Removed a couple of uses of multiplication registers + fixed a bug wi…

…th screen clearin gin the SA-1 build

patches/shared/shared.asm

@@ -201,6 +201,12 @@ function dma_source_indirect_word(address, bank) = ((bank&$FF)<<24)|(address&$FF
 function dma_size_indirect(address) = $01000000|(address&$FFFFFF)
 
 
+; Same as above, but also indicates that the passed in address lies in the direct page.
+; This lets the %dma_transfer() macro know to use an LDA.b instead of another LDA.
+
+function dma_size_indirect_dp(address) = $02000000|(address&$FFFFFF)
+
+
 ; A simple macro for preparing access to VRAM.
 
 macro configure_vram_access(access_mode, increment_mode, increment_size, address_remap, vram_address)
@@ -254,12 +260,14 @@ macro dma_transfer(channel, dma_settings, destination, source, num_bytes)
 	!temp_num_bytes_resolved #= <num_bytes>
 	!temp_dma_num_bytes_mode #= ((!temp_num_bytes_resolved>>24)&$FF)
 	
-	assert !temp_dma_num_bytes_mode <= 1, "Invalid argument passed to num_bytes of %dma_transfer()."
+	assert !temp_dma_num_bytes_mode <= 2, "Invalid argument passed to num_bytes of %dma_transfer()."
 	
 	if !temp_dma_num_bytes_mode == 0	
 		!temp_dma_num_bytes = "lda.w #<num_bytes>"
 	elseif !temp_dma_num_bytes_mode == 1
 		!temp_dma_num_bytes = "lda <num_bytes>"
+	elseif !temp_dma_num_bytes_mode == 2
+		!temp_dma_num_bytes = "lda.b <num_bytes>"
 	endif
 
 	lda.b #<dma_settings>

patches/vwf_dialogues/vwf_dialogues.asm

@@ -266,22 +266,55 @@ macro vwf_mvn_transfer(bytes, source, destination, current_cpu, ...)
 
 	phb
 if sizeof(...) > 0
-	lda.b #<bytes>	; Calculate starting index
-	sta.w select(!temp_currently_on_sa1_cpu,$2251,$211B)
-	lda.b #<bytes>>>8
-	sta.w select(!temp_currently_on_sa1_cpu,$2252,$211B)
-	stz.w select(!temp_currently_on_sa1_cpu,$2250,$211C)
-	lda <...[0]>
-	sta.w select(!temp_currently_on_sa1_cpu,$2253,$211C)
-	if !temp_currently_on_sa1_cpu
-		stz $2254
-		nop
+	!temp_start_index := <...[0]>
+	!temp_log2 #= floor(log2(<bytes>))
+
+	if 2**!temp_log2 == <bytes>
+		; RPG Hacker: The multiplier is logarithmic - we can use a number of ASL to do the multiplication.
+		; If the exponent gets too large, this might be less performant than an actual multiplication,
+		; but I don't think we'll ever get to those severe dimensions, so I won't optimize for them.
+		lda.b #$00
+		xba
+		lda !temp_start_index
+		rep #$31
+		asl #!temp_log2
+		adc.w #<source>
+		tax
+	elseif !temp_currently_on_sa1_cpu == !vwf_cpu_snes && <bytes> < $100
+		; If the number of bytes is representable as a single byte (we assume the start index to always be
+		; representable as a single byte), and we're currently on the SNES CPU, do the multiplication via the
+		; regular multiplication registers instead of the mode 7 registers (this can prevent certain quirks).
+		lda.b #<bytes>	; Calculate starting index
+		sta $4202
+		lda !temp_start_index
+		sta $4203
+		rep #$31
+		lda.w #<source>
+		nop #2
+		adc $4216
+		tax
+	else
+		; If none of the above is possible, fall back to a generic solution.
+		lda.b #<bytes>	; Calculate starting index
+		sta.w select(!temp_currently_on_sa1_cpu,$2251,$211B)
+		lda.b #<bytes>>>8
+		sta.w select(!temp_currently_on_sa1_cpu,$2252,$211B)
+		stz.w select(!temp_currently_on_sa1_cpu,$2250,$211C)
+		lda !temp_start_index
+		sta.w select(!temp_currently_on_sa1_cpu,$2253,$211C)
+		if !temp_currently_on_sa1_cpu
+			stz $2254
+			nop
+		endif
+
+		rep #$31
+		lda.w select(!temp_currently_on_sa1_cpu,$2306,$2134)	; Add source address
+		adc.w #<source>	; to get new source address
+		tax
 	endif
 
