freedreno/ir3: add cat7 instructions

Needed for memory and execution barriers.

Signed-off-by: Rob Clark <[email protected]>

4 files changed, 79 insertions, 2 deletions

diff --git a/src/gallium/drivers/freedreno/ir3/disasm-a3xx.c b/src/gallium/drivers/freedreno/ir3/disasm-a3xx.c
index 4685ed6deae..96b34649c37 100644
--- a/src/gallium/drivers/freedreno/ir3/disasm-a3xx.c
+++ b/src/gallium/drivers/freedreno/ir3/disasm-a3xx.c
@@ -714,6 +714,23 @@ static void print_instr_cat6(instr_t *instr)
 	}
 }
 
+static void print_instr_cat7(instr_t *instr)
+{
+	instr_cat7_t *cat7 = &instr->cat7;
+
+	if (cat7->g)
+		printf(".g");
+	if (cat7->l)
+		printf(".l");
+
+	if (_OPC(7, cat7->opc) == OPC_FENCE) {
+		if (cat7->r)
+			printf(".r");
+		if (cat7->w)
+			printf(".w");
+	}
+}
+
 /* size of largest OPC field of all the instruction categories: */
 #define NOPC_BITS 6
 
@@ -879,6 +896,8 @@ static const struct opc_info {
 	OPC(6, OPC_LDC,          ldc),
 	OPC(6, OPC_LDLV,         ldlv),
 
+	OPC(7, OPC_BAR,          bar),
+	OPC(7, OPC_FENCE,        fence),
 
 #undef OPC
 };
@@ -909,7 +928,7 @@ static void print_instr(uint32_t *dwords, int level, int n)
 
 	if (instr->sync)
 		printf("(sy)");
-	if (instr->ss && (instr->opc_cat <= 4))
+	if (instr->ss && ((instr->opc_cat <= 4) || (instr->opc_cat == 7)))
 		printf("(ss)");
 	if (instr->jmp_tgt)
 		printf("(jp)");
diff --git a/src/gallium/drivers/freedreno/ir3/instr-a3xx.h b/src/gallium/drivers/freedreno/ir3/instr-a3xx.h
index b429b3b9ffc..9edcc58ce9c 100644
--- a/src/gallium/drivers/freedreno/ir3/instr-a3xx.h
+++ b/src/gallium/drivers/freedreno/ir3/instr-a3xx.h
@@ -195,6 +195,10 @@ typedef enum {
 	OPC_LDC             = _OPC(6, 30),
 	OPC_LDLV            = _OPC(6, 31),
 
+	/* category 7: */
+	OPC_BAR             = _OPC(7, 0),
+	OPC_FENCE           = _OPC(7, 1),
+
 	/* meta instructions (category -1): */
 	/* placeholder instr to mark shader inputs: */
 	OPC_META_INPUT      = _OPC(-1, 0),
@@ -715,6 +719,24 @@ typedef union PACKED {
 	};
 } instr_cat6_t;
 
