doc: add Meson coding style to contributors guide
[dpdk.git] / lib / librte_pipeline / rte_swx_pipeline.c
index 3dd5784..a2732a1 100644 (file)
@@ -12,6 +12,8 @@
 #include <rte_common.h>
 #include <rte_prefetch.h>
 #include <rte_byteorder.h>
+#include <rte_cycles.h>
+#include <rte_meter.h>
 
 #include "rte_swx_pipeline.h"
 #include "rte_swx_ctl.h"
@@ -279,8 +281,11 @@ enum instruction_type {
        /* rx m.port_in */
        INSTR_RX,
 
-       /* tx m.port_out */
-       INSTR_TX,
+       /* tx port_out
+        * port_out = MI
+        */
+       INSTR_TX,   /* port_out = M */
+       INSTR_TX_I, /* port_out = I */
 
        /* extract h.header */
        INSTR_HDR_EXTRACT,
@@ -313,9 +318,11 @@ enum instruction_type {
         * dst = src
         * dst = HMEF, src = HMEFTI
         */
-       INSTR_MOV,   /* dst = MEF, src = MEFT */
-       INSTR_MOV_S, /* (dst, src) = (MEF, H) or (dst, src) = (H, MEFT) */
-       INSTR_MOV_I, /* dst = HMEF, src = I */
+       INSTR_MOV,    /* dst = MEF, src = MEFT */
+       INSTR_MOV_MH, /* dst = MEF, src = H */
+       INSTR_MOV_HM, /* dst = H, src = MEFT */
+       INSTR_MOV_HH, /* dst = H, src = H */
+       INSTR_MOV_I,  /* dst = HMEF, src = I */
 
        /* dma h.header t.field
         * memcpy(h.header, t.field, sizeof(h.header))
@@ -369,25 +376,31 @@ enum instruction_type {
         * dst &= src
         * dst = HMEF, src = HMEFTI
         */
-       INSTR_ALU_AND,   /* dst = MEF, src = MEFT */
-       INSTR_ALU_AND_S, /* (dst, src) = (MEF, H) or (dst, src) = (H, MEFT) */
-       INSTR_ALU_AND_I, /* dst = HMEF, src = I */
+       INSTR_ALU_AND,    /* dst = MEF, src = MEFT */
+       INSTR_ALU_AND_MH, /* dst = MEF, src = H */
+       INSTR_ALU_AND_HM, /* dst = H, src = MEFT */
+       INSTR_ALU_AND_HH, /* dst = H, src = H */
+       INSTR_ALU_AND_I,  /* dst = HMEF, src = I */
 
        /* or dst src
         * dst |= src
         * dst = HMEF, src = HMEFTI
         */
-       INSTR_ALU_OR,   /* dst = MEF, src = MEFT */
-       INSTR_ALU_OR_S, /* (dst, src) = (MEF, H) or (dst, src) = (H, MEFT) */
-       INSTR_ALU_OR_I, /* dst = HMEF, src = I */
+       INSTR_ALU_OR,    /* dst = MEF, src = MEFT */
+       INSTR_ALU_OR_MH, /* dst = MEF, src = H */
+       INSTR_ALU_OR_HM, /* dst = H, src = MEFT */
+       INSTR_ALU_OR_HH, /* dst = H, src = H */
+       INSTR_ALU_OR_I,  /* dst = HMEF, src = I */
 
        /* xor dst src
         * dst ^= src
         * dst = HMEF, src = HMEFTI
         */
-       INSTR_ALU_XOR,   /* dst = MEF, src = MEFT */
-       INSTR_ALU_XOR_S, /* (dst, src) = (MEF, H) or (dst, src) = (H, MEFT) */
-       INSTR_ALU_XOR_I, /* dst = HMEF, src = I */
+       INSTR_ALU_XOR,    /* dst = MEF, src = MEFT */
+       INSTR_ALU_XOR_MH, /* dst = MEF, src = H */
+       INSTR_ALU_XOR_HM, /* dst = H, src = MEFT */
+       INSTR_ALU_XOR_HH, /* dst = H, src = H */
+       INSTR_ALU_XOR_I,  /* dst = HMEF, src = I */
 
        /* shl dst src
         * dst <<= src
@@ -458,6 +471,31 @@ enum instruction_type {
        INSTR_REGADD_RIM, /* index = I, src = MEFT */
        INSTR_REGADD_RII, /* index = I, src = I */
 
+       /* metprefetch METARRAY index
+        * prefetch METARRAY[index]
+        * index = HMEFTI
+        */
+       INSTR_METPREFETCH_H, /* index = H */
+       INSTR_METPREFETCH_M, /* index = MEFT */
+       INSTR_METPREFETCH_I, /* index = I */
+
+       /* meter METARRAY index length color_in color_out
+        * color_out = meter(METARRAY[index], length, color_in)
+        * index = HMEFTI, length = HMEFT, color_in = MEFTI, color_out = MEF
+        */
+       INSTR_METER_HHM, /* index = H, length = H, color_in = MEFT */
+       INSTR_METER_HHI, /* index = H, length = H, color_in = I */
+       INSTR_METER_HMM, /* index = H, length = MEFT, color_in = MEFT */
+       INSTR_METER_HMI, /* index = H, length = MEFT, color_in = I */
+       INSTR_METER_MHM, /* index = MEFT, length = H, color_in = MEFT */
+       INSTR_METER_MHI, /* index = MEFT, length = H, color_in = I */
+       INSTR_METER_MMM, /* index = MEFT, length = MEFT, color_in = MEFT */
+       INSTR_METER_MMI, /* index = MEFT, length = MEFT, color_in = I */
+       INSTR_METER_IHM, /* index = I, length = H, color_in = MEFT */
+       INSTR_METER_IHI, /* index = I, length = H, color_in = I */
+       INSTR_METER_IMM, /* index = I, length = MEFT, color_in = MEFT */
+       INSTR_METER_IMI, /* index = I, length = MEFT, color_in = I */
+
        /* table TABLE */
        INSTR_TABLE,
 
@@ -503,41 +541,45 @@ enum instruction_type {
        INSTR_JMP_ACTION_MISS,
 
        /* jmpeq LABEL a b
-        * Jump is a is equal to b
+        * Jump if a is equal to b
         * a = HMEFT, b = HMEFTI
         */
-       INSTR_JMP_EQ,   /* (a, b) = (MEFT, MEFT) or (a, b) = (H, H) */
-       INSTR_JMP_EQ_S, /* (a, b) = (MEFT, H) or (a, b) = (H, MEFT) */
-       INSTR_JMP_EQ_I, /* (a, b) = (MEFT, I) or (a, b) = (H, I) */
+       INSTR_JMP_EQ,    /* a = MEFT, b = MEFT */
+       INSTR_JMP_EQ_MH, /* a = MEFT, b = H */
+       INSTR_JMP_EQ_HM, /* a = H, b = MEFT */
+       INSTR_JMP_EQ_HH, /* a = H, b = H */
+       INSTR_JMP_EQ_I,  /* (a, b) = (MEFT, I) or (a, b) = (H, I) */
 
        /* jmpneq LABEL a b
-        * Jump is a is not equal to b
+        * Jump if a is not equal to b
         * a = HMEFT, b = HMEFTI
         */
-       INSTR_JMP_NEQ,   /* (a, b) = (MEFT, MEFT) or (a, b) = (H, H) */
-       INSTR_JMP_NEQ_S, /* (a, b) = (MEFT, H) or (a, b) = (H, MEFT) */
-       INSTR_JMP_NEQ_I, /* (a, b) = (MEFT, I) or (a, b) = (H, I) */
+       INSTR_JMP_NEQ,    /* a = MEFT, b = MEFT */
+       INSTR_JMP_NEQ_MH, /* a = MEFT, b = H */
+       INSTR_JMP_NEQ_HM, /* a = H, b = MEFT */
+       INSTR_JMP_NEQ_HH, /* a = H, b = H */
+       INSTR_JMP_NEQ_I,  /* (a, b) = (MEFT, I) or (a, b) = (H, I) */
 
        /* jmplt LABEL a b
         * Jump if a is less than b
         * a = HMEFT, b = HMEFTI
         */
-       INSTR_JMP_LT,    /* a = MEF, b = MEF */
-       INSTR_JMP_LT_MH, /* a = MEF, b = H */
-       INSTR_JMP_LT_HM, /* a = H, b = MEF */
+       INSTR_JMP_LT,    /* a = MEFT, b = MEFT */
+       INSTR_JMP_LT_MH, /* a = MEFT, b = H */
+       INSTR_JMP_LT_HM, /* a = H, b = MEFT */
        INSTR_JMP_LT_HH, /* a = H, b = H */
-       INSTR_JMP_LT_MI, /* a = MEF, b = I */
+       INSTR_JMP_LT_MI, /* a = MEFT, b = I */
        INSTR_JMP_LT_HI, /* a = H, b = I */
 
        /* jmpgt LABEL a b
         * Jump if a is greater than b
         * a = HMEFT, b = HMEFTI
         */
-       INSTR_JMP_GT,    /* a = MEF, b = MEF */
-       INSTR_JMP_GT_MH, /* a = MEF, b = H */
-       INSTR_JMP_GT_HM, /* a = H, b = MEF */
+       INSTR_JMP_GT,    /* a = MEFT, b = MEFT */
+       INSTR_JMP_GT_MH, /* a = MEFT, b = H */
+       INSTR_JMP_GT_HM, /* a = H, b = MEFT */
        INSTR_JMP_GT_HH, /* a = H, b = H */
-       INSTR_JMP_GT_MI, /* a = MEF, b = I */
+       INSTR_JMP_GT_MI, /* a = MEFT, b = I */
        INSTR_JMP_GT_HI, /* a = H, b = I */
 
        /* return
@@ -555,9 +597,15 @@ struct instr_operand {
 
 struct instr_io {
        struct {
-               uint8_t offset;
-               uint8_t n_bits;
-               uint8_t pad[2];
+               union {
+                       struct {
+                               uint8_t offset;
+                               uint8_t n_bits;
+                               uint8_t pad[2];
+                       };
+
+                       uint32_t val;
+               };
        } io;
 
        struct {
@@ -607,6 +655,25 @@ struct instr_regarray {
        };
 };
 
+struct instr_meter {
+       uint8_t metarray_id;
+       uint8_t pad[3];
+
+       union {
+               struct instr_operand idx;
+               uint32_t idx_val;
+       };
+
+       struct instr_operand length;
+
+       union {
+               struct instr_operand color_in;
+               uint32_t color_in_val;
+       };
+
+       struct instr_operand color_out;
+};
+
 struct instr_dma {
        struct {
                uint8_t header_id[8];
@@ -642,6 +709,7 @@ struct instruction {
                struct instr_hdr_validity valid;
                struct instr_dst_src mov;
                struct instr_regarray regarray;
+               struct instr_meter meter;
                struct instr_dma dma;
                struct instr_dst_src alu;
                struct instr_table table;
@@ -665,6 +733,7 @@ struct action {
        TAILQ_ENTRY(action) node;
        char name[RTE_SWX_NAME_SIZE];
        struct struct_type *st;
+       int *args_endianness; /* 0 = Host Byte Order (HBO). */
        struct instruction *instructions;
        uint32_t n_instructions;
        uint32_t id;
@@ -698,7 +767,6 @@ struct table {
        /* Match. */
        struct match_field *fields;
        uint32_t n_fields;
-       int is_header; /* Only valid when n_fields > 0. */
        struct header *header; /* Only valid when n_fields > 0. */
 
        /* Action. */
@@ -721,6 +789,11 @@ struct table_runtime {
        uint8_t **key;
 };
 
+struct table_statistics {
+       uint64_t n_pkts_hit[2]; /* 0 = Miss, 1 = Hit. */
+       uint64_t *n_pkts_action;
+};
+
 /*
  * Register array.
  */
@@ -739,6 +812,43 @@ struct regarray_runtime {
        uint32_t size_mask;
 };
 
