BI_MOV,
BI_REDUCE_FMA,
BI_SELECT,
- BI_SHIFT,
BI_STORE,
BI_STORE_VAR,
BI_SPECIAL, /* _FAST on supported GPUs */
BI_TABLE,
BI_TEX,
BI_ROUND,
+ BI_IMUL,
BI_NUM_CLASSES
};
BI_IMATH_SUB,
};
+enum bi_imul_op {
+ BI_IMUL_IMUL,
+};
+
enum bi_table_op {
/* fp32 log2() with low precision, suitable for GL or half_log2() in
* CL. In the first argument, takes x. Letting u be such that x =
* exp2() in GL. In the first argument, it takes f2i_rte(x * 2^24). In
* the second, it takes x itself. */
BI_SPECIAL_EXP2_LOW,
+ BI_SPECIAL_IABS,
};
enum bi_tex_op {
enum bi_frexp_op frexp;
enum bi_tex_op texture;
enum bi_imath_op imath;
+ enum bi_imul_op imul;
/* For FMA/ADD, should we add a biased exponent? */
bool mscale;
/* Represents the assignment of ports for a given bi_bundle */
-struct bi_registers {
+typedef struct {
/* Register to assign to each port */
unsigned port[4];
/* Whether writes are actually for the last instruction */
bool first_instruction;
-};
+} bi_registers;
/* A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
* leave it NULL; the emitter will fill in a nop. Instructions reference
*/
typedef struct {
- struct bi_registers regs;
+ bi_registers regs;
bi_instruction *fma;
bi_instruction *add;
} bi_bundle;
+struct bi_block;
+
typedef struct {
struct list_head link;
+ /* Link back up for branch calculations */
+ struct bi_block *block;
+
/* A clause can have 8 instructions in bundled FMA/ADD sense, so there
- * can be 8 bundles. But each bundle can have both an FMA and an ADD,
- * so a clause can have up to 16 bi_instructions. Whether bundles or
- * instructions are used depends on where in scheduling we are. */
+ * can be 8 bundles. */
- unsigned instruction_count;
unsigned bundle_count;
-
- union {
- bi_instruction *instructions[16];
- bi_bundle bundles[8];
- };
+ bi_bundle bundles[8];
/* For scoreboarding -- the clause ID (this is not globally unique!)
* and its dependencies in terms of other clauses, computed during
/* Corresponds to the usual bit but shifted by a clause */
bool data_register_write_barrier;
- /* Constants read by this clause. ISA limit. */
+ /* Constants read by this clause. ISA limit. Must satisfy:
+ *
+ * constant_count + bundle_count <= 13
+ *
+ * Also implicitly constant_count <= bundle_count since a bundle only
+ * reads a single constant.
+ */
uint64_t constants[8];
unsigned constant_count;
+ /* Branches encode a constant offset relative to the program counter
+ * with some magic flags. By convention, if there is a branch, its
+ * constant will be last. Set this flag to indicate this is required.
+ */
+ bool branch_constant;
+
/* What type of high latency instruction is here, basically */
unsigned clause_type;
} bi_clause;
/* During NIR->BIR */
nir_function_impl *impl;
bi_block *current_block;
- unsigned block_name_count;
bi_block *after_block;
bi_block *break_block;
bi_block *continue_block;
#define bi_foreach_block_from(ctx, from, v) \
list_for_each_entry_from(pan_block, v, from, &ctx->blocks, link)
+#define bi_foreach_block_from_rev(ctx, from, v) \
+ list_for_each_entry_from_rev(pan_block, v, from, &ctx->blocks, link)
+
#define bi_foreach_instr_in_block(block, v) \
list_for_each_entry(bi_instruction, v, &(block)->base.instructions, link)
#define bi_foreach_clause_in_block(block, v) \
list_for_each_entry(bi_clause, v, &(block)->clauses, link)
+#define bi_foreach_clause_in_block_from(block, v, from) \
+ list_for_each_entry_from(bi_clause, v, from, &(block)->clauses, link)
+
+#define bi_foreach_clause_in_block_from_rev(block, v, from) \
+ list_for_each_entry_from_rev(bi_clause, v, from, &(block)->clauses, link)
+
#define bi_foreach_instr_global(ctx, v) \
bi_foreach_block(ctx, v_block) \
bi_foreach_instr_in_block((bi_block *) v_block, v)
void bi_invalidate_liveness(bi_context *ctx);
bool bi_is_live_after(bi_context *ctx, bi_block *block, bi_instruction *start, int src);
+/* Layout */
+
+bool bi_can_insert_bundle(bi_clause *clause, bool constant);
+unsigned bi_clause_quadwords(bi_clause *clause);
+signed bi_block_offset(bi_context *ctx, bi_clause *start, bi_block *target);
+
/* Code emit */
void bi_pack(bi_context *ctx, struct util_dynarray *emission);