From 13efb349646e8de132b78e58d9b703c94758113b Mon Sep 17 00:00:00 2001
From: Christoph Oelckers <c.oelckers@users.noreply.github.com>
Date: Sun, 4 Dec 2016 18:05:34 +0100
Subject: [PATCH] - removed the asm version of R_DrawColumn because it doesn't
provide any significant benefit.
On modern systems it is actually slower than the C version, only on old ones it is marginally faster - but the overall execution time for this function is so low that even in the worst case scenario the minor loss of performance on older systems is still not relevant.
---
src/asm_ia32/tmap.asm | 103 ------------------------------------------
src/r_draw.cpp | 25 +++-------
src/r_draw.h | 17 +++----
3 files changed, 13 insertions(+), 132 deletions(-)
diff --git a/src/asm_ia32/tmap.asm b/src/asm_ia32/tmap.asm
index d9e689ee1..4ffcc370d 100644
--- a/src/asm_ia32/tmap.asm
+++ b/src/asm_ia32/tmap.asm
@@ -598,106 +598,6 @@ dmsdone add esp,8
-;*----------------------------------------------------------------------
-;*
-;* R_DrawColumnP
-;*
-;*----------------------------------------------------------------------
-
-GLOBAL @R_DrawColumnP_ASM@0
-GLOBAL _R_DrawColumnP_ASM
-GLOBAL R_DrawColumnP_ASM
-
- align 16
-
-R_DrawColumnP_ASM:
-_R_DrawColumnP_ASM:
-@R_DrawColumnP_ASM@0:
-
-; count = dc_yh - dc_yl;
-
- mov ecx,[dc_count]
- test ecx,ecx
- jle near rdcpret ; count <= 0: nothing to do, so leave
-
- push ebp ; save registers
- push ebx
- push edi
- push esi
-
-; dest = ylookup[dc_yl] + dc_x + dc_destorg;
-
- mov edi,[dc_dest]
- mov ebp,ecx
- mov ebx,[dc_texturefrac] ; ebx = frac
-rdcp1: sub edi,SPACEFILLER4
- mov ecx,ebx
- shr ecx,16
- mov esi,[dc_source]
- mov edx,[dc_iscale]
- mov eax,[dc_colormap]
-
- cmp BYTE [CPU+66],byte 5
- jg rdcploop2
-
- align 16
-
-; The registers should now look like this:
-;
-; [31 .. 16][15 .. 8][7 .. 0]
-; eax [colormap ]
-; ebx [yi ][yf ]
-; ecx [scratch ]
-; edx [dyi ][dyf ]
-; esi [source texture column ]
-; edi [destination screen pointer ]
-; ebp [counter ]
-;
-
-
-; Note the partial register stalls on anything better than a Pentium
-; That's why there are two versions of this loop.
-
-rdcploop:
- mov cl,[esi+ecx] ; Fetch texel
- xor ch,ch
- add ebx,edx ; increment frac
-rdcp2: add edi,SPACEFILLER4 ; increment destination pointer
- mov cl,[eax+ecx] ; colormap texel
- mov [edi],cl ; Store texel
- mov ecx,ebx
- shr ecx,16
- dec ebp
- jnz rdcploop ; loop
-
- pop esi
- pop edi
- pop ebx
- pop ebp
-rdcpret:
- ret
-
- align 16
-
-rdcploop2:
- movzx ecx,byte [esi+ecx] ; Fetch texel
- add ebx,edx ; increment frac
- mov cl,[eax+ecx] ; colormap texel
-rdcp3: add edi,SPACEFILLER4 ; increment destination pointer
- mov [edi],cl ; Store texel
- mov ecx,ebx
- shr ecx,16
- dec ebp
- jnz rdcploop2 ; loop
-
- pop esi
- pop edi
- pop ebx
- pop ebp
- ret
-
-
-
;*----------------------------------------------------------------------
;*
;* R_DrawFuzzColumnP
@@ -1648,9 +1548,6 @@ ASM_PatchPitch:
_ASM_PatchPitch:
@ASM_PatchPitch@0:
mov eax,[dc_pitch]
- mov [rdcp1+2],eax
- mov [rdcp2+2],eax
- mov [rdcp3+2],eax
mov [s4p+1],eax
mov [a4p+1],eax
mov [ac4p+1],eax
diff --git a/src/r_draw.cpp b/src/r_draw.cpp
index 099904f25..7d8ebe036 100644
--- a/src/r_draw.cpp
+++ b/src/r_draw.cpp
@@ -69,16 +69,11 @@ int scaledviewwidth;
// These get changed depending on the current
// screen depth and asm/no asm.
void (*R_DrawColumnHoriz)(void);
-void (*R_DrawColumn)(void);
void (*R_DrawFuzzColumn)(void);
void (*R_DrawTranslatedColumn)(void);
void (*R_DrawShadedColumn)(void);
void (*R_DrawSpan)(void);
void (*R_DrawSpanMasked)(void);
-void (*R_DrawSpanTranslucent)(void);
-void (*R_DrawSpanMaskedTranslucent)(void);
-void (*R_DrawSpanAddClamp)(void);
-void (*R_DrawSpanMaskedAddClamp)(void);
void (*rt_map4cols)(int,int,int);
//
@@ -171,7 +166,6 @@ void R_InitShadeMaps()
/* */
/************************************/
-#ifndef X86_ASM
//
// A column is a vertical slice/span from a wall texture that,
// given the DOOM style restrictions on the view orientation,
@@ -179,7 +173,7 @@ void R_InitShadeMaps()
// Thus a special case loop for very fast rendering can
// be used. It has also been used with Wolfenstein 3D.