+/*
+ * Meter array.
+ */
+struct meter_profile {
+       TAILQ_ENTRY(meter_profile) node;
+       char name[RTE_SWX_NAME_SIZE];
+       struct rte_meter_trtcm_params params;
+       struct rte_meter_trtcm_profile profile;
+       uint32_t n_users;
+};
+
+TAILQ_HEAD(meter_profile_tailq, meter_profile);
+
+struct metarray {
+       TAILQ_ENTRY(metarray) node;
+       char name[RTE_SWX_NAME_SIZE];
+       uint32_t size;
+       uint32_t id;
+};
+
+TAILQ_HEAD(metarray_tailq, metarray);
+
+struct meter {
+       struct rte_meter_trtcm m;
+       struct meter_profile *profile;
+       enum rte_color color_mask;
+       uint8_t pad[20];
+
+       uint64_t n_pkts[RTE_COLORS];
+       uint64_t n_bytes[RTE_COLORS];
+};
+
+struct metarray_runtime {
+       struct meter *metarray;
+       uint32_t size_mask;
+};
+
 /*
  * Pipeline.
  */
@@ -804,7 +914,7 @@ struct thread {
 
 #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
 
-#define ALU_S(thread, ip, operator)  \
+#define ALU_MH(thread, ip, operator)  \
 {                                                                              \
        uint8_t *dst_struct = (thread)->structs[(ip)->alu.dst.struct_id];      \
        uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->alu.dst.offset];   \
@@ -822,8 +932,6 @@ struct thread {
        *dst64_ptr = (dst64 & ~dst64_mask) | (result & dst64_mask);            \
 }
 
-#define ALU_MH ALU_S
-
 #define ALU_HM(thread, ip, operator)  \
 {                                                                              \
        uint8_t *dst_struct = (thread)->structs[(ip)->alu.dst.struct_id];      \
@@ -844,6 +952,25 @@ struct thread {
        *dst64_ptr = (dst64 & ~dst64_mask) | result;                           \
 }
 
+#define ALU_HM_FAST(thread, ip, operator)  \
+{                                                                                 \
+       uint8_t *dst_struct = (thread)->structs[(ip)->alu.dst.struct_id];         \
+       uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->alu.dst.offset];      \
+       uint64_t dst64 = *dst64_ptr;                                              \
+       uint64_t dst64_mask = UINT64_MAX >> (64 - (ip)->alu.dst.n_bits);          \
+       uint64_t dst = dst64 & dst64_mask;                                        \
+                                                                                 \
+       uint8_t *src_struct = (thread)->structs[(ip)->alu.src.struct_id];         \
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[(ip)->alu.src.offset];      \
+       uint64_t src64 = *src64_ptr;                                              \
+       uint64_t src64_mask = UINT64_MAX >> (64 - (ip)->alu.src.n_bits);          \
+       uint64_t src = hton64(src64 & src64_mask) >> (64 - (ip)->alu.dst.n_bits); \
+                                                                                 \
+       uint64_t result = dst operator src;                                       \
+                                                                                 \
+       *dst64_ptr = (dst64 & ~dst64_mask) | result;                              \
+}
+
 #define ALU_HH(thread, ip, operator)  \
 {                                                                              \
        uint8_t *dst_struct = (thread)->structs[(ip)->alu.dst.struct_id];      \
@@ -863,12 +990,31 @@ struct thread {
        *dst64_ptr = (dst64 & ~dst64_mask) | result;                           \
 }
 
+#define ALU_HH_FAST(thread, ip, operator)  \
+{                                                                                             \
+       uint8_t *dst_struct = (thread)->structs[(ip)->alu.dst.struct_id];                     \
+       uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->alu.dst.offset];                  \
+       uint64_t dst64 = *dst64_ptr;                                                          \
+       uint64_t dst64_mask = UINT64_MAX >> (64 - (ip)->alu.dst.n_bits);                      \
+       uint64_t dst = dst64 & dst64_mask;                                                    \
+                                                                                             \
+       uint8_t *src_struct = (thread)->structs[(ip)->alu.src.struct_id];                     \
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[(ip)->alu.src.offset];                  \
+       uint64_t src64 = *src64_ptr;                                                          \
+       uint64_t src = (src64 << (64 - (ip)->alu.src.n_bits)) >> (64 - (ip)->alu.dst.n_bits); \
+                                                                                             \
+       uint64_t result = dst operator src;                                                   \
+                                                                                             \
+       *dst64_ptr = (dst64 & ~dst64_mask) | result;                                          \
+}
+
 #else
 
-#define ALU_S ALU
 #define ALU_MH ALU
 #define ALU_HM ALU
+#define ALU_HM_FAST ALU
 #define ALU_HH ALU
+#define ALU_HH_FAST ALU
 
 #endif
 
@@ -931,7 +1077,7 @@ struct thread {
 
 #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
 
-#define MOV_S(thread, ip)  \
+#define MOV_MH(thread, ip)  \
 {                                                                              \
        uint8_t *dst_struct = (thread)->structs[(ip)->mov.dst.struct_id];      \
        uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->mov.dst.offset];   \
@@ -946,9 +1092,44 @@ struct thread {
        *dst64_ptr = (dst64 & ~dst64_mask) | (src & dst64_mask);               \
 }
 
+#define MOV_HM(thread, ip)  \
+{                                                                              \
+       uint8_t *dst_struct = (thread)->structs[(ip)->mov.dst.struct_id];      \
+       uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->mov.dst.offset];   \
+       uint64_t dst64 = *dst64_ptr;                                           \
+       uint64_t dst64_mask = UINT64_MAX >> (64 - (ip)->mov.dst.n_bits);       \
+                                                                              \
+       uint8_t *src_struct = (thread)->structs[(ip)->mov.src.struct_id];      \
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[(ip)->mov.src.offset];   \
+       uint64_t src64 = *src64_ptr;                                           \
+       uint64_t src64_mask = UINT64_MAX >> (64 - (ip)->mov.src.n_bits);       \
+       uint64_t src = src64 & src64_mask;                                     \
+                                                                              \
+       src = hton64(src) >> (64 - (ip)->mov.dst.n_bits);                      \
+       *dst64_ptr = (dst64 & ~dst64_mask) | src;                              \
+}
+
+#define MOV_HH(thread, ip)  \
+{                                                                              \
+       uint8_t *dst_struct = (thread)->structs[(ip)->mov.dst.struct_id];      \
+       uint64_t *dst64_ptr = (uint64_t *)&dst_struct[(ip)->mov.dst.offset];   \
+       uint64_t dst64 = *dst64_ptr;                                           \
+       uint64_t dst64_mask = UINT64_MAX >> (64 - (ip)->mov.dst.n_bits);       \
+                                                                              \
+       uint8_t *src_struct = (thread)->structs[(ip)->mov.src.struct_id];      \
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[(ip)->mov.src.offset];   \
+       uint64_t src64 = *src64_ptr;                                           \
+                                                                              \
+       uint64_t src = src64 << (64 - (ip)->mov.src.n_bits);                   \
+       src = src >> (64 - (ip)->mov.dst.n_bits);                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | src;                              \
+}
+
 #else
 
-#define MOV_S MOV
+#define MOV_MH MOV
+#define MOV_HM MOV
+#define MOV_HH MOV
 
 #endif
 
@@ -983,7 +1164,7 @@ struct thread {
 
 #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
 
-#define JMP_CMP_S(thread, ip, operator)  \
+#define JMP_CMP_MH(thread, ip, operator)  \
 {                                                                              \
        uint8_t *a_struct = (thread)->structs[(ip)->jmp.a.struct_id];          \
        uint64_t *a64_ptr = (uint64_t *)&a_struct[(ip)->jmp.a.offset];         \
@@ -999,8 +1180,6 @@ struct thread {
        (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
 }
 
-#define JMP_CMP_MH JMP_CMP_S
-
 #define JMP_CMP_HM(thread, ip, operator)  \
 {                                                                              \
        uint8_t *a_struct = (thread)->structs[(ip)->jmp.a.struct_id];          \
@@ -1032,12 +1211,27 @@ struct thread {
        (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
 }
 
+#define JMP_CMP_HH_FAST(thread, ip, operator)  \
+{                                                                              \
+       uint8_t *a_struct = (thread)->structs[(ip)->jmp.a.struct_id];          \
+       uint64_t *a64_ptr = (uint64_t *)&a_struct[(ip)->jmp.a.offset];         \
+       uint64_t a64 = *a64_ptr;                                               \
+       uint64_t a = a64 << (64 - (ip)->jmp.a.n_bits);                         \
+                                                                              \
+       uint8_t *b_struct = (thread)->structs[(ip)->jmp.b.struct_id];          \
+       uint64_t *b64_ptr = (uint64_t *)&b_struct[(ip)->jmp.b.offset];         \
+       uint64_t b64 = *b64_ptr;                                               \
+       uint64_t b = b64 << (64 - (ip)->jmp.b.n_bits);                         \
+                                                                              \
+       (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
+}
+
 #else
 
-#define JMP_CMP_S JMP_CMP
 #define JMP_CMP_MH JMP_CMP
 #define JMP_CMP_HM JMP_CMP
 #define JMP_CMP_HH JMP_CMP
+#define JMP_CMP_HH_FAST JMP_CMP
 
 #endif
 
@@ -1115,12 +1309,16 @@ struct rte_swx_pipeline {
        struct table_type_tailq table_types;
        struct table_tailq tables;
        struct regarray_tailq regarrays;
+       struct meter_profile_tailq meter_profiles;
+       struct metarray_tailq metarrays;
 
        struct port_in_runtime *in;
        struct port_out_runtime *out;
        struct instruction **action_instructions;
        struct rte_swx_table_state *table_state;
+       struct table_statistics *table_stats;
        struct regarray_runtime *regarray_runtime;
+       struct metarray_runtime *metarray_runtime;
        struct instruction *instructions;
        struct thread threads[RTE_SWX_PIPELINE_THREADS_MAX];
 
@@ -1132,6 +1330,7 @@ struct rte_swx_pipeline {
        uint32_t n_actions;
        uint32_t n_tables;
        uint32_t n_regarrays;
+       uint32_t n_metarrays;
        uint32_t n_headers;
        uint32_t thread_id;
        uint32_t port_id;
@@ -2149,6 +2348,18 @@ header_find(struct rte_swx_pipeline *p, const char *name)
        return NULL;
 }
 
+static struct header *
+header_find_by_struct_id(struct rte_swx_pipeline *p, uint32_t struct_id)
+{
+       struct header *elem;
+
+       TAILQ_FOREACH(elem, &p->headers, node)
+               if (elem->struct_id == struct_id)
+                       return elem;
+
+       return NULL;
+}
+
 static struct header *
 header_parse(struct rte_swx_pipeline *p,
             const char *name)
@@ -2402,6 +2613,19 @@ metadata_free(struct rte_swx_pipeline *p)
 /*
  * Instruction.
  */
+static int
+instruction_is_tx(enum instruction_type type)
+{
+       switch (type) {
+       case INSTR_TX:
+       case INSTR_TX_I:
+               return 1;
+
+       default:
+               return 0;
+       }
+}
+
 static int
 instruction_is_jmp(struct instruction *instr)
 {
@@ -2414,10 +2638,14 @@ instruction_is_jmp(struct instruction *instr)
        case INSTR_JMP_ACTION_HIT:
        case INSTR_JMP_ACTION_MISS:
        case INSTR_JMP_EQ:
-       case INSTR_JMP_EQ_S:
+       case INSTR_JMP_EQ_MH:
+       case INSTR_JMP_EQ_HM:
+       case INSTR_JMP_EQ_HH:
        case INSTR_JMP_EQ_I:
        case INSTR_JMP_NEQ:
-       case INSTR_JMP_NEQ_S:
+       case INSTR_JMP_NEQ_MH:
+       case INSTR_JMP_NEQ_HM:
+       case INSTR_JMP_NEQ_HH:
        case INSTR_JMP_NEQ_I:
        case INSTR_JMP_LT:
        case INSTR_JMP_LT_MH:
@@ -2644,16 +2872,42 @@ instr_tx_translate(struct rte_swx_pipeline *p,
                   struct instruction *instr,
                   struct instruction_data *data __rte_unused)
 {
+       char *port = tokens[1];
        struct field *f;
+       uint32_t port_val;
 
