZE
+
+	loc 0
+	lae array + EM_WSIZE*1
+	sti EM_WSIZE
+
+	loc 0
+	lae array + EM_WSIZE*2
+	sti EM_WSIZE
+
+	ret 0
+	end
+