Fragment asm: initial support of PSEQ fetch instruction and couple of minor changes

src/libgrate/fragment_asm.h

@@ -252,6 +252,18 @@ typedef union fragment_alu_instruction_x4 {
 } alu_instr;
 
 typedef union fragment_pseq_instruction {
+	struct __attribute__((packed)) {
+		unsigned unk_0:1;
+		unsigned dst_regs_select:1;
+		unsigned unk_2:1;
+		unsigned enable0:1;
+		unsigned unk_4_15:12;
+		unsigned rt_select:4;
+		unsigned unk_20_22:3;
+		unsigned enable1:1;
+		unsigned unk_24_31:8;
+	};
+
 	uint32_t data;
 } pseq_instr;
 

src/libgrate/fragment_asm.l

@@ -39,7 +39,14 @@ int yywrap(void)
 
 %%
 [ \t]+			/* skip */;
-"/*"([^*]|(\*+[^*/]))*\*+\/ /* multi line comment */;
+"/*"([^*]|(\*+[^*/]))*\*+\/ {
+				/* multi line comment */
+				unsigned i;
+
+				for (i = 0; i < strlen(yytext); i++)
+					if (yytext[i] == '\n')
+						yylineno++;
+			}
 "//".*			/* single line comment */;
 [\n]			yylineno++;
 
@@ -209,10 +216,12 @@ TXB			return T_TXB_OPCODE;
 
 rt[0-9]{1,2}		{
 				fragment_asmlval.u = atoi(yytext + 2);
-				return T_DW_RENDER_TARGET;
+				return T_RENDER_TARGET;
 			}
 
 "stencil"		return T_DW_STENCIL;
 
+FETCH			return T_FETCH;
+
 .			return T_SYNTAX_ERROR;
 %%

src/libgrate/fragment_asm.y

@@ -21,6 +21,7 @@
  */
 
 %{
+#include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -65,6 +66,8 @@ unsigned asm_pseq_to_dw_exec_nb;
 
 int asm_discards_fragment;
 
+static bool asm_alu_used[4];
+
 static void reset_fragment_asm_parser_state(void)
 {
 	int i, k;
@@ -94,6 +97,8 @@ static void reset_fragment_asm_parser_state(void)
 	asm_discards_fragment = 0;
 
 	fragment_asmlineno = 1;
+
+	memset(asm_alu_used, 0, sizeof(asm_alu_used));
 }
 
 static uint32_t float_to_fp20(float f)
@@ -228,7 +233,7 @@ static uint32_t float_to_fx10(float f)
 
 %token T_DW_STORE
 %token T_DW_STENCIL
-%token <u> T_DW_RENDER_TARGET
+%token <u> T_RENDER_TARGET
 
 %token <u> T_HEX
 %token <f> T_FLOAT
@@ -263,6 +268,8 @@ static uint32_t float_to_fx10(float f)
 %token T_ALU_BUFFER_SIZE
 %token T_PSEQ_DW_EXEC_NB
 
+%token T_FETCH
+
 %union {
 	char c;
 	float f;
@@ -481,6 +488,27 @@ PSEQ_INSTRUCTION:
 		asm_pseq_instructions[asm_fs_instructions_nb].data = $2;
 	}
 	|
+	T_PSEQ T_FETCH T_ROW_REGISTER ',' T_ROW_REGISTER ',' T_RENDER_TARGET
+	{
+		if ($3 == 0 && $5 == 1) {
+			asm_pseq_instructions[asm_fs_instructions_nb].dst_regs_select =  0;
+		}
+		else if ($3 == 2 && $5 == 3) {
+			asm_pseq_instructions[asm_fs_instructions_nb].dst_regs_select =  1;
+		}
+		else {
+			PARSE_ERROR("PSEQ destination registers should be either \"r0,r1\" or \"r2,r3\"");
+		}
+
+		if ($7 > 15) {
+			PARSE_ERROR("Invalid PSEQ render target, 15 is maximum");
+		}
+
+		asm_pseq_instructions[asm_fs_instructions_nb].rt_select = $7;
+		asm_pseq_instructions[asm_fs_instructions_nb].enable0 = 1;
+		asm_pseq_instructions[asm_fs_instructions_nb].enable1 = 1;
+	}
+	|
 	T_PSEQ T_OPCODE_NOP
 	|
 	;
@@ -831,6 +859,8 @@ ALU_INSTRUCTION: T_ALU ALUX_INSTRUCTIONS
 		asm_alu_sched[asm_fs_instructions_nb].instructions_nb++;
 		asm_alu_sched[asm_fs_instructions_nb].address = address;
 		asm_alu_instructions_nb++;
+
+		memset(asm_alu_used, 0, sizeof(asm_alu_used));
 	}
 	;
 