        CHECK(n_tokens == 2, EINVAL);
 
-       f = metadata_field_parse(p, tokens[1]);
-       CHECK(f, EINVAL);
+       f = metadata_field_parse(p, port);
+       if (f) {
+               instr->type = INSTR_TX;
+               instr->io.io.offset = f->offset / 8;
+               instr->io.io.n_bits = f->n_bits;
+               return 0;
+       }
 
-       instr->type = INSTR_TX;
-       instr->io.io.offset = f->offset / 8;
-       instr->io.io.n_bits = f->n_bits;
+       /* TX_I. */
+       port_val = strtoul(port, &port, 0);
+       CHECK(!port[0], EINVAL);
+
+       instr->type = INSTR_TX_I;
+       instr->io.io.val = port_val;
+       return 0;
+}
+
+static int
+instr_drop_translate(struct rte_swx_pipeline *p,
+                    struct action *action __rte_unused,
+                    char **tokens __rte_unused,
+                    int n_tokens,
+                    struct instruction *instr,
+                    struct instruction_data *data __rte_unused)
+{
+       CHECK(n_tokens == 1, EINVAL);
+
+       /* TX_I. */
+       instr->type = INSTR_TX_I;
+       instr->io.io.val = p->n_ports_out - 1;
        return 0;
 }
 
@@ -2741,6 +2995,30 @@ instr_tx_exec(struct rte_swx_pipeline *p)
        instr_rx_exec(p);
 }
 
+static inline void
+instr_tx_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint64_t port_id = ip->io.io.val;
+       struct port_out_runtime *port = &p->out[port_id];
+       struct rte_swx_pkt *pkt = &t->pkt;
+
+       TRACE("[Thread %2u]: tx (i) 1 pkt to port %u\n",
+             p->thread_id,
+             (uint32_t)port_id);
+
+       /* Headers. */
+       emit_handler(t);
+
+       /* Packet. */
+       port->pkt_tx(port->obj, pkt);
+
+       /* Thread. */
+       thread_ip_reset(p, t);
+       instr_rx_exec(p);
+}
+
 /*
  * extract.
  */
@@ -2936,10 +3214,11 @@ __instr_hdr_emit_exec(struct rte_swx_pipeline *p, uint32_t n_emit)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
+       uint64_t valid_headers = t->valid_headers;
        uint32_t n_headers_out = t->n_headers_out;
        struct header_out_runtime *ho = &t->headers_out[n_headers_out - 1];
        uint8_t *ho_ptr = NULL;
-       uint32_t ho_nbytes = 0, i;
+       uint32_t ho_nbytes = 0, first = 1, i;
 
        for (i = 0; i < n_emit; i++) {
                uint32_t header_id = ip->io.hdr.header_id[i];
@@ -2949,12 +3228,17 @@ __instr_hdr_emit_exec(struct rte_swx_pipeline *p, uint32_t n_emit)
                struct header_runtime *hi = &t->headers[header_id];
                uint8_t *hi_ptr = t->structs[struct_id];
 
+               if (!MASK64_BIT_GET(valid_headers, header_id))
+                       continue;
+
                TRACE("[Thread %2u]: emit header %u\n",
                      p->thread_id,
                      header_id);
 
                /* Headers. */
-               if (!i) {
+               if (first) {
+                       first = 0;
+
                        if (!t->n_headers_out) {
                                ho = &t->headers_out[0];
 
@@ -3195,7 +3479,8 @@ instr_table_exec(struct rte_swx_pipeline *p)
        uint32_t table_id = ip->table.table_id;
        struct rte_swx_table_state *ts = &t->table_state[table_id];
        struct table_runtime *table = &t->tables[table_id];
-       uint64_t action_id;
+       struct table_statistics *stats = &p->table_stats[table_id];
+       uint64_t action_id, n_pkts_hit, n_pkts_action;
        uint8_t *action_data;
        int done, hit;
 
@@ -3218,6 +3503,8 @@ instr_table_exec(struct rte_swx_pipeline *p)
 
        action_id = hit ? action_id : ts->default_action_id;
        action_data = hit ? action_data : ts->default_action_data;
+       n_pkts_hit = stats->n_pkts_hit[hit];
+       n_pkts_action = stats->n_pkts_action[action_id];
 
        TRACE("[Thread %2u] table %u (%s, action %u)\n",
              p->thread_id,
@@ -3228,6 +3515,8 @@ instr_table_exec(struct rte_swx_pipeline *p)
        t->action_id = action_id;
        t->structs[0] = action_data;
        t->hit = hit;
+       stats->n_pkts_hit[hit] = n_pkts_hit + 1;
+       stats->n_pkts_action[action_id] = n_pkts_action + 1;
 
        /* Thread. */
        thread_ip_action_call(p, t, action_id);
@@ -3335,20 +3624,23 @@ instr_mov_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
        fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
        CHECK(fdst, EINVAL);
 
-       /* MOV or MOV_S. */
+       /* MOV, MOV_MH, MOV_HM or MOV_HH. */
        fsrc = struct_field_parse(p, action, src, &src_struct_id);
        if (fsrc) {
                instr->type = INSTR_MOV;
-               if ((dst[0] == 'h' && src[0] != 'h') ||
-                   (dst[0] != 'h' && src[0] == 'h'))
-                       instr->type = INSTR_MOV_S;
+               if (dst[0] != 'h' && src[0] == 'h')
+                       instr->type = INSTR_MOV_MH;
+               if (dst[0] == 'h' && src[0] != 'h')
+                       instr->type = INSTR_MOV_HM;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_MOV_HH;
 
                instr->mov.dst.struct_id = (uint8_t)dst_struct_id;
                instr->mov.dst.n_bits = fdst->n_bits;
@@ -3390,70 +3682,69 @@ instr_mov_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_mov_s_exec(struct rte_swx_pipeline *p)
+instr_mov_mh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] mov (s)\n",
+       TRACE("[Thread %2u] mov (mh)\n",
              p->thread_id);
 
-       MOV_S(t, ip);
+       MOV_MH(t, ip);
 
        /* Thread. */
        thread_ip_inc(p);
 }
 
 static inline void
-instr_mov_i_exec(struct rte_swx_pipeline *p)
+instr_mov_hm_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] mov m.f %" PRIx64 "\n",
-             p->thread_id,
-             ip->mov.src_val);
+       TRACE("[Thread %2u] mov (hm)\n",
+             p->thread_id);
 
-       MOV_I(t, ip);
+       MOV_HM(t, ip);
 
        /* Thread. */
        thread_ip_inc(p);
 }
 
-/*
- * dma.
- */
-static int
-instr_dma_translate(struct rte_swx_pipeline *p,
-                   struct action *action,
-                   char **tokens,
-                   int n_tokens,
-                   struct instruction *instr,
-                   struct instruction_data *data __rte_unused)
+static inline void
+instr_mov_hh_exec(struct rte_swx_pipeline *p)
 {
-       char *dst = tokens[1];
-       char *src = tokens[2];
-       struct header *h;
-       struct field *tf;
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
 
-       CHECK(action, EINVAL);
-       CHECK(n_tokens == 3, EINVAL);
+       TRACE("[Thread %2u] mov (hh)\n",
+             p->thread_id);
 
-       h = header_parse(p, dst);
-       CHECK(h, EINVAL);
+       MOV_HH(t, ip);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_mov_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
 
-       tf = action_field_parse(action, src);
-       CHECK(tf, EINVAL);
+       TRACE("[Thread %2u] mov m.f %" PRIx64 "\n",
+             p->thread_id,
+             ip->mov.src_val);
 
-       instr->type = INSTR_DMA_HT;
-       instr->dma.dst.header_id[0] = h->id;
-       instr->dma.dst.struct_id[0] = h->struct_id;
-       instr->dma.n_bytes[0] = h->st->n_bits / 8;
-       instr->dma.src.offset[0] = tf->offset / 8;
+       MOV_I(t, ip);
 
-       return 0;
+       /* Thread. */
+       thread_ip_inc(p);
 }
 
+/*
+ * dma.
+ */
 static inline void
 __instr_dma_ht_exec(struct rte_swx_pipeline *p, uint32_t n_dma);
 
@@ -3598,7 +3889,7 @@ instr_alu_add_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
@@ -3651,7 +3942,7 @@ instr_alu_sub_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
@@ -3781,7 +4072,7 @@ instr_alu_shl_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
@@ -3834,7 +4125,7 @@ instr_alu_shr_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
@@ -3887,20 +4178,23 @@ instr_alu_and_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
        fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
        CHECK(fdst, EINVAL);
 
-       /* AND or AND_S. */
+       /* AND, AND_MH, AND_HM, AND_HH. */
        fsrc = struct_field_parse(p, action, src, &src_struct_id);
        if (fsrc) {
                instr->type = INSTR_ALU_AND;
-               if ((dst[0] == 'h' && src[0] != 'h') ||
-                   (dst[0] != 'h' && src[0] == 'h'))
-                       instr->type = INSTR_ALU_AND_S;
+               if (dst[0] != 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_AND_MH;
+               if (dst[0] == 'h' && src[0] != 'h')
+                       instr->type = INSTR_ALU_AND_HM;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_AND_HH;
 
                instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
                instr->alu.dst.n_bits = fdst->n_bits;
@@ -3937,20 +4231,23 @@ instr_alu_or_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
        fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
        CHECK(fdst, EINVAL);
 
-       /* OR or OR_S. */
+       /* OR, OR_MH, OR_HM, OR_HH. */
        fsrc = struct_field_parse(p, action, src, &src_struct_id);
        if (fsrc) {
                instr->type = INSTR_ALU_OR;
-               if ((dst[0] == 'h' && src[0] != 'h') ||
-                   (dst[0] != 'h' && src[0] == 'h'))
-                       instr->type = INSTR_ALU_OR_S;
+               if (dst[0] != 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_OR_MH;
+               if (dst[0] == 'h' && src[0] != 'h')
+                       instr->type = INSTR_ALU_OR_HM;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_OR_HH;
 
                instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
                instr->alu.dst.n_bits = fdst->n_bits;
@@ -3987,20 +4284,23 @@ instr_alu_xor_translate(struct rte_swx_pipeline *p,
        char *dst = tokens[1], *src = tokens[2];
        struct field *fdst, *fsrc;
        uint64_t src_val;
-       uint32_t dst_struct_id, src_struct_id;
+       uint32_t dst_struct_id = 0, src_struct_id = 0;
 
        CHECK(n_tokens == 3, EINVAL);
 
        fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
        CHECK(fdst, EINVAL);
 
-       /* XOR or XOR_S. */
+       /* XOR, XOR_MH, XOR_HM, XOR_HH. */
        fsrc = struct_field_parse(p, action, src, &src_struct_id);
        if (fsrc) {
                instr->type = INSTR_ALU_XOR;
-               if ((dst[0] == 'h' && src[0] != 'h') ||
-                   (dst[0] != 'h' && src[0] == 'h'))
-                       instr->type = INSTR_ALU_XOR_S;
+               if (dst[0] != 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_XOR_MH;
+               if (dst[0] == 'h' && src[0] != 'h')
+                       instr->type = INSTR_ALU_XOR_HM;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_XOR_HH;
 
                instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
                instr->alu.dst.n_bits = fdst->n_bits;
@@ -4402,15 +4702,45 @@ instr_alu_and_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_alu_and_s_exec(struct rte_swx_pipeline *p)
+instr_alu_and_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] and (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, &);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_and_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] and (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM_FAST(t, ip, &);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_and_hh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] and (s)\n", p->thread_id);
+       TRACE("[Thread %2u] and (hh)\n", p->thread_id);
 
        /* Structs. */
-       ALU_S(t, ip, &);
+       ALU_HH_FAST(t, ip, &);
 
        /* Thread. */
        thread_ip_inc(p);
@@ -4447,15 +4777,45 @@ instr_alu_or_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_alu_or_s_exec(struct rte_swx_pipeline *p)
+instr_alu_or_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] or (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, |);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_or_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] or (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM_FAST(t, ip, |);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_or_hh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] or (s)\n", p->thread_id);
+       TRACE("[Thread %2u] or (hh)\n", p->thread_id);
 
