panfrost: Use pack for Bifrost test state

[mesa.git] / src / panfrost / bifrost / compiler.h

diff --git a/src/panfrost/bifrost/compiler.h b/src/panfrost/bifrost/compiler.h
index 43bd8bbc09facfa5fe8720464b8687fc82ba5f40..32361cc37e1c44d16776bb90eb64332f92c1a7f9 100644 (file)
--- a/src/panfrost/bifrost/compiler.h
+++ b/src/panfrost/bifrost/compiler.h
@@ -70,13 +70,13 @@ enum bi_class {
         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
 };
 
@@ -172,6 +172,10 @@ enum bi_imath_op {
         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 =
@@ -200,6 +204,7 @@ enum bi_special_op {
          * 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 {
@@ -282,6 +287,7 @@ typedef struct {
                 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;
@@ -335,21 +341,19 @@ typedef struct {
         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
@@ -373,10 +377,22 @@ typedef struct {
         /* 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;
@@ -399,7 +415,6 @@ typedef struct {
        /* 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;
@@ -489,6 +504,9 @@ bi_make_temp_reg(bi_context *ctx)
 #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)
 
@@ -510,6 +528,12 @@ bi_make_temp_reg(bi_context *ctx)
 #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)
@@ -582,6 +606,12 @@ void bi_liveness_ins_update(uint16_t *live, bi_instruction *ins, unsigned max);
 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);