//
-void R_DrawColumnP_C (void)
+void R_DrawColumn (void)
{
int count;
BYTE* dest;
@@ -222,7 +216,7 @@ void R_DrawColumnP_C (void)
} while (--count);
}
}
-#endif
+
// [RH] Just fills a column with a color
void R_FillColumnP (void)
@@ -1192,7 +1186,7 @@ void R_DrawSpanMaskedP_C (void)
}
#endif
-void R_DrawSpanTranslucentP_C (void)
+void R_DrawSpanTranslucent (void)
{
dsfixed_t xfrac;
dsfixed_t yfrac;
@@ -1252,7 +1246,7 @@ void R_DrawSpanTranslucentP_C (void)
}
}
-void R_DrawSpanMaskedTranslucentP_C (void)
+void R_DrawSpanMaskedTranslucent (void)
{
dsfixed_t xfrac;
dsfixed_t yfrac;
@@ -1326,7 +1320,7 @@ void R_DrawSpanMaskedTranslucentP_C (void)
}
}
-void R_DrawSpanAddClampP_C (void)
+void R_DrawSpanAddClamp (void)
{
dsfixed_t xfrac;
dsfixed_t yfrac;
@@ -1392,7 +1386,7 @@ void R_DrawSpanAddClampP_C (void)
}
}
-void R_DrawSpanMaskedAddClampP_C (void)
+void R_DrawSpanMaskedAddClamp (void)
{
dsfixed_t xfrac;
dsfixed_t yfrac;
@@ -2550,12 +2544,10 @@ const BYTE *R_GetColumn (FTexture *tex, int col)
return tex->GetColumn (col, NULL);
}
-
// [RH] Initialize the column drawer pointers
void R_InitColumnDrawers ()
{
#ifdef X86_ASM
- R_DrawColumn = R_DrawColumnP_ASM;
R_DrawColumnHoriz = R_DrawColumnHorizP_C;
R_DrawFuzzColumn = R_DrawFuzzColumnP_ASM;
R_DrawTranslatedColumn = R_DrawTranslatedColumnP_C;
@@ -2572,7 +2564,6 @@ void R_InitColumnDrawers ()
}
#else
R_DrawColumnHoriz = R_DrawColumnHorizP_C;
- R_DrawColumn = R_DrawColumnP_C;
R_DrawFuzzColumn = R_DrawFuzzColumnP_C;
R_DrawTranslatedColumn = R_DrawTranslatedColumnP_C;
R_DrawShadedColumn = R_DrawShadedColumnP_C;
@@ -2580,10 +2571,6 @@ void R_InitColumnDrawers ()
R_DrawSpanMasked = R_DrawSpanMaskedP_C;
rt_map4cols = rt_map4cols_c;
#endif
- R_DrawSpanTranslucent = R_DrawSpanTranslucentP_C;
- R_DrawSpanMaskedTranslucent = R_DrawSpanMaskedTranslucentP_C;
- R_DrawSpanAddClamp = R_DrawSpanAddClampP_C;
- R_DrawSpanMaskedAddClamp = R_DrawSpanMaskedAddClampP_C;
}
// [RH] Choose column drawers in a single place
diff --git a/src/r_draw.h b/src/r_draw.h
index 7d574c1c0..c5e99e610 100644
--- a/src/r_draw.h
+++ b/src/r_draw.h
@@ -65,7 +65,6 @@ extern "C" unsigned int horizspans[4];
// The span blitting interface.
// Hook in assembler or system specific BLT here.
-extern void (*R_DrawColumn)(void);
extern DWORD (*dovline1) ();
extern DWORD (*doprevline1) ();
@@ -103,16 +102,16 @@ void R_SetSpanSource(const BYTE *pixels);
extern void (*R_DrawSpanMasked)(void);
// Span drawing for translucent textures.
-extern void (*R_DrawSpanTranslucent)(void);
+void R_DrawSpanTranslucent(void);
// Span drawing for masked, translucent textures.
-extern void (*R_DrawSpanMaskedTranslucent)(void);
+void R_DrawSpanMaskedTranslucent(void);
// Span drawing for translucent, additive textures.
-extern void (*R_DrawSpanAddClamp)(void);
+void R_DrawSpanAddClamp(void);
// Span drawing for masked, translucent, additive textures.
-extern void (*R_DrawSpanMaskedAddClamp)(void);
+void R_DrawSpanMaskedAddClamp(void);
// [RH] Span blit into an interleaved intermediate buffer
extern void (*R_DrawColumnHoriz)(void);
@@ -193,8 +192,6 @@ void R_DrawFogBoundary (int x1, int x2, short *uclip, short *dclip);
#ifdef X86_ASM
-extern "C" void R_DrawColumnP_Unrolled (void);
-extern "C" void R_DrawColumnP_ASM (void);
extern "C" void R_DrawFuzzColumnP_ASM (void);
void R_DrawShadedColumnP_C (void);
extern "C" void R_DrawSpanP_ASM (void);
@@ -204,7 +201,6 @@ void R_DrawColumnHorizP_C(void);
#else
-void R_DrawColumnP_C (void);
void R_DrawFuzzColumnP_C (void);
void R_DrawShadedColumnP_C (void);
void R_DrawSpanP_C (void);
@@ -212,10 +208,11 @@ void R_DrawSpanMaskedP_C (void);
#endif
+void R_DrawColumn();
void R_DrawColumnHorizP_C(void);
void R_DrawTranslatedColumnP_C(void);
-void R_DrawSpanTranslucentP_C (void);
-void R_DrawSpanMaskedTranslucentP_C (void);
+void R_DrawSpanTranslucent (void);
+void R_DrawSpanMaskedTranslucent (void);
void R_DrawTlatedLucentColumnP_C (void);
#define R_DrawTlatedLucentColumn R_DrawTlatedLucentColumnP_C