@@ -848,13 +878,25 @@ ALUX_INSTRUCTIONS:
 ALUX_INSTRUCTION:
 	T_ALUX ALU_OPERATION
 	{
+		if (asm_alu_used[$1]) {
+			PARSE_ERROR("ALU has been overridden");
+		} else {
+			asm_alu_used[$1] = true;
+		}
+
 		asm_alu_instructions[asm_alu_instructions_nb].a[$1] = $2;
 	}
 	|
 	T_ALUX ALU3_IMMEDIATES
 	{
 		uint32_t swap = asm_alu_instructions[asm_alu_instructions_nb].part7;
 
+		if (asm_alu_used[$1]) {
+			PARSE_ERROR("ALU has been overridden");
+		} else {
+			asm_alu_used[$1] = true;
+		}
+
 		if ($1 != 3) {
 			PARSE_ERROR("ALU immediates can override ALU3 only");
 		}
@@ -1488,7 +1530,7 @@ DW_INSTRUCTION:
 		asm_dw_instructions[asm_fs_instructions_nb].data = $2;
 	}
 	|
-	T_DW T_DW_STORE T_DW_RENDER_TARGET ',' T_ROW_REGISTER ',' T_ROW_REGISTER
+	T_DW T_DW_STORE T_RENDER_TARGET ',' T_ROW_REGISTER ',' T_ROW_REGISTER
 	{
 		asm_dw_instructions[asm_fs_instructions_nb].enable = 1;
 

src/libgrate/fragment_disasm.c

@@ -704,6 +704,29 @@ static const char * disassemble_dw(const dw_instr *dw)
 	return ret;
 }
 
+static const char * disassemble_pseq(const pseq_instr *pseq)
+{
+	static char ret[64];
+	char *buf = ret;
+
+	buf += sprintf(buf, "fetch ");
+
+	if (pseq->dst_regs_select) {
+		buf += sprintf(buf, "r2, r3, ");
+	} else {
+		buf += sprintf(buf, "r0, r1, ");
+	}
+
+	buf += sprintf(buf, "rt%u", pseq->rt_select);
+
+	if (pseq->unk_0 || pseq->unk_2 || pseq->unk_4_15 ||
+	    pseq->unk_20_22 || pseq->unk_24_31) {
+		sprintf(buf, " // 0x%08X", pseq->data);
+	}
+
+	return ret;
+}
+
 const char * fragment_pipeline_disassemble(
 	const pseq_instr *pseq,
 	const mfu_instr *mfu, unsigned mfu_nb,
@@ -726,7 +749,7 @@ const char * fragment_pipeline_disassemble(
 	buf += sprintf(buf, "EXEC\n");
 
 	if (pseq->data != 0x00000000) {
-		buf += sprintf(buf, "\tPSEQ:\t0x%08X\n", pseq->data);
+		buf += sprintf(buf, "\tPSEQ:\t%s\n", disassemble_pseq(pseq));
 	}
 
 	for (i = 0; i < mfu_nb; i++) {

src/libgrate/linker_asm.l

@@ -39,7 +39,14 @@ int yywrap(void)
 
 %%
 [ \t]+				/* skip */;
-"/*"([^*]|(\*+[^*/]))*\*+\/ 	/* multi line comment */;
+"/*"([^*]|(\*+[^*/]))*\*+\/ {
+				/* multi line comment */
+				unsigned i;
+
+				for (i = 0; i < strlen(yytext); i++)
+					if (yytext[i] == '\n')
+						yylineno++;
+			}
 "//".*				/* single line comment */;
 [\n]				yylineno++;
 "."				return '.';

src/libgrate/vertex_asm.l

@@ -38,7 +38,14 @@ int yywrap(void)
 
 %%
 [ \t]+			/* skip */;
-"/*"([^*]|(\*+[^*/]))*\*+\/ /* multi line comment */;
+"/*"([^*]|(\*+[^*/]))*\*+\/ {
+				/* multi line comment */
+				unsigned i;
+
+				for (i = 0; i < strlen(yytext); i++)
+					if (yytext[i] == '\n')
+						yylineno++;
+			}
 "//".*			/* single line comment */;
 [\n]			yylineno++;