        /* Structs. */
-       ALU_S(t, ip, |);
+       ALU_HH_FAST(t, ip, |);
 
        /* Thread. */
        thread_ip_inc(p);
@@ -4492,50 +4852,80 @@ instr_alu_xor_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_alu_xor_s_exec(struct rte_swx_pipeline *p)
+instr_alu_xor_mh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] xor (s)\n", p->thread_id);
+       TRACE("[Thread %2u] xor (mh)\n", p->thread_id);
 
        /* Structs. */
-       ALU_S(t, ip, ^);
+       ALU_MH(t, ip, ^);
 
        /* Thread. */
        thread_ip_inc(p);
 }
 
 static inline void
-instr_alu_xor_i_exec(struct rte_swx_pipeline *p)
+instr_alu_xor_hm_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] xor (i)\n", p->thread_id);
+       TRACE("[Thread %2u] xor (hm)\n", p->thread_id);
 
        /* Structs. */
-       ALU_I(t, ip, ^);
+       ALU_HM_FAST(t, ip, ^);
 
        /* Thread. */
        thread_ip_inc(p);
 }
 
 static inline void
-instr_alu_ckadd_field_exec(struct rte_swx_pipeline *p)
+instr_alu_xor_hh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
-       uint8_t *dst_struct, *src_struct;
-       uint16_t *dst16_ptr, dst;
-       uint64_t *src64_ptr, src64, src64_mask, src;
-       uint64_t r;
 
-       TRACE("[Thread %2u] ckadd (field)\n", p->thread_id);
+       TRACE("[Thread %2u] xor (hh)\n", p->thread_id);
 
        /* Structs. */
-       dst_struct = t->structs[ip->alu.dst.struct_id];
-       dst16_ptr = (uint16_t *)&dst_struct[ip->alu.dst.offset];
+       ALU_HH_FAST(t, ip, ^);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_xor_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] xor (i)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_I(t, ip, ^);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_ckadd_field_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint8_t *dst_struct, *src_struct;
+       uint16_t *dst16_ptr, dst;
+       uint64_t *src64_ptr, src64, src64_mask, src;
+       uint64_t r;
+
+       TRACE("[Thread %2u] ckadd (field)\n", p->thread_id);
+
+       /* Structs. */
+       dst_struct = t->structs[ip->alu.dst.struct_id];
+       dst16_ptr = (uint16_t *)&dst_struct[ip->alu.dst.offset];
        dst = *dst16_ptr;
 
        src_struct = t->structs[ip->alu.src.struct_id];
@@ -5658,253 +6048,1082 @@ instr_regadd_rii_exec(struct rte_swx_pipeline *p)
 }
 
 /*
- * jmp.
+ * metarray.
  */
-static struct action *
-action_find(struct rte_swx_pipeline *p, const char *name);
+static struct metarray *
+metarray_find(struct rte_swx_pipeline *p, const char *name);
 
 static int
-instr_jmp_translate(struct rte_swx_pipeline *p __rte_unused,
-                   struct action *action __rte_unused,
-                   char **tokens,
-                   int n_tokens,
-                   struct instruction *instr,
-                   struct instruction_data *data)
+instr_metprefetch_translate(struct rte_swx_pipeline *p,
+                           struct action *action,
+                           char **tokens,
+                           int n_tokens,
+                           struct instruction *instr,
+                           struct instruction_data *data __rte_unused)
 {
-       CHECK(n_tokens == 2, EINVAL);
-
-       strcpy(data->jmp_label, tokens[1]);
+       char *metarray = tokens[1], *idx = tokens[2];
+       struct metarray *m;
+       struct field *fidx;
+       uint32_t idx_struct_id, idx_val;
 
-       instr->type = INSTR_JMP;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       return 0;
-}
+       CHECK(n_tokens == 3, EINVAL);
 
-static int
-instr_jmp_valid_translate(struct rte_swx_pipeline *p,
-                         struct action *action __rte_unused,
-                         char **tokens,
-                         int n_tokens,
-                         struct instruction *instr,
-                         struct instruction_data *data)
-{
-       struct header *h;
+       m = metarray_find(p, metarray);
+       CHECK(m, EINVAL);
 
-       CHECK(n_tokens == 3, EINVAL);
+       /* METPREFETCH_H, METPREFETCH_M. */
+       fidx = struct_field_parse(p, action, idx, &idx_struct_id);
+       if (fidx) {
+               instr->type = INSTR_METPREFETCH_M;
+               if (idx[0] == 'h')
+                       instr->type = INSTR_METPREFETCH_H;
 
-       strcpy(data->jmp_label, tokens[1]);
+               instr->meter.metarray_id = m->id;
+               instr->meter.idx.struct_id = (uint8_t)idx_struct_id;
+               instr->meter.idx.n_bits = fidx->n_bits;
+               instr->meter.idx.offset = fidx->offset / 8;
+               return 0;
+       }
 
-       h = header_parse(p, tokens[2]);
-       CHECK(h, EINVAL);
+       /* METPREFETCH_I. */
+       idx_val = strtoul(idx, &idx, 0);
+       CHECK(!idx[0], EINVAL);
 
-       instr->type = INSTR_JMP_VALID;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       instr->jmp.header_id = h->id;
+       instr->type = INSTR_METPREFETCH_I;
+       instr->meter.metarray_id = m->id;
+       instr->meter.idx_val = idx_val;
        return 0;
 }
 
 static int
-instr_jmp_invalid_translate(struct rte_swx_pipeline *p,
-                           struct action *action __rte_unused,
-                           char **tokens,
-                           int n_tokens,
-                           struct instruction *instr,
-                           struct instruction_data *data)
+instr_meter_translate(struct rte_swx_pipeline *p,
+                     struct action *action,
+                     char **tokens,
+                     int n_tokens,
+                     struct instruction *instr,
+                     struct instruction_data *data __rte_unused)
 {
-       struct header *h;
+       char *metarray = tokens[1], *idx = tokens[2], *length = tokens[3];
+       char *color_in = tokens[4], *color_out = tokens[5];
+       struct metarray *m;
+       struct field *fidx, *flength, *fcin, *fcout;
+       uint32_t idx_struct_id, length_struct_id;
+       uint32_t color_in_struct_id, color_out_struct_id;
 
-       CHECK(n_tokens == 3, EINVAL);
+       CHECK(n_tokens == 6, EINVAL);
 
-       strcpy(data->jmp_label, tokens[1]);
+       m = metarray_find(p, metarray);
+       CHECK(m, EINVAL);
 
-       h = header_parse(p, tokens[2]);
-       CHECK(h, EINVAL);
+       fidx = struct_field_parse(p, action, idx, &idx_struct_id);
 
-       instr->type = INSTR_JMP_INVALID;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       instr->jmp.header_id = h->id;
-       return 0;
-}
+       flength = struct_field_parse(p, action, length, &length_struct_id);
+       CHECK(flength, EINVAL);
 
-static int
-instr_jmp_hit_translate(struct rte_swx_pipeline *p __rte_unused,
-                       struct action *action,
-                       char **tokens,
-                       int n_tokens,
-                       struct instruction *instr,
-                       struct instruction_data *data)
-{
-       CHECK(!action, EINVAL);
-       CHECK(n_tokens == 2, EINVAL);
+       fcin = struct_field_parse(p, action, color_in, &color_in_struct_id);
 
-       strcpy(data->jmp_label, tokens[1]);
+       fcout = struct_field_parse(p, NULL, color_out, &color_out_struct_id);
+       CHECK(fcout, EINVAL);
 
-       instr->type = INSTR_JMP_HIT;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       return 0;
-}
+       /* index = HMEFT, length = HMEFT, color_in = MEFT, color_out = MEF. */
+       if (fidx && fcin) {
+               instr->type = INSTR_METER_MMM;
+               if (idx[0] == 'h' && length[0] == 'h')
+                       instr->type = INSTR_METER_HHM;
+               if (idx[0] == 'h' && length[0] != 'h')
+                       instr->type = INSTR_METER_HMM;
+               if (idx[0] != 'h' && length[0] == 'h')
+                       instr->type = INSTR_METER_MHM;
 
-static int
-instr_jmp_miss_translate(struct rte_swx_pipeline *p __rte_unused,
-                        struct action *action,
-                        char **tokens,
-                        int n_tokens,
-                        struct instruction *instr,
-                        struct instruction_data *data)
-{
-       CHECK(!action, EINVAL);
-       CHECK(n_tokens == 2, EINVAL);
+               instr->meter.metarray_id = m->id;
 
-       strcpy(data->jmp_label, tokens[1]);
+               instr->meter.idx.struct_id = (uint8_t)idx_struct_id;
+               instr->meter.idx.n_bits = fidx->n_bits;
+               instr->meter.idx.offset = fidx->offset / 8;
 
-       instr->type = INSTR_JMP_MISS;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       return 0;
-}
+               instr->meter.length.struct_id = (uint8_t)length_struct_id;
+               instr->meter.length.n_bits = flength->n_bits;
+               instr->meter.length.offset = flength->offset / 8;
 
-static int
-instr_jmp_action_hit_translate(struct rte_swx_pipeline *p,
-                              struct action *action,
-                              char **tokens,
-                              int n_tokens,
-                              struct instruction *instr,
-                              struct instruction_data *data)
-{
-       struct action *a;
+               instr->meter.color_in.struct_id = (uint8_t)color_in_struct_id;
+               instr->meter.color_in.n_bits = fcin->n_bits;
+               instr->meter.color_in.offset = fcin->offset / 8;
 
-       CHECK(!action, EINVAL);
-       CHECK(n_tokens == 3, EINVAL);
+               instr->meter.color_out.struct_id = (uint8_t)color_out_struct_id;
+               instr->meter.color_out.n_bits = fcout->n_bits;
+               instr->meter.color_out.offset = fcout->offset / 8;
 
-       strcpy(data->jmp_label, tokens[1]);
+               return 0;
+       }
 
-       a = action_find(p, tokens[2]);
-       CHECK(a, EINVAL);
+       /* index = HMEFT, length = HMEFT, color_in = I, color_out = MEF. */
+       if (fidx && !fcin) {
+               uint32_t color_in_val = strtoul(color_in, &color_in, 0);
+               CHECK(!color_in[0], EINVAL);
 
-       instr->type = INSTR_JMP_ACTION_HIT;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       instr->jmp.action_id = a->id;
-       return 0;
-}
+               instr->type = INSTR_METER_MMI;
+               if (idx[0] == 'h' && length[0] == 'h')
+                       instr->type = INSTR_METER_HHI;
+               if (idx[0] == 'h' && length[0] != 'h')
+                       instr->type = INSTR_METER_HMI;
+               if (idx[0] != 'h' && length[0] == 'h')
+                       instr->type = INSTR_METER_MHI;
 
-static int
-instr_jmp_action_miss_translate(struct rte_swx_pipeline *p,
-                               struct action *action,
-                               char **tokens,
-                               int n_tokens,
-                               struct instruction *instr,
-                               struct instruction_data *data)
-{
-       struct action *a;
+               instr->meter.metarray_id = m->id;
 
-       CHECK(!action, EINVAL);
-       CHECK(n_tokens == 3, EINVAL);
+               instr->meter.idx.struct_id = (uint8_t)idx_struct_id;
+               instr->meter.idx.n_bits = fidx->n_bits;
+               instr->meter.idx.offset = fidx->offset / 8;
 
-       strcpy(data->jmp_label, tokens[1]);
+               instr->meter.length.struct_id = (uint8_t)length_struct_id;
+               instr->meter.length.n_bits = flength->n_bits;
+               instr->meter.length.offset = flength->offset / 8;
 
-       a = action_find(p, tokens[2]);
-       CHECK(a, EINVAL);
+               instr->meter.color_in_val = color_in_val;
 
-       instr->type = INSTR_JMP_ACTION_MISS;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       instr->jmp.action_id = a->id;
-       return 0;
-}
+               instr->meter.color_out.struct_id = (uint8_t)color_out_struct_id;
+               instr->meter.color_out.n_bits = fcout->n_bits;
+               instr->meter.color_out.offset = fcout->offset / 8;
 