+typedef struct PACKED {
+	/* dword0: */
+	uint32_t pad1     : 32;
+
+	/* dword1: */
+	uint32_t pad2     : 12;
+	uint32_t ss       : 1;  /* maybe in the encoding, but blob only uses (sy) */
+	uint32_t pad3     : 6;
+	uint32_t w        : 1;  /* write */
+	uint32_t r        : 1;  /* read */
+	uint32_t l        : 1;  /* local */
+	uint32_t g        : 1;  /* global */
+	uint32_t opc      : 4;  /* presumed, but only a couple known OPCs */
+	uint32_t jmp_tgt  : 1;  /* (jp) */
+	uint32_t sync     : 1;  /* (sy) */
+	uint32_t opc_cat  : 3;
+} instr_cat7_t;
+
 typedef union PACKED {
 	instr_cat0_t cat0;
 	instr_cat1_t cat1;
@@ -723,12 +745,13 @@ typedef union PACKED {
 	instr_cat4_t cat4;
 	instr_cat5_t cat5;
 	instr_cat6_t cat6;
+	instr_cat7_t cat7;
 	struct PACKED {
 		/* dword0: */
 		uint64_t pad1     : 40;
 		uint32_t repeat   : 3;  /* cat0-cat4 */
 		uint32_t pad2     : 1;
-		uint32_t ss       : 1;  /* cat1-cat4 (cat0??) */
+		uint32_t ss       : 1;  /* cat1-cat4 (cat0??) and cat7 (?) */
 		uint32_t ul       : 1;  /* cat2-cat4 (and cat1 in blob.. which may be bug??) */
 		uint32_t pad3     : 13;
 		uint32_t jmp_tgt  : 1;
@@ -748,6 +771,7 @@ static inline uint32_t instr_opc(instr_t *instr)
 	case 4:  return instr->cat4.opc;
 	case 5:  return instr->cat5.opc;
 	case 6:  return instr->cat6.opc;
+	case 7:  return instr->cat7.opc;
 	default: return 0;
 	}
 }
diff --git a/src/gallium/drivers/freedreno/ir3/ir3.c b/src/gallium/drivers/freedreno/ir3/ir3.c
index d703f4e7f38..61bc3b5ad47 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3.c
@@ -604,9 +604,28 @@ static int emit_cat6(struct ir3_instruction *instr, void *ptr,
 	return 0;
 }
 
+static int emit_cat7(struct ir3_instruction *instr, void *ptr,
+		struct ir3_info *info)
+{
+	instr_cat7_t *cat7 = ptr;
+
+	cat7->ss      = !!(instr->flags & IR3_INSTR_SS);
+	cat7->w       = instr->cat7.w;
+	cat7->r       = instr->cat7.r;
+	cat7->l       = instr->cat7.l;
+	cat7->g       = instr->cat7.g;
+	cat7->opc     = instr->opc;
+	cat7->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
+	cat7->sync    = !!(instr->flags & IR3_INSTR_SY);
+	cat7->opc_cat = 7;
+
+	return 0;
+}
+
 static int (*emit[])(struct ir3_instruction *instr, void *ptr,
 		struct ir3_info *info) = {
 	emit_cat0, emit_cat1, emit_cat2, emit_cat3, emit_cat4, emit_cat5, emit_cat6,
+	emit_cat7,
 };
 
 void * ir3_assemble(struct ir3 *shader, struct ir3_info *info,
diff --git a/src/gallium/drivers/freedreno/ir3/ir3.h b/src/gallium/drivers/freedreno/ir3/ir3.h
index de7a2a88733..25fddbf00f4 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3.h
+++ b/src/gallium/drivers/freedreno/ir3/ir3.h
@@ -228,6 +228,12 @@ struct ir3_instruction {
 			int dst_offset;
 			int iim_val;          /* for ldgb/stgb, # of components */
 		} cat6;
+		struct {
+			unsigned w : 1;       /* write */
+			unsigned r : 1;       /* read */
+			unsigned l : 1;       /* local */
+			unsigned g : 1;       /* global */
+		} cat7;
 		/* for meta-instructions, just used to hold extra data
 		 * before instruction scheduling, etc
 		 */
@@ -605,6 +611,11 @@ static inline bool is_mem(struct ir3_instruction *instr)
 	return (opc_cat(instr->opc) == 6);
 }
 
+static inline bool is_barrier(struct ir3_instruction *instr)
+{
+	return (opc_cat(instr->opc) == 7);
+}
+
 static inline bool
 is_store(struct ir3_instruction *instr)
 {
@@ -1179,6 +1190,10 @@ INSTR4(ATOMIC_AND);
 INSTR4(ATOMIC_OR);
 INSTR4(ATOMIC_XOR);
 
+/* cat7 instructions: */
+INSTR0(BAR);
+INSTR0(FENCE);
+
 /* ************************************************************************* */
 /* split this out or find some helper to use.. like main/bitset.h.. */