diff --git a/programs/develop/fasm/trunk/expressi.inc b/programs/develop/fasm/trunk/expressi.inc
index 2bb11e785e..46af470242 100644
--- a/programs/develop/fasm/trunk/expressi.inc
+++ b/programs/develop/fasm/trunk/expressi.inc
@@ -1014,11 +1014,8 @@ calculate_expression:
 	jne	operand_sizes_do_not_match
 	jmp	calculation_loop
       current_offset_label:
-	mov	al,[labels_type]
-	mov	[edi+12],al
-	mov	eax,[org_symbol]
-	mov	[edi+16],eax
 	mov	eax,[current_offset]
+      make_current_offset_label:
 	xor	edx,edx
 	sub	eax,dword [org_origin]
 	sbb	edx,dword [org_origin+4]
@@ -1027,24 +1024,15 @@ calculate_expression:
 	stos	dword [edi]
 	mov	eax,[org_registers]
 	stos	dword [edi]
+	mov	al,[labels_type]
+	mov	[edi-12+12],al
+	mov	eax,[org_symbol]
+	mov	[edi-12+16],eax
 	add	edi,8
 	jmp	calculation_loop
       org_origin_label:
-	mov	al,[labels_type]
-	mov	[edi+12],al
-	mov	eax,[org_symbol]
-	mov	[edi+16],eax
 	mov	eax,[org_start]
-	xor	edx,edx
-	sub	eax,dword [org_origin]
-	sbb	edx,dword [org_origin+4]
-	stos	dword [edi]
-	mov	eax,edx
-	stos	dword [edi]
-	mov	eax,[org_registers]
-	stos	dword [edi]
-	add	edi,8
-	jmp	calculation_loop
+	jmp	make_current_offset_label
       counter_label:
 	mov	eax,[counter]
       make_dword_label_value:
diff --git a/programs/develop/fasm/trunk/tables.inc b/programs/develop/fasm/trunk/tables.inc
index b935927004..5bc98d18e3 100644
--- a/programs/develop/fasm/trunk/tables.inc
+++ b/programs/develop/fasm/trunk/tables.inc
@@ -1714,9 +1714,9 @@ instructions_7:
  db 'roundps',8
  dw sse4_instruction_3a_imm8-assembler
  db 'roundsd',0Bh
- dw sse4_instruction_3a_imm8-assembler
+ dw sse4_sd_instruction_3a_imm8-assembler
  db 'roundss',0Ah
- dw sse4_instruction_3a_imm8-assembler
+ dw sse4_ss_instruction_3a_imm8-assembler
  db 'rsqrtps',52h
  dw sse_ps_instruction-assembler
  db 'rsqrtss',52h
diff --git a/programs/develop/fasm/trunk/version.inc b/programs/develop/fasm/trunk/version.inc
index 3052a9d83f..caf2ce54af 100644
--- a/programs/develop/fasm/trunk/version.inc
+++ b/programs/develop/fasm/trunk/version.inc
@@ -33,7 +33,7 @@
 ; cannot simply be copied and put under another distribution licence
 ; (including the GNU Public Licence).
 
-VERSION_STRING equ "1.67.37"
+VERSION_STRING equ "1.67.38"
 
 VERSION_MAJOR = 1
 VERSION_MINOR = 67
diff --git a/programs/develop/fasm/trunk/x86_64.inc b/programs/develop/fasm/trunk/x86_64.inc
index 462840641c..093b9cb0f6 100644
--- a/programs/develop/fasm/trunk/x86_64.inc
+++ b/programs/develop/fasm/trunk/x86_64.inc
@@ -4821,12 +4821,12 @@ movdq2q_instruction:
 	jmp	nomem_instruction_ready
 
 sse_ps_instruction_imm8:
-	mov	[immediate_size],8
+	mov	[immediate_size],1
 sse_ps_instruction:
 	mov	[mmx_size],16
 	jmp	sse_instruction
 sse_pd_instruction_imm8:
-	mov	[immediate_size],8
+	mov	[immediate_size],1
 sse_pd_instruction:
 	mov	[mmx_size],16
 	mov	[opcode_prefix],66h
@@ -4888,7 +4888,7 @@ sse_instruction:
 	cmp	[operand_size],al
 	jne	invalid_operand_size
       sse_mem_size_ok:
-	cmp	[immediate_size],8
+	cmp	[immediate_size],1
 	je	mmx_imm8
 	cmp	[immediate_size],-1
 	jne	sse_ok
@@ -4908,7 +4908,7 @@ sse_instruction:
 	cmp	ah,16
 	jne	invalid_operand_size
 	mov	bl,al
-	cmp	[immediate_size],8
+	cmp	[immediate_size],1
 	je	mmx_nomem_imm8
 	cmp	[immediate_size],-1
 	jne	sse_nomem_ok
@@ -5405,22 +5405,31 @@ sse4_instruction_38_xmm0:
 	mov	[immediate_size],-1
 	jmp	sse4_instruction_38
 sse4_instruction_38_imm8:
-	mov	[immediate_size],8
+	mov	[immediate_size],1
 sse4_instruction_38:
+	mov	[mmx_size],16
 	mov	[opcode_prefix],66h
 	mov	[base_code],0Fh
 	mov	[supplemental_code],al
 	mov	al,38h
-	mov	[mmx_size],16
 	jmp	sse_instruction
+sse4_ss_instruction_3a_imm8:
+	mov	[immediate_size],1
+	mov	[mmx_size],4
+	jmp	sse4_instruction_3a_setup
+sse4_sd_instruction_3a_imm8:
+	mov	[immediate_size],1
+	mov	[mmx_size],8
+	jmp	sse4_instruction_3a_setup
 sse4_instruction_3a_imm8:
-	mov	[immediate_size],8
+	mov	[immediate_size],1
 sse4_instruction_3a:
+	mov	[mmx_size],16
+      sse4_instruction_3a_setup:
 	mov	[opcode_prefix],66h
 	mov	[base_code],0Fh
 	mov	[supplemental_code],al
 	mov	al,3Ah
-	mov	[mmx_size],16
 	jmp	sse_instruction
 extractps_instruction:
 	mov	[opcode_prefix],66h