-static int
-instr_jmp_eq_translate(struct rte_swx_pipeline *p,
-                      struct action *action,
-                      char **tokens,
-                      int n_tokens,
-                      struct instruction *instr,
-                      struct instruction_data *data)
-{
-       char *a = tokens[2], *b = tokens[3];
-       struct field *fa, *fb;
-       uint64_t b_val;
-       uint32_t a_struct_id, b_struct_id;
+               return 0;
+       }
 
-       CHECK(n_tokens == 4, EINVAL);
+       /* index = I, length = HMEFT, color_in = MEFT, color_out = MEF. */
+       if (!fidx && fcin) {
+               uint32_t idx_val;
 
-       strcpy(data->jmp_label, tokens[1]);
+               idx_val = strtoul(idx, &idx, 0);
+               CHECK(!idx[0], EINVAL);
 
-       fa = struct_field_parse(p, action, a, &a_struct_id);
-       CHECK(fa, EINVAL);
+               instr->type = INSTR_METER_IMM;
+               if (length[0] == 'h')
+                       instr->type = INSTR_METER_IHM;
 
-       /* JMP_EQ or JMP_EQ_S. */
-       fb = struct_field_parse(p, action, b, &b_struct_id);
-       if (fb) {
-               instr->type = INSTR_JMP_EQ;
-               if ((a[0] == 'h' && b[0] != 'h') ||
-                   (a[0] != 'h' && b[0] == 'h'))
-                       instr->type = INSTR_JMP_EQ_S;
-               instr->jmp.ip = NULL; /* Resolved later. */
+               instr->meter.metarray_id = m->id;
+
+               instr->meter.idx_val = idx_val;
+
+               instr->meter.length.struct_id = (uint8_t)length_struct_id;
+               instr->meter.length.n_bits = flength->n_bits;
+               instr->meter.length.offset = flength->offset / 8;
+
+               instr->meter.color_in.struct_id = (uint8_t)color_in_struct_id;
+               instr->meter.color_in.n_bits = fcin->n_bits;
+               instr->meter.color_in.offset = fcin->offset / 8;
+
+               instr->meter.color_out.struct_id = (uint8_t)color_out_struct_id;
+               instr->meter.color_out.n_bits = fcout->n_bits;
+               instr->meter.color_out.offset = fcout->offset / 8;
 
-               instr->jmp.a.struct_id = (uint8_t)a_struct_id;
-               instr->jmp.a.n_bits = fa->n_bits;
-               instr->jmp.a.offset = fa->offset / 8;
-               instr->jmp.b.struct_id = (uint8_t)b_struct_id;
-               instr->jmp.b.n_bits = fb->n_bits;
-               instr->jmp.b.offset = fb->offset / 8;
                return 0;
        }
 
-       /* JMP_EQ_I. */
-       b_val = strtoull(b, &b, 0);
-       CHECK(!b[0], EINVAL);
+       /* index = I, length = HMEFT, color_in = I, color_out = MEF. */
+       if (!fidx && !fcin) {
+               uint32_t idx_val, color_in_val;
 
-       if (a[0] == 'h')
-               b_val = hton64(b_val) >> (64 - fa->n_bits);
+               idx_val = strtoul(idx, &idx, 0);
+               CHECK(!idx[0], EINVAL);
 
-       instr->type = INSTR_JMP_EQ_I;
-       instr->jmp.ip = NULL; /* Resolved later. */
-       instr->jmp.a.struct_id = (uint8_t)a_struct_id;
-       instr->jmp.a.n_bits = fa->n_bits;
-       instr->jmp.a.offset = fa->offset / 8;
-       instr->jmp.b_val = b_val;
-       return 0;
-}
+               color_in_val = strtoul(color_in, &color_in, 0);
+               CHECK(!color_in[0], EINVAL);
 
-static int
-instr_jmp_neq_translate(struct rte_swx_pipeline *p,
-                       struct action *action,
-                       char **tokens,
-                       int n_tokens,
-                       struct instruction *instr,
-                       struct instruction_data *data)
-{
-       char *a = tokens[2], *b = tokens[3];
-       struct field *fa, *fb;
-       uint64_t b_val;
-       uint32_t a_struct_id, b_struct_id;
+               instr->type = INSTR_METER_IMI;
+               if (length[0] == 'h')
+                       instr->type = INSTR_METER_IHI;
 
-       CHECK(n_tokens == 4, EINVAL);
+               instr->meter.metarray_id = m->id;
 
-       strcpy(data->jmp_label, tokens[1]);
+               instr->meter.idx_val = idx_val;
 
-       fa = struct_field_parse(p, action, a, &a_struct_id);
-       CHECK(fa, EINVAL);
+               instr->meter.length.struct_id = (uint8_t)length_struct_id;
+               instr->meter.length.n_bits = flength->n_bits;
+               instr->meter.length.offset = flength->offset / 8;
 
-       /* JMP_NEQ or JMP_NEQ_S. */
-       fb = struct_field_parse(p, action, b, &b_struct_id);
-       if (fb) {
-               instr->type = INSTR_JMP_NEQ;
-               if ((a[0] == 'h' && b[0] != 'h') ||
-                   (a[0] != 'h' && b[0] == 'h'))
-                       instr->type = INSTR_JMP_NEQ_S;
-               instr->jmp.ip = NULL; /* Resolved later. */
+               instr->meter.color_in_val = color_in_val;
+
+               instr->meter.color_out.struct_id = (uint8_t)color_out_struct_id;
+               instr->meter.color_out.n_bits = fcout->n_bits;
+               instr->meter.color_out.offset = fcout->offset / 8;
 
-               instr->jmp.a.struct_id = (uint8_t)a_struct_id;
-               instr->jmp.a.n_bits = fa->n_bits;
-               instr->jmp.a.offset = fa->offset / 8;
-               instr->jmp.b.struct_id = (uint8_t)b_struct_id;
-               instr->jmp.b.n_bits = fb->n_bits;
-               instr->jmp.b.offset = fb->offset / 8;
                return 0;
        }
 
