gpu_neon: fix apparent missing msb setting in blend_blocks_textured_add_fourth

[pcsx_rearmed.git] / plugins / gpu_neon / psx_gpu / psx_gpu_arm_neon.S

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
index 110c868..c0199a0 100644 (file)
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S
@@ -194,26 +194,18 @@
 
 .align 4
 
-#ifndef __MACH__
+#include "arm_features.h"
 
-#define function(name)                                                         \
-  .global name;                                                                \
-  .type name, %function;                                                       \
-  name:                                                                        \
+#define function(name) FUNCTION(name):
+
+#ifndef TEXRELS_FORBIDDEN
 
 #define JT_OP_REL(table_label, index_reg, temp)
 #define JT_OP(x...) x
 #define JTE(start, target) target
 
-#define EXTRA_UNSAVED_REGS
-
 #else
 
-#define function(name)                                                         \
-  .globl _##name;                                                              \
-  name:                                                                        \
-  _##name:                                                                     \
-
 #define JT_OP_REL(table_label, index_reg, temp)                                \
   adr temp, table_label;                                                       \
   ldr temp, [temp, index_reg, lsl #2];                                         \
@@ -222,13 +214,12 @@
 #define JT_OP(x...)
 #define JTE(start, target) (target - start)
 
-// r7 is preserved, but add it for EABI alignment..
-#define EXTRA_UNSAVED_REGS r7, r9,
+#endif
 
+#ifdef __MACH__
 #define flush_render_block_buffer _flush_render_block_buffer
 #define setup_sprite_untextured_simple _setup_sprite_untextured_simple
 #define update_texture_8bpp_cache _update_texture_8bpp_cache
-
 #endif
 
 @ r0: psx_gpu
@@ -3944,7 +3935,7 @@ blend_blocks_add_untextured_builder(on)
 #define blend_blocks_subtract_combine_textured()                               \
   vbif.u16 blend_pixels, pixels, blend_mask                                    \
 
-#define blend_blocks_subtract_set_stb_textured()                               \
+#define blend_blocks_subtract_set_stp_textured()                               \
   vorr.u16 blend_pixels, #0x8000                                               \
 
 #define blend_blocks_subtract_msb_mask_textured()                              \
@@ -3954,7 +3945,7 @@ blend_blocks_add_untextured_builder(on)
 
 #define blend_blocks_subtract_combine_untextured()                             \
 
-#define blend_blocks_subtract_set_stb_untextured()                             \
+#define blend_blocks_subtract_set_stp_untextured()                             \
   vorr.u16 blend_pixels, blend_pixels, msb_mask                                \
 
 #define blend_blocks_subtract_msb_mask_untextured()                            \
@@ -4019,7 +4010,7 @@ function(blend_blocks_##texturing##_subtract_##mask_evaluate)                  \
   vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64;                           \
   vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g;                            \
   vand.u16 pixels_rb, pixels_next, d128_0x7C1F;                                \
-  blend_blocks_subtract_set_stb_##texturing();                                 \
+  blend_blocks_subtract_set_stp_##texturing();                                 \
   vand.u16 pixels_g, pixels_next, d128_0x03E0;                                 \
   blend_blocks_subtract_combine_##texturing();                                 \
   blend_blocks_subtract_set_blend_mask_##texturing();                          \
@@ -4047,7 +4038,7 @@ function(blend_blocks_##texturing##_subtract_##mask_evaluate)                  \
                                                                                \
   blend_blocks_subtract_msb_mask_##texturing();                                \
   vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g;                            \
-  blend_blocks_subtract_set_stb_##texturing();                                 \
+  blend_blocks_subtract_set_stp_##texturing();                                 \
   blend_blocks_subtract_combine_##texturing();                                 \
   vbit.u16 blend_pixels, fb_pixels, draw_mask;                                 \
   vst1.u16 { blend_pixels }, [fb_ptr_next];                                    \
@@ -4119,6 +4110,7 @@ function(blend_blocks_textured_add_fourth_##mask_evaluate)                     \
   ldr fb_ptr_next, [pixel_ptr, #28];                                           \
                                                                                \
   vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g;                            \
+  vorr.u16 blend_pixels, #0x8000;  /* stp */                                   \
   vbif.u16 blend_pixels, pixels, blend_mask;                                   \
                                                                                \
   vld1.u32 { pixels }, [pixel_ptr, :128], c_64;                                \
@@ -4154,8 +4146,9 @@ function(blend_blocks_textured_add_fourth_##mask_evaluate)                     \
                                                                                \
  1:                                                                            \
   vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g;                            \
-  vorr.u16 blend_pixels, blend_pixels, msb_mask;                               \
+  vorr.u16 blend_pixels, #0x8000;  /* stp */                                   \
   vbif.u16 blend_pixels, pixels, blend_mask;                                   \
+  vorr.u16 blend_pixels, blend_pixels, msb_mask;                               \
   vbit.u16 blend_pixels, fb_pixels, draw_mask;                                 \
   vst1.u16 { blend_pixels }, [fb_ptr_next];                                    \
                                                                                \