Fix some typos and misc. cleanup.

1 files changed, 11 insertions, 10 deletions

diff --git a/unpremultiply-sse2-test.S b/unpremultiply-sse2-test.S
index 1fce594..e8bef21 100644
--- a/unpremultiply-sse2-test.S
+++ b/unpremultiply-sse2-test.S
@@ -1,14 +1,17 @@
-	section .text
 ;;;
 ;;; Unpremultiply routine for SSE2/AMD64.
 ;;;
+;;; This file exports a function unpremultiply_with_sse2_test() that
+;;; can be used to unpremultiply a contiguous buffer of 32 bit pixels.
+;;;
+	section .text
 
 ; We're only using rax-rbp in this file so that
 ; conversion to 32 bit SSE2 would be easier by
 ; updating the register names and the
 ; argument extraction to the calling convention.
 
-; Location of alpha in a 32 bit pixel.
+; Location of alpha in a 32 bit pixel.  Alpha measures opaqueness.
 %define ASHIFT 24
 ;%define ASHIFT 0
 
@@ -45,8 +48,7 @@ reciprocal_table_D:
 %endrep
 
 unpremultiply_single_pixels:
-;; Slower version for the odd pixels at the beginning and
-;; and.
+;; Slower version for the odd pixels at the ends.
 ;;
 ;; In:
 ;;   uint32_t *dst/rdi:		Destination pixels.
@@ -76,7 +78,7 @@ unpremultiply_single_pixels:
 	mov	ebx, eax
 	mov	ebp, eax		; Initialise result pixel register.
 	and	ebp, 0xFF000000		; Mask off non-alpha from result pix.
-	jz	.prolog
+	jz	.next
 	shr	ebx, 24			; Load alpha.
 	mov	ebx, DWORD [reciprocal_table_D + 4*ebx] ; Load reciprocal.
 
@@ -109,7 +111,7 @@ unpremultiply_single_pixels:
 	shr	eax, 8			; Shift out alpha.
 	and	ebp, 255		; Mask off non-alpha.
 	mov	ebx, ebp		; Initialise result pixel.
-	jz	.prolog
+	jz	.next
 	mov	ebx, DWORD [reciprocal_table_D + 4*ebx] ; Load reciprocal.
 
 	; Do the component from bits 8..15.
@@ -135,7 +137,7 @@ unpremultiply_single_pixels:
 	shl	eax, 8
 	or	ebp, eax
 %endif
-.prolog:
+.next:
 	; Write the result pixel.
 	mov	[rdi + rcx*4], ebp
 
@@ -165,7 +167,6 @@ unpremultiply_single_pixels:
 ;;   rsi:		rsi + 16*floor(num_pixels/4); advanced past src.
 ;;
 ;; Scratched: xmm1-xmm4, rax-rdx, rbx
-
 	; Advance the src and dst pointers to the end.  The bias
 	; of +-15 is used to have the loop condition trigger an exit
 	; just before we access the last incomplete block.
@@ -239,7 +240,7 @@ global unpremultiply_with_sse2_test
 
 unpremultiply_with_sse2_test:
 ;;
-;; void unpremultiply_with_sse2(
+;; void unpremultiply_with_sse2_test(
 ;;	uint32_t *dst/rdi,
 ;;	uint32_t const *src/rsi,
 ;;	ulong n/rdx);
@@ -251,7 +252,7 @@ unpremultiply_with_sse2_test:
 	push	rbp
 	push	rbx
 
-	; Save start dst for alignment tests later.
+	; Save start of dst for alignment tests later.
 	mov	rcx, rdi
 
 	; If we don't have enough pixels for at least a few iterations