-       /* JMP_NEQ_I. */
-       b_val = strtoull(b, &b, 0);
-       CHECK(!b[0], EINVAL);
+       CHECK(0, EINVAL);
+}
+
+static inline struct meter *
+instr_meter_idx_hbo(struct rte_swx_pipeline *p, struct thread *t, struct instruction *ip)
+{
+       struct metarray_runtime *r = &p->metarray_runtime[ip->meter.metarray_id];
+
+       uint8_t *idx_struct = t->structs[ip->meter.idx.struct_id];
+       uint64_t *idx64_ptr = (uint64_t *)&idx_struct[ip->meter.idx.offset];
+       uint64_t idx64 = *idx64_ptr;
+       uint64_t idx64_mask = UINT64_MAX >> (64 - (ip)->meter.idx.n_bits);
+       uint64_t idx = idx64 & idx64_mask & r->size_mask;
+
+       return &r->metarray[idx];
+}
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+static inline struct meter *
+instr_meter_idx_nbo(struct rte_swx_pipeline *p, struct thread *t, struct instruction *ip)
+{
+       struct metarray_runtime *r = &p->metarray_runtime[ip->meter.metarray_id];
+
+       uint8_t *idx_struct = t->structs[ip->meter.idx.struct_id];
+       uint64_t *idx64_ptr = (uint64_t *)&idx_struct[ip->meter.idx.offset];
+       uint64_t idx64 = *idx64_ptr;
+       uint64_t idx = (ntoh64(idx64) >> (64 - ip->meter.idx.n_bits)) & r->size_mask;
+
+       return &r->metarray[idx];
+}
+
+#else
+
+#define instr_meter_idx_nbo instr_meter_idx_hbo
+
+#endif
+
+static inline struct meter *
+instr_meter_idx_imm(struct rte_swx_pipeline *p, struct instruction *ip)
+{
+       struct metarray_runtime *r = &p->metarray_runtime[ip->meter.metarray_id];
+
+       uint64_t idx =  ip->meter.idx_val & r->size_mask;
+
+       return &r->metarray[idx];
+}
+
+static inline uint32_t
+instr_meter_length_hbo(struct thread *t, struct instruction *ip)
+{
+       uint8_t *src_struct = t->structs[ip->meter.length.struct_id];
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[ip->meter.length.offset];
+       uint64_t src64 = *src64_ptr;
+       uint64_t src64_mask = UINT64_MAX >> (64 - (ip)->meter.length.n_bits);
+       uint64_t src = src64 & src64_mask;
+
+       return (uint32_t)src;
+}
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+static inline uint32_t
+instr_meter_length_nbo(struct thread *t, struct instruction *ip)
+{
+       uint8_t *src_struct = t->structs[ip->meter.length.struct_id];
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[ip->meter.length.offset];
+       uint64_t src64 = *src64_ptr;
+       uint64_t src = ntoh64(src64) >> (64 - ip->meter.length.n_bits);
+
+       return (uint32_t)src;
+}
+
+#else
+
+#define instr_meter_length_nbo instr_meter_length_hbo
+
+#endif
+
+static inline enum rte_color
+instr_meter_color_in_hbo(struct thread *t, struct instruction *ip)
+{
+       uint8_t *src_struct = t->structs[ip->meter.color_in.struct_id];
+       uint64_t *src64_ptr = (uint64_t *)&src_struct[ip->meter.color_in.offset];
+       uint64_t src64 = *src64_ptr;
+       uint64_t src64_mask = UINT64_MAX >> (64 - ip->meter.color_in.n_bits);
+       uint64_t src = src64 & src64_mask;
+
+       return (enum rte_color)src;
+}
+
+static inline void
+instr_meter_color_out_hbo_set(struct thread *t, struct instruction *ip, enum rte_color color_out)
+{
+       uint8_t *dst_struct = t->structs[ip->meter.color_out.struct_id];
+       uint64_t *dst64_ptr = (uint64_t *)&dst_struct[ip->meter.color_out.offset];
+       uint64_t dst64 = *dst64_ptr;
+       uint64_t dst64_mask = UINT64_MAX >> (64 - ip->meter.color_out.n_bits);
+
+       uint64_t src = (uint64_t)color_out;
+
+       *dst64_ptr = (dst64 & ~dst64_mask) | (src & dst64_mask);
+}
+
+static inline void
+instr_metprefetch_h_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+
+       TRACE("[Thread %2u] metprefetch (h)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_nbo(p, t, ip);
+       rte_prefetch0(m);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_metprefetch_m_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+
+       TRACE("[Thread %2u] metprefetch (m)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_hbo(p, t, ip);
+       rte_prefetch0(m);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_metprefetch_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+
+       TRACE("[Thread %2u] metprefetch (i)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_imm(p, ip);
+       rte_prefetch0(m);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_hhm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (hhm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_nbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_hhi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (hhi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_nbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_hmm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (hmm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_nbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+static inline void
+instr_meter_hmi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (hmi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_nbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_mhm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (mhm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_hbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_mhi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (mhi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_hbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_mmm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (mmm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_hbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_mmi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (mmi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_hbo(p, t, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_ihm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (ihm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_imm(p, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_ihi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (ihi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_imm(p, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_nbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_meter_imm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (imm)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_imm(p, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = instr_meter_color_in_hbo(t, ip);
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+static inline void
+instr_meter_imi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct meter *m;
+       uint64_t time, n_pkts, n_bytes;
+       uint32_t length;
+       enum rte_color color_in, color_out;
+
+       TRACE("[Thread %2u] meter (imi)\n", p->thread_id);
+
+       /* Structs. */
+       m = instr_meter_idx_imm(p, ip);
+       rte_prefetch0(m->n_pkts);
+       time = rte_get_tsc_cycles();
+       length = instr_meter_length_hbo(t, ip);
+       color_in = (enum rte_color)ip->meter.color_in_val;
+
+       color_out = rte_meter_trtcm_color_aware_check(&m->m,
+               &m->profile->profile,
+               time,
+               length,
+               color_in);
+
+       color_out &= m->color_mask;
+
+       n_pkts = m->n_pkts[color_out];
+       n_bytes = m->n_bytes[color_out];
+
+       instr_meter_color_out_hbo_set(t, ip, color_out);
+
+       m->n_pkts[color_out] = n_pkts + 1;
+       m->n_bytes[color_out] = n_bytes + length;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+/*
+ * jmp.
+ */
+static struct action *
+action_find(struct rte_swx_pipeline *p, const char *name);
+
+static int
+instr_jmp_translate(struct rte_swx_pipeline *p __rte_unused,
+                   struct action *action __rte_unused,
+                   char **tokens,
+                   int n_tokens,
+                   struct instruction *instr,
+                   struct instruction_data *data)
+{
+       CHECK(n_tokens == 2, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       instr->type = INSTR_JMP;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       return 0;
+}
+
+static int
+instr_jmp_valid_translate(struct rte_swx_pipeline *p,
+                         struct action *action __rte_unused,
+                         char **tokens,
+                         int n_tokens,
+                         struct instruction *instr,
+                         struct instruction_data *data)
+{
+       struct header *h;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       h = header_parse(p, tokens[2]);
+       CHECK(h, EINVAL);
+
+       instr->type = INSTR_JMP_VALID;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       instr->jmp.header_id = h->id;
+       return 0;
+}
+
+static int
+instr_jmp_invalid_translate(struct rte_swx_pipeline *p,
+                           struct action *action __rte_unused,
+                           char **tokens,
+                           int n_tokens,
+                           struct instruction *instr,
+                           struct instruction_data *data)
+{
+       struct header *h;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       h = header_parse(p, tokens[2]);
+       CHECK(h, EINVAL);
+
+       instr->type = INSTR_JMP_INVALID;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       instr->jmp.header_id = h->id;
+       return 0;
+}
+
+static int
+instr_jmp_hit_translate(struct rte_swx_pipeline *p __rte_unused,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data)
+{
+       CHECK(!action, EINVAL);
+       CHECK(n_tokens == 2, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       instr->type = INSTR_JMP_HIT;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       return 0;
+}
+
+static int
+instr_jmp_miss_translate(struct rte_swx_pipeline *p __rte_unused,
+                        struct action *action,
+                        char **tokens,
+                        int n_tokens,
+                        struct instruction *instr,
+                        struct instruction_data *data)
+{
+       CHECK(!action, EINVAL);
+       CHECK(n_tokens == 2, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       instr->type = INSTR_JMP_MISS;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       return 0;
+}
+
+static int
+instr_jmp_action_hit_translate(struct rte_swx_pipeline *p,
+                              struct action *action,
+                              char **tokens,
+                              int n_tokens,
+                              struct instruction *instr,
+                              struct instruction_data *data)
+{
+       struct action *a;
+
+       CHECK(!action, EINVAL);
+       CHECK(n_tokens == 3, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       a = action_find(p, tokens[2]);
+       CHECK(a, EINVAL);
+
+       instr->type = INSTR_JMP_ACTION_HIT;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       instr->jmp.action_id = a->id;
+       return 0;
+}
+
+static int
+instr_jmp_action_miss_translate(struct rte_swx_pipeline *p,
+                               struct action *action,
+                               char **tokens,
+                               int n_tokens,
+                               struct instruction *instr,
+                               struct instruction_data *data)
+{
+       struct action *a;
+
+       CHECK(!action, EINVAL);
+       CHECK(n_tokens == 3, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       a = action_find(p, tokens[2]);
+       CHECK(a, EINVAL);
+
+       instr->type = INSTR_JMP_ACTION_MISS;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       instr->jmp.action_id = a->id;
+       return 0;
+}
+
+static int
+instr_jmp_eq_translate(struct rte_swx_pipeline *p,
+                      struct action *action,
+                      char **tokens,
+                      int n_tokens,
+                      struct instruction *instr,
+                      struct instruction_data *data)
+{
+       char *a = tokens[2], *b = tokens[3];
+       struct field *fa, *fb;
+       uint64_t b_val;
+       uint32_t a_struct_id, b_struct_id;
+
+       CHECK(n_tokens == 4, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       fa = struct_field_parse(p, action, a, &a_struct_id);
+       CHECK(fa, EINVAL);
+
+       /* JMP_EQ, JMP_EQ_MH, JMP_EQ_HM, JMP_EQ_HH. */
+       fb = struct_field_parse(p, action, b, &b_struct_id);
+       if (fb) {
+               instr->type = INSTR_JMP_EQ;
+               if (a[0] != 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_EQ_MH;
+               if (a[0] == 'h' && b[0] != 'h')
+                       instr->type = INSTR_JMP_EQ_HM;
+               if (a[0] == 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_EQ_HH;
+               instr->jmp.ip = NULL; /* Resolved later. */
+
+               instr->jmp.a.struct_id = (uint8_t)a_struct_id;
+               instr->jmp.a.n_bits = fa->n_bits;
+               instr->jmp.a.offset = fa->offset / 8;
+               instr->jmp.b.struct_id = (uint8_t)b_struct_id;
+               instr->jmp.b.n_bits = fb->n_bits;
+               instr->jmp.b.offset = fb->offset / 8;
+               return 0;
+       }
+
+       /* JMP_EQ_I. */
+       b_val = strtoull(b, &b, 0);
+       CHECK(!b[0], EINVAL);
+
+       if (a[0] == 'h')
+               b_val = hton64(b_val) >> (64 - fa->n_bits);
+
+       instr->type = INSTR_JMP_EQ_I;
+       instr->jmp.ip = NULL; /* Resolved later. */
+       instr->jmp.a.struct_id = (uint8_t)a_struct_id;
+       instr->jmp.a.n_bits = fa->n_bits;
+       instr->jmp.a.offset = fa->offset / 8;
+       instr->jmp.b_val = b_val;
+       return 0;
+}
+
+static int
+instr_jmp_neq_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data)
+{
+       char *a = tokens[2], *b = tokens[3];
+       struct field *fa, *fb;
+       uint64_t b_val;
+       uint32_t a_struct_id, b_struct_id;
+
+       CHECK(n_tokens == 4, EINVAL);
+
+       strcpy(data->jmp_label, tokens[1]);
+
+       fa = struct_field_parse(p, action, a, &a_struct_id);
+       CHECK(fa, EINVAL);
+
+       /* JMP_NEQ, JMP_NEQ_MH, JMP_NEQ_HM, JMP_NEQ_HH. */
+       fb = struct_field_parse(p, action, b, &b_struct_id);
+       if (fb) {
+               instr->type = INSTR_JMP_NEQ;
+               if (a[0] != 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_NEQ_MH;
+               if (a[0] == 'h' && b[0] != 'h')
+                       instr->type = INSTR_JMP_NEQ_HM;
+               if (a[0] == 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_NEQ_HH;
+               instr->jmp.ip = NULL; /* Resolved later. */
+
+               instr->jmp.a.struct_id = (uint8_t)a_struct_id;
+               instr->jmp.a.n_bits = fa->n_bits;
+               instr->jmp.a.offset = fa->offset / 8;
+               instr->jmp.b.struct_id = (uint8_t)b_struct_id;
+               instr->jmp.b.n_bits = fb->n_bits;
+               instr->jmp.b.offset = fb->offset / 8;
+               return 0;
+       }
+
+       /* JMP_NEQ_I. */
+       b_val = strtoull(b, &b, 0);
+       CHECK(!b[0], EINVAL);
 
        if (a[0] == 'h')
                b_val = hton64(b_val) >> (64 - fa->n_bits);
@@ -6125,14 +7344,36 @@ instr_jmp_eq_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_jmp_eq_s_exec(struct rte_swx_pipeline *p)
+instr_jmp_eq_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpeq (mh)\n", p->thread_id);
+
+       JMP_CMP_MH(t, ip, ==);
+}
+
+static inline void
+instr_jmp_eq_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpeq (hm)\n", p->thread_id);
+
+       JMP_CMP_HM(t, ip, ==);
+}
+
+static inline void
+instr_jmp_eq_hh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] jmpeq (s)\n", p->thread_id);
+       TRACE("[Thread %2u] jmpeq (hh)\n", p->thread_id);
 
-       JMP_CMP_S(t, ip, ==);
+       JMP_CMP_HH_FAST(t, ip, ==);
 }
 
 static inline void
@@ -6158,14 +7399,36 @@ instr_jmp_neq_exec(struct rte_swx_pipeline *p)
 }
 
 static inline void
-instr_jmp_neq_s_exec(struct rte_swx_pipeline *p)
+instr_jmp_neq_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpneq (mh)\n", p->thread_id);
+
+       JMP_CMP_MH(t, ip, !=);
+}
+
+static inline void
+instr_jmp_neq_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpneq (hm)\n", p->thread_id);
+
+       JMP_CMP_HM(t, ip, !=);
+}
+
+static inline void
+instr_jmp_neq_hh_exec(struct rte_swx_pipeline *p)
 {
        struct thread *t = &p->threads[p->thread_id];
        struct instruction *ip = t->ip;
 
-       TRACE("[Thread %2u] jmpneq (s)\n", p->thread_id);
+       TRACE("[Thread %2u] jmpneq (hh)\n", p->thread_id);
 
-       JMP_CMP_S(t, ip, !=);
+       JMP_CMP_HH_FAST(t, ip, !=);
 }
 
 static inline void
@@ -6391,6 +7654,14 @@ instr_translate(struct rte_swx_pipeline *p,
                                          instr,
                                          data);
 
+       if (!strcmp(tokens[tpos], "drop"))
+               return instr_drop_translate(p,
+                                           action,
+                                           &tokens[tpos],
+                                           n_tokens - tpos,
+                                           instr,
+                                           data);
+
        if (!strcmp(tokens[tpos], "extract"))
                return instr_hdr_extract_translate(p,
                                                   action,
@@ -6431,14 +7702,6 @@ instr_translate(struct rte_swx_pipeline *p,
                                           instr,
                                           data);
 
-       if (!strcmp(tokens[tpos], "dma"))
-               return instr_dma_translate(p,
-                                          action,
-                                          &tokens[tpos],
-                                          n_tokens - tpos,
-                                          instr,
-                                          data);
-
        if (!strcmp(tokens[tpos], "add"))
                return instr_alu_add_translate(p,
                                               action,
@@ -6543,6 +7806,22 @@ instr_translate(struct rte_swx_pipeline *p,
                                              instr,
                                              data);
 
+       if (!strcmp(tokens[tpos], "metprefetch"))
+               return instr_metprefetch_translate(p,
+                                                  action,
+                                                  &tokens[tpos],
+                                                  n_tokens - tpos,
+                                                  instr,
+                                                  data);
+
+       if (!strcmp(tokens[tpos], "meter"))
+               return instr_meter_translate(p,
+                                            action,
+                                            &tokens[tpos],
+                                            n_tokens - tpos,
+                                            instr,
+                                            data);
+
        if (!strcmp(tokens[tpos], "table"))
                return instr_table_translate(p,
                                             action,
@@ -6761,7 +8040,7 @@ instr_verify(struct rte_swx_pipeline *p __rte_unused,
                for (i = 0; i < n_instructions; i++) {
                        type = instr[i].type;
 
-                       if (type == INSTR_TX)
+                       if (instruction_is_tx(type))
                                break;
                }
                CHECK(i < n_instructions, EINVAL);
@@ -6770,7 +8049,7 @@ instr_verify(struct rte_swx_pipeline *p __rte_unused,
                 * jump.
                 */
                type = instr[n_instructions - 1].type;
-               CHECK((type == INSTR_TX) || (type == INSTR_JMP), EINVAL);
+               CHECK(instruction_is_tx(type) || (type == INSTR_JMP), EINVAL);
        }
 
        if (a) {
@@ -6781,7 +8060,7 @@ instr_verify(struct rte_swx_pipeline *p __rte_unused,
                for (i = 0; i < n_instructions; i++) {
                        type = instr[i].type;
 
-                       if ((type == INSTR_RETURN) || (type == INSTR_TX))
+                       if ((type == INSTR_RETURN) || instruction_is_tx(type))
                                break;
                }
                CHECK(i < n_instructions, EINVAL);
@@ -6790,8 +8069,34 @@ instr_verify(struct rte_swx_pipeline *p __rte_unused,
        return 0;
 }
 
