wined3d: Implement SM5 gather4_po_c opcode.

dlls/wined3d/arb_program_shader.c

@@ -5128,6 +5128,7 @@ static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABL
     /* WINED3DSIH_GATHER4                          */ NULL,
     /* WINED3DSIH_GATHER4_C                        */ NULL,
     /* WINED3DSIH_GATHER4_PO                       */ NULL,
+    /* WINED3DSIH_GATHER4_PO_C                     */ NULL,
     /* WINED3DSIH_GE                               */ NULL,
     /* WINED3DSIH_HS_CONTROL_POINT_PHASE           */ NULL,
     /* WINED3DSIH_HS_DECLS                         */ NULL,

dlls/wined3d/glsl_shader.c

@@ -5877,9 +5877,11 @@ static void shader_glsl_gather4(const struct wined3d_shader_instruction *ins)
     }
 
     has_offset = ins->handler_idx == WINED3DSIH_GATHER4_PO
+            || ins->handler_idx == WINED3DSIH_GATHER4_PO_C
             || wined3d_shader_instruction_has_texel_offset(ins);
 
-    resource_param_idx = ins->handler_idx == WINED3DSIH_GATHER4_PO ? 2 : 1;
+    resource_param_idx =
+            (ins->handler_idx == WINED3DSIH_GATHER4_PO || ins->handler_idx == WINED3DSIH_GATHER4_PO_C) ? 2 : 1;
     resource_idx = ins->src[resource_param_idx].reg.idx[0].offset;
     sampler_idx = ins->src[resource_param_idx + 1].reg.idx[0].offset;
     component_idx = shader_glsl_swizzle_get_component(ins->src[resource_param_idx + 1].swizzle, 0);
@@ -5902,12 +5904,12 @@ static void shader_glsl_gather4(const struct wined3d_shader_instruction *ins)
 
     shader_addline(buffer, "textureGather%s(%s_sampler%u, %s",
             has_offset ? "Offset" : "", prefix, sampler_bind_idx, coord_param.param_str);
-    if (ins->handler_idx == WINED3DSIH_GATHER4_C)
+    if (ins->handler_idx == WINED3DSIH_GATHER4_C || ins->handler_idx == WINED3DSIH_GATHER4_PO_C)
     {
-        shader_glsl_add_src_param(ins, &ins->src[3], WINED3DSP_WRITEMASK_0, &compare_param);
+        shader_glsl_add_src_param(ins, &ins->src[resource_param_idx + 2], WINED3DSP_WRITEMASK_0, &compare_param);
         shader_addline(buffer, ", %s", compare_param.param_str);
     }
-    if (ins->handler_idx == WINED3DSIH_GATHER4_PO)
+    if (ins->handler_idx == WINED3DSIH_GATHER4_PO || ins->handler_idx == WINED3DSIH_GATHER4_PO_C)
     {
         shader_glsl_add_src_param(ins, &ins->src[1], (1u << offset_size) - 1, &offset_param);
         shader_addline(buffer, ", %s", offset_param.param_str);
@@ -10681,6 +10683,7 @@ static const SHADER_HANDLER shader_glsl_instruction_handler_table[WINED3DSIH_TAB
     /* WINED3DSIH_GATHER4                          */ shader_glsl_gather4,
     /* WINED3DSIH_GATHER4_C                        */ shader_glsl_gather4,
     /* WINED3DSIH_GATHER4_PO                       */ shader_glsl_gather4,
+    /* WINED3DSIH_GATHER4_PO_C                     */ shader_glsl_gather4,
     /* WINED3DSIH_GE                               */ shader_glsl_relop,
     /* WINED3DSIH_HS_CONTROL_POINT_PHASE           */ shader_glsl_nop,
     /* WINED3DSIH_HS_DECLS                         */ shader_glsl_nop,

dlls/wined3d/shader.c

@@ -148,6 +148,7 @@ static const char * const shader_opcode_names[] =
     /* WINED3DSIH_GATHER4                          */ "gather4",
     /* WINED3DSIH_GATHER4_C                        */ "gather4_c",
     /* WINED3DSIH_GATHER4_PO                       */ "gather4_po",
+    /* WINED3DSIH_GATHER4_PO_C                     */ "gather4_po_c",
     /* WINED3DSIH_GE                               */ "ge",
     /* WINED3DSIH_HS_CONTROL_POINT_PHASE           */ "hs_control_point_phase",
     /* WINED3DSIH_HS_DECLS                         */ "hs_decls",
@@ -1586,7 +1587,8 @@ static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const st
                 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
                         ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
             }
-            else if (ins.handler_idx == WINED3DSIH_GATHER4_PO)
+            else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
+                    || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
             {
                 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
                         ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);

dlls/wined3d/shader_sm4.c

@@ -235,6 +235,7 @@ enum wined3d_sm4_opcode
     WINED3D_SM5_OP_DERIV_RTY_FINE                   = 0x7d,
     WINED3D_SM5_OP_GATHER4_C                        = 0x7e,
     WINED3D_SM5_OP_GATHER4_PO                       = 0x7f,
+    WINED3D_SM5_OP_GATHER4_PO_C                     = 0x80,
     WINED3D_SM5_OP_RCP                              = 0x81,
     WINED3D_SM5_OP_F32TOF16                         = 0x82,
     WINED3D_SM5_OP_F16TOF32                         = 0x83,
@@ -1014,6 +1015,7 @@ static const struct wined3d_sm4_opcode_info opcode_table[] =
     {WINED3D_SM5_OP_DERIV_RTY_FINE,                   WINED3DSIH_DSY_FINE,                         "f",    "f"},
     {WINED3D_SM5_OP_GATHER4_C,                        WINED3DSIH_GATHER4_C,                        "f",    "fRSf"},
     {WINED3D_SM5_OP_GATHER4_PO,                       WINED3DSIH_GATHER4_PO,                       "f",    "fiRS"},
+    {WINED3D_SM5_OP_GATHER4_PO_C,                     WINED3DSIH_GATHER4_PO_C,                     "f",    "fiRSf"},
     {WINED3D_SM5_OP_RCP,                              WINED3DSIH_RCP,                              "f",    "f"},
     {WINED3D_SM5_OP_F32TOF16,                         WINED3DSIH_F32TOF16,                         "u",    "f"},
     {WINED3D_SM5_OP_F16TOF32,                         WINED3DSIH_F16TOF32,                         "f",    "u"},

dlls/wined3d/wined3d_private.h

@@ -773,6 +773,7 @@ enum WINED3D_SHADER_INSTRUCTION_HANDLER
     WINED3DSIH_GATHER4,
     WINED3DSIH_GATHER4_C,
     WINED3DSIH_GATHER4_PO,
+    WINED3DSIH_GATHER4_PO_C,
     WINED3DSIH_GE,
     WINED3DSIH_HS_CONTROL_POINT_PHASE,
     WINED3DSIH_HS_DECLS,