-	rep #$31
-	lda.w select(!temp_currently_on_sa1_cpu,$2306,$2134)	; Add source address
-	adc.w #<source>	; to get new source address
-	tax
+	undef "temp_log2"
+	undef "temp_start_index"
 else
 	rep #$30
 	ldx.w #<source>
@@ -5710,7 +5743,7 @@ endif
 	sep #$20
 
 	%configure_vram_access(VRamAccessMode.Write, VRamIncrementMode.OnHighByte, VRamIncrementSize.1Byte, VRamAddressRemap.None, $02)
-	%dma_transfer(!vwf_dma_channel_nmi, DmaMode.WriteOnce, $2118, dma_source_indirect_word($00, bank(!vwf_buffer_text_box_tilemap)), dma_size_indirect($04))
+	%dma_transfer(!vwf_dma_channel_nmi, DmaMode.WriteOnce, $2118, dma_source_indirect_word($00, bank(!vwf_buffer_text_box_tilemap)), dma_size_indirect_dp($04))
 
 	rep #$20
 	lda !vwf_clear_box_remaining_bytes

patches/vwf_dialogues/vwfroutines.asm

@@ -76,6 +76,7 @@ endmacro
 ;      During processing: Pointer (24-bit) to font, offset by two 8x8 tile (to access right half of 16x16 tiles)
 ; $0E: Current pixel offset into destination memory buffer
 ; $0F: #$01 if this is writing first tile in the current buffer, otherwise #$00
+;      Used as temporary memory during processing
 %vwf_register_shared_routine(VWF_GenerateVWF)
 	!vwf_var_ram_and_gfx_ram_in_same_bank #= equal(bank(!vwf_var_ram), bank(!vwf_gfx_ram))
 
@@ -195,28 +196,30 @@ endif
 	dex
 	bpl.b ..Check2
 .Skip:
-if !use_sa1_mapping
-	lda.b #$01
-	sta.l $2250
-endif
 if !vwf_var_ram_and_gfx_ram_in_same_bank != false
 	ldy.b #$00
 else
 	ldx.b #$00
 	txy
 endif
+	lda.b #$00
+	pha
+
 	lda.w !vwf_current_pixel	; Adjust destination address
 	clc
 	adc.w !vwf_char_width
-	sta.l select(!use_sa1_mapping,$2251,$4204)
-	lda.b #$00
-	sta.l select(!use_sa1_mapping,$2252,$4205)
-	lda.b #$08
-	sta.l select(!use_sa1_mapping,$2253,$4206)
-if !use_sa1_mapping
+	sta $0F
+
+	lsr #3						; Compute quotient of division by 8
+	pha
+
 	lda.b #$00
-	sta.l $2254
-endif
+	pha	
+
+	lda $0F						; Compute remainder of division by 8
+	and.b #%00000111
+	pha
+
 	lda.w !vwf_first_tile		; Skip one step if first tile in line
 	beq .Combine
 	stz.w !vwf_first_tile
@@ -255,7 +258,7 @@ endif
 	iny
 	cpy.b #$60
 	bne.b .Copy
-	lda.l select(!use_sa1_mapping,$2306,$4214)
+	lda $03,s
 	asl
 	pha
 	clc
@@ -266,25 +269,18 @@ endif
 	sep #$20
 	adc.w !vwf_buffer_index
 	sta.w !vwf_buffer_index
-if !use_sa1_mapping
-	lda.b #$00
-	sta.l $2250
-endif
-	lda.l select(!use_sa1_mapping,$2308,$4216)
+	lda $01,s
 	sta.w !vwf_current_pixel
-	lda.b #$20
-	sta.l select(!use_sa1_mapping,$2251,$4202)
-	lda.l select(!use_sa1_mapping,$2306,$4214)
-	sta.l select(!use_sa1_mapping,$2253,$4203)
+	lda $03,s
 	clc
 	adc.w !vwf_current_x
 	sta.w !vwf_current_x
-if !use_sa1_mapping
-	lda.b #$00
-	sta.l $2254
-endif
 	rep #$21
-	lda.l select(!use_sa1_mapping,$2308,$4216)
+	pla
+	pla
+	; The XBA here is necessary, because bytes are stored on the stack in reverse order
+	; due to the fact we use stack-relative indexing to access them.
+	xba
 	adc.b $09
 	sta.b $09
 	dec.w !vwf_num_bytes		; Adjust number of bytes