+static uint32_t
+instr_compact(struct instruction *instructions,
+             struct instruction_data *instruction_data,
+             uint32_t n_instructions)
+{
+       uint32_t i, pos = 0;
+
+       /* Eliminate the invalid instructions that have been optimized out. */
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+
+               if (data->invalid)
+                       continue;
+
+               if (i != pos) {
+                       memcpy(&instructions[pos], instr, sizeof(*instr));
+                       memcpy(&instruction_data[pos], data, sizeof(*data));
+               }
+
+               pos++;
+       }
+
+       return pos;
+}
+
 static int
-instr_pattern_extract_many_detect(struct instruction *instr,
+instr_pattern_extract_many_search(struct instruction *instr,
                                  struct instruction_data *data,
                                  uint32_t n_instr,
                                  uint32_t *n_pattern_instr)
@@ -6820,9 +8125,9 @@ instr_pattern_extract_many_detect(struct instruction *instr,
 }
 
 static void
-instr_pattern_extract_many_optimize(struct instruction *instr,
-                                   struct instruction_data *data,
-                                   uint32_t n_instr)
+instr_pattern_extract_many_replace(struct instruction *instr,
+                                  struct instruction_data *data,
+                                  uint32_t n_instr)
 {
        uint32_t i;
 
@@ -6836,8 +8141,46 @@ instr_pattern_extract_many_optimize(struct instruction *instr,
        }
 }
 
+static uint32_t
+instr_pattern_extract_many_optimize(struct instruction *instructions,
+                                   struct instruction_data *instruction_data,
+                                   uint32_t n_instructions)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instructions; ) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+               uint32_t n_instr = 0;
+               int detected;
+
+               /* Extract many. */
+               detected = instr_pattern_extract_many_search(instr,
+                                                            data,
+                                                            n_instructions - i,
+                                                            &n_instr);
+               if (detected) {
+                       instr_pattern_extract_many_replace(instr,
+                                                          data,
+                                                          n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* No pattern starting at the current instruction. */
+               i++;
+       }
+
+       /* Eliminate the invalid instructions that have been optimized out. */
+       n_instructions = instr_compact(instructions,
+                                      instruction_data,
+                                      n_instructions);
+
+       return n_instructions;
+}
+
 static int
-instr_pattern_emit_many_tx_detect(struct instruction *instr,
+instr_pattern_emit_many_tx_search(struct instruction *instr,
                                  struct instruction_data *data,
                                  uint32_t n_instr,
                                  uint32_t *n_pattern_instr)
@@ -6861,7 +8204,7 @@ instr_pattern_emit_many_tx_detect(struct instruction *instr,
        if (!i)
                return 0;
 
-       if (instr[i].type != INSTR_TX)
+       if (!instruction_is_tx(instr[i].type))
                return 0;
 
        if (data[i].n_users)
@@ -6874,9 +8217,9 @@ instr_pattern_emit_many_tx_detect(struct instruction *instr,
 }
 
 static void
-instr_pattern_emit_many_tx_optimize(struct instruction *instr,
-                                   struct instruction_data *data,
-                                   uint32_t n_instr)
+instr_pattern_emit_many_tx_replace(struct instruction *instr,
+                                  struct instruction_data *data,
+                                  uint32_t n_instr)
 {
        uint32_t i;
 
@@ -6897,8 +8240,220 @@ instr_pattern_emit_many_tx_optimize(struct instruction *instr,
        data[i].invalid = 1;
 }
 
+static uint32_t
+instr_pattern_emit_many_tx_optimize(struct instruction *instructions,
+                                   struct instruction_data *instruction_data,
+                                   uint32_t n_instructions)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instructions; ) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+               uint32_t n_instr = 0;
+               int detected;
+
+               /* Emit many + TX. */
+               detected = instr_pattern_emit_many_tx_search(instr,
+                                                            data,
+                                                            n_instructions - i,
+                                                            &n_instr);
+               if (detected) {
+                       instr_pattern_emit_many_tx_replace(instr,
+                                                          data,
+                                                          n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* No pattern starting at the current instruction. */
+               i++;
+       }
+
+       /* Eliminate the invalid instructions that have been optimized out. */
+       n_instructions = instr_compact(instructions,
+                                      instruction_data,
+                                      n_instructions);
+
+       return n_instructions;
+}
+
+static uint32_t
+action_arg_src_mov_count(struct action *a,
+                        uint32_t arg_id,
+                        struct instruction *instructions,
+                        struct instruction_data *instruction_data,
+                        uint32_t n_instructions);
+
+static int
+instr_pattern_mov_all_validate_search(struct rte_swx_pipeline *p,
+                                     struct action *a,
+                                     struct instruction *instr,
+                                     struct instruction_data *data,
+                                     uint32_t n_instr,
+                                     struct instruction *instructions,
+                                     struct instruction_data *instruction_data,
+                                     uint32_t n_instructions,
+                                     uint32_t *n_pattern_instr)
+{
+       struct header *h;
+       uint32_t src_field_id, i, j;
+
+       /* Prerequisites. */
+       if (!a || !a->st)
+               return 0;
+
+       /* First instruction: MOV_HM. */
+       if (data[0].invalid || (instr[0].type != INSTR_MOV_HM))
+               return 0;
+
+       h = header_find_by_struct_id(p, instr[0].mov.dst.struct_id);
+       if (!h)
+               return 0;
+
+       for (src_field_id = 0; src_field_id < a->st->n_fields; src_field_id++)
+               if (instr[0].mov.src.offset == a->st->fields[src_field_id].offset / 8)
+                       break;
+
+       if (src_field_id == a->st->n_fields)
+               return 0;
+
+       if (instr[0].mov.dst.offset ||
+           (instr[0].mov.dst.n_bits != h->st->fields[0].n_bits) ||
+           instr[0].mov.src.struct_id ||
+           (instr[0].mov.src.n_bits != a->st->fields[src_field_id].n_bits) ||
+           (instr[0].mov.dst.n_bits != instr[0].mov.src.n_bits))
+               return 0;
+
+       if ((n_instr < h->st->n_fields + 1) ||
+            (a->st->n_fields < src_field_id + h->st->n_fields + 1))
+               return 0;
+
+       /* Subsequent instructions: MOV_HM. */
+       for (i = 1; i < h->st->n_fields; i++)
+               if (data[i].invalid ||
+                   data[i].n_users ||
+                   (instr[i].type != INSTR_MOV_HM) ||
+                   (instr[i].mov.dst.struct_id != h->struct_id) ||
+                   (instr[i].mov.dst.offset != h->st->fields[i].offset / 8) ||
+                   (instr[i].mov.dst.n_bits != h->st->fields[i].n_bits) ||
+                   instr[i].mov.src.struct_id ||
+                   (instr[i].mov.src.offset != a->st->fields[src_field_id + i].offset / 8) ||
+                   (instr[i].mov.src.n_bits != a->st->fields[src_field_id + i].n_bits) ||
+                   (instr[i].mov.dst.n_bits != instr[i].mov.src.n_bits))
+                       return 0;
+
+       /* Last instruction: HDR_VALIDATE. */
+       if ((instr[i].type != INSTR_HDR_VALIDATE) ||
+           (instr[i].valid.header_id != h->id))
+               return 0;
+
+       /* Check that none of the action args that are used as source for this
+        * DMA transfer are not used as source in any other mov instruction.
+        */
+       for (j = src_field_id; j < src_field_id + h->st->n_fields; j++) {
+               uint32_t n_users;
+
+               n_users = action_arg_src_mov_count(a,
+                                                  j,
+                                                  instructions,
+                                                  instruction_data,
+                                                  n_instructions);
+               if (n_users > 1)
+                       return 0;
+       }
+
+       *n_pattern_instr = 1 + i;
+       return 1;
+}
+
+static void
+instr_pattern_mov_all_validate_replace(struct rte_swx_pipeline *p,
+                                      struct action *a,
+                                      struct instruction *instr,
+                                      struct instruction_data *data,
+                                      uint32_t n_instr)
+{
+       struct header *h;
+       uint32_t src_field_id, src_offset, i;
+
+       /* Read from the instructions before they are modified. */
+       h = header_find_by_struct_id(p, instr[0].mov.dst.struct_id);
+       if (!h)
+               return;
+
+       for (src_field_id = 0; src_field_id < a->st->n_fields; src_field_id++)
+               if (instr[0].mov.src.offset == a->st->fields[src_field_id].offset / 8)
+                       break;
+
+       if (src_field_id == a->st->n_fields)
+               return;
+
+       src_offset = instr[0].mov.src.offset;
+
+       /* Modify the instructions. */
+       instr[0].type = INSTR_DMA_HT;
+       instr[0].dma.dst.header_id[0] = h->id;
+       instr[0].dma.dst.struct_id[0] = h->struct_id;
+       instr[0].dma.src.offset[0] = (uint8_t)src_offset;
+       instr[0].dma.n_bytes[0] = h->st->n_bits / 8;
+
+       for (i = 1; i < n_instr; i++)
+               data[i].invalid = 1;
+
+       /* Update the endianness of the action arguments to header endianness. */
+       for (i = 0; i < h->st->n_fields; i++)
+               a->args_endianness[src_field_id + i] = 1;
+}
+
+static uint32_t
+instr_pattern_mov_all_validate_optimize(struct rte_swx_pipeline *p,
+                                       struct action *a,
+                                       struct instruction *instructions,
+                                       struct instruction_data *instruction_data,
+                                       uint32_t n_instructions)
+{
+       uint32_t i;
+
+       if (!a || !a->st)
+               return n_instructions;
+
+       for (i = 0; i < n_instructions; ) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+               uint32_t n_instr = 0;
+               int detected;
+
+               /* Mov all + validate. */
+               detected = instr_pattern_mov_all_validate_search(p,
+                                                                a,
+                                                                instr,
+                                                                data,
+                                                                n_instructions - i,
+                                                                instructions,
+                                                                instruction_data,
+                                                                n_instructions,
+                                                                &n_instr);
+               if (detected) {
+                       instr_pattern_mov_all_validate_replace(p, a, instr, data, n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* No pattern starting at the current instruction. */
+               i++;
+       }
+
+       /* Eliminate the invalid instructions that have been optimized out. */
+       n_instructions = instr_compact(instructions,
+                                      instruction_data,
+                                      n_instructions);
+
+       return n_instructions;
+}
+
 static int
-instr_pattern_dma_many_detect(struct instruction *instr,
+instr_pattern_dma_many_search(struct instruction *instr,
                              struct instruction_data *data,
                              uint32_t n_instr,
                              uint32_t *n_pattern_instr)
@@ -6927,9 +8482,9 @@ instr_pattern_dma_many_detect(struct instruction *instr,
 }
 
 static void
-instr_pattern_dma_many_optimize(struct instruction *instr,
-                               struct instruction_data *data,
-                               uint32_t n_instr)
+instr_pattern_dma_many_replace(struct instruction *instr,
+                              struct instruction_data *data,
+                              uint32_t n_instr)
 {
        uint32_t i;
 
@@ -6945,11 +8500,11 @@ instr_pattern_dma_many_optimize(struct instruction *instr,
 }
 
 static uint32_t
-instr_optimize(struct instruction *instructions,
+instr_pattern_dma_many_optimize(struct instruction *instructions,
               struct instruction_data *instruction_data,
               uint32_t n_instructions)
 {
-       uint32_t i, pos = 0;
+       uint32_t i;
 
        for (i = 0; i < n_instructions; ) {
                struct instruction *instr = &instructions[i];
@@ -6957,39 +8512,13 @@ instr_optimize(struct instruction *instructions,
                uint32_t n_instr = 0;
                int detected;
 
-               /* Extract many. */
-               detected = instr_pattern_extract_many_detect(instr,
-                                                            data,
-                                                            n_instructions - i,
-                                                            &n_instr);
-               if (detected) {
-                       instr_pattern_extract_many_optimize(instr,
-                                                           data,
-                                                           n_instr);
-                       i += n_instr;
-                       continue;
-               }
-
-               /* Emit many + TX. */
-               detected = instr_pattern_emit_many_tx_detect(instr,
-                                                            data,
-                                                            n_instructions - i,
-                                                            &n_instr);
-               if (detected) {
-                       instr_pattern_emit_many_tx_optimize(instr,
-                                                           data,
-                                                           n_instr);
-                       i += n_instr;
-                       continue;
-               }
-
                /* DMA many. */
-               detected = instr_pattern_dma_many_detect(instr,
+               detected = instr_pattern_dma_many_search(instr,
                                                         data,
                                                         n_instructions - i,
                                                         &n_instr);
                if (detected) {
-                       instr_pattern_dma_many_optimize(instr, data, n_instr);
+                       instr_pattern_dma_many_replace(instr, data, n_instr);
                        i += n_instr;
                        continue;
                }
@@ -6999,22 +8528,43 @@ instr_optimize(struct instruction *instructions,
        }
 
        /* Eliminate the invalid instructions that have been optimized out. */
-       for (i = 0; i < n_instructions; i++) {
-               struct instruction *instr = &instructions[i];
-               struct instruction_data *data = &instruction_data[i];
+       n_instructions = instr_compact(instructions,
+                                      instruction_data,
+                                      n_instructions);
 
-               if (data->invalid)
-                       continue;
+       return n_instructions;
+}
 
-               if (i != pos) {
-                       memcpy(&instructions[pos], instr, sizeof(*instr));
-                       memcpy(&instruction_data[pos], data, sizeof(*data));
-               }
+static uint32_t
+instr_optimize(struct rte_swx_pipeline *p,
+              struct action *a,
+              struct instruction *instructions,
+              struct instruction_data *instruction_data,
+              uint32_t n_instructions)
+{
+       /* Extract many. */
+       n_instructions = instr_pattern_extract_many_optimize(instructions,
+                                                            instruction_data,
+                                                            n_instructions);
 
-               pos++;
-       }
+       /* Emit many + TX. */
+       n_instructions = instr_pattern_emit_many_tx_optimize(instructions,
+                                                            instruction_data,
+                                                            n_instructions);
 
-       return pos;
+       /* Mov all + validate. */
+       n_instructions = instr_pattern_mov_all_validate_optimize(p,
+                                                                a,
+                                                                instructions,
+                                                                instruction_data,
+                                                                n_instructions);
+
+       /* DMA many. */
+       n_instructions = instr_pattern_dma_many_optimize(instructions,
+                                                        instruction_data,
+                                                        n_instructions);
+
+       return n_instructions;
 }
 
 static int
@@ -7036,20 +8586,20 @@ instruction_config(struct rte_swx_pipeline *p,
        /* Memory allocation. */
        instr = calloc(n_instructions, sizeof(struct instruction));
        if (!instr) {
-               err = ENOMEM;
+               err = -ENOMEM;
                goto error;
        }
 
        data = calloc(n_instructions, sizeof(struct instruction_data));
        if (!data) {
-               err = ENOMEM;
+               err = -ENOMEM;
                goto error;
        }
 
        for (i = 0; i < n_instructions; i++) {
                char *string = strdup(instructions[i]);
                if (!string) {
-                       err = ENOMEM;
+                       err = -ENOMEM;
                        goto error;
                }
 
@@ -7070,7 +8620,7 @@ instruction_config(struct rte_swx_pipeline *p,
        if (err)
                goto error;
 
-       n_instructions = instr_optimize(instr, data, n_instructions);
+       n_instructions = instr_optimize(p, a, instr, data, n_instructions);
 
        err = instr_jmp_resolve(instr, data, n_instructions);
        if (err)
@@ -7098,6 +8648,7 @@ typedef void (*instr_exec_t)(struct rte_swx_pipeline *);
 static instr_exec_t instruction_table[] = {
        [INSTR_RX] = instr_rx_exec,
        [INSTR_TX] = instr_tx_exec,
+       [INSTR_TX_I] = instr_tx_i_exec,
 
        [INSTR_HDR_EXTRACT] = instr_hdr_extract_exec,
        [INSTR_HDR_EXTRACT2] = instr_hdr_extract2_exec,
@@ -7122,7 +8673,9 @@ static instr_exec_t instruction_table[] = {
        [INSTR_HDR_INVALIDATE] = instr_hdr_invalidate_exec,
 
        [INSTR_MOV] = instr_mov_exec,
-       [INSTR_MOV_S] = instr_mov_s_exec,
+       [INSTR_MOV_MH] = instr_mov_mh_exec,
+       [INSTR_MOV_HM] = instr_mov_hm_exec,
+       [INSTR_MOV_HH] = instr_mov_hh_exec,
        [INSTR_MOV_I] = instr_mov_i_exec,
 
        [INSTR_DMA_HT] = instr_dma_ht_exec,
@@ -7154,15 +8707,21 @@ static instr_exec_t instruction_table[] = {
        [INSTR_ALU_CKSUB_FIELD] = instr_alu_cksub_field_exec,
 
        [INSTR_ALU_AND] = instr_alu_and_exec,
-       [INSTR_ALU_AND_S] = instr_alu_and_s_exec,
+       [INSTR_ALU_AND_MH] = instr_alu_and_mh_exec,
+       [INSTR_ALU_AND_HM] = instr_alu_and_hm_exec,
+       [INSTR_ALU_AND_HH] = instr_alu_and_hh_exec,
        [INSTR_ALU_AND_I] = instr_alu_and_i_exec,
 
        [INSTR_ALU_OR] = instr_alu_or_exec,
-       [INSTR_ALU_OR_S] = instr_alu_or_s_exec,
+       [INSTR_ALU_OR_MH] = instr_alu_or_mh_exec,
+       [INSTR_ALU_OR_HM] = instr_alu_or_hm_exec,
+       [INSTR_ALU_OR_HH] = instr_alu_or_hh_exec,
        [INSTR_ALU_OR_I] = instr_alu_or_i_exec,
 
        [INSTR_ALU_XOR] = instr_alu_xor_exec,
-       [INSTR_ALU_XOR_S] = instr_alu_xor_s_exec,
+       [INSTR_ALU_XOR_MH] = instr_alu_xor_mh_exec,
+       [INSTR_ALU_XOR_HM] = instr_alu_xor_hm_exec,
+       [INSTR_ALU_XOR_HH] = instr_alu_xor_hh_exec,
        [INSTR_ALU_XOR_I] = instr_alu_xor_i_exec,
 
        [INSTR_ALU_SHL] = instr_alu_shl_exec,
@@ -7210,6 +8769,23 @@ static instr_exec_t instruction_table[] = {
        [INSTR_REGADD_RIM] = instr_regadd_rim_exec,
        [INSTR_REGADD_RII] = instr_regadd_rii_exec,
 
+       [INSTR_METPREFETCH_H] = instr_metprefetch_h_exec,
+       [INSTR_METPREFETCH_M] = instr_metprefetch_m_exec,
+       [INSTR_METPREFETCH_I] = instr_metprefetch_i_exec,
+
+       [INSTR_METER_HHM] = instr_meter_hhm_exec,
+       [INSTR_METER_HHI] = instr_meter_hhi_exec,
+       [INSTR_METER_HMM] = instr_meter_hmm_exec,
+       [INSTR_METER_HMI] = instr_meter_hmi_exec,
+       [INSTR_METER_MHM] = instr_meter_mhm_exec,
+       [INSTR_METER_MHI] = instr_meter_mhi_exec,
+       [INSTR_METER_MMM] = instr_meter_mmm_exec,
+       [INSTR_METER_MMI] = instr_meter_mmi_exec,
+       [INSTR_METER_IHM] = instr_meter_ihm_exec,
+       [INSTR_METER_IHI] = instr_meter_ihi_exec,
+       [INSTR_METER_IMM] = instr_meter_imm_exec,
+       [INSTR_METER_IMI] = instr_meter_imi_exec,
+
        [INSTR_TABLE] = instr_table_exec,
        [INSTR_EXTERN_OBJ] = instr_extern_obj_exec,
        [INSTR_EXTERN_FUNC] = instr_extern_func_exec,
@@ -7223,11 +8799,15 @@ static instr_exec_t instruction_table[] = {
        [INSTR_JMP_ACTION_MISS] = instr_jmp_action_miss_exec,
 
        [INSTR_JMP_EQ] = instr_jmp_eq_exec,
-       [INSTR_JMP_EQ_S] = instr_jmp_eq_s_exec,
+       [INSTR_JMP_EQ_MH] = instr_jmp_eq_mh_exec,
+       [INSTR_JMP_EQ_HM] = instr_jmp_eq_hm_exec,
+       [INSTR_JMP_EQ_HH] = instr_jmp_eq_hh_exec,
        [INSTR_JMP_EQ_I] = instr_jmp_eq_i_exec,
 
        [INSTR_JMP_NEQ] = instr_jmp_neq_exec,
-       [INSTR_JMP_NEQ_S] = instr_jmp_neq_s_exec,
+       [INSTR_JMP_NEQ_MH] = instr_jmp_neq_mh_exec,
+       [INSTR_JMP_NEQ_HM] = instr_jmp_neq_hm_exec,
+       [INSTR_JMP_NEQ_HH] = instr_jmp_neq_hh_exec,
        [INSTR_JMP_NEQ_I] = instr_jmp_neq_i_exec,
 
        [INSTR_JMP_LT] = instr_jmp_lt_exec,
@@ -7329,6 +8909,13 @@ rte_swx_pipeline_action_config(struct rte_swx_pipeline *p,
        /* Node allocation. */
        a = calloc(1, sizeof(struct action));
        CHECK(a, ENOMEM);
+       if (args_struct_type) {
+               a->args_endianness = calloc(args_struct_type->n_fields, sizeof(int));
+               if (!a->args_endianness) {
+                       free(a);
+                       CHECK(0, ENOMEM);
+               }
+       }
 
        /* Node initialization. */
        strcpy(a->name, name);
@@ -7338,6 +8925,7 @@ rte_swx_pipeline_action_config(struct rte_swx_pipeline *p,
        /* Instruction translation. */
        err = instruction_config(p, a, instructions, n_instructions);
        if (err) {
+               free(a->args_endianness);
                free(a);
                return err;
        }
@@ -7389,6 +8977,40 @@ action_free(struct rte_swx_pipeline *p)
        }
 }
 
+static uint32_t
+action_arg_src_mov_count(struct action *a,
+                        uint32_t arg_id,
+                        struct instruction *instructions,
+                        struct instruction_data *instruction_data,
+                        uint32_t n_instructions)
+{
+       uint32_t offset, n_users = 0, i;
+
+       if (!a->st ||
+           (arg_id >= a->st->n_fields) ||
+           !instructions ||
+           !instruction_data ||
+           !n_instructions)
+               return 0;
+
+       offset = a->st->fields[arg_id].offset / 8;
+
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+
+               if (data->invalid ||
+                   ((instr->type != INSTR_MOV) && (instr->type != INSTR_MOV_HM)) ||
+                   instr->mov.src.struct_id ||
+                   (instr->mov.src.offset != offset))
+                       continue;
+
+               n_users++;
+       }
+
+       return n_users;
+}
+
 /*
  * Table.
  */
@@ -7470,39 +9092,146 @@ rte_swx_pipeline_table_type_register(struct rte_swx_pipeline *p,
        CHECK(ops->lkp, EINVAL);
        CHECK(ops->free, EINVAL);
 
-       /* Node allocation. */
-       elem = calloc(1, sizeof(struct table_type));
-       CHECK(elem, ENOMEM);
+       /* Node allocation. */
+       elem = calloc(1, sizeof(struct table_type));
+       CHECK(elem, ENOMEM);
+
+       /* Node initialization. */
+       strcpy(elem->name, name);
+       elem->match_type = match_type;
+       memcpy(&elem->ops, ops, sizeof(*ops));
+
+       /* Node add to tailq. */
+       TAILQ_INSERT_TAIL(&p->table_types, elem, node);
+
+       return 0;
+}
+
+static enum rte_swx_table_match_type
+table_match_type_resolve(struct rte_swx_match_field_params *fields,
+                        uint32_t n_fields,
+                        uint32_t max_offset_field_id)
+{
+       uint32_t n_fields_em = 0, i;
+
+       for (i = 0; i < n_fields; i++)
+               if (fields[i].match_type == RTE_SWX_TABLE_MATCH_EXACT)
+                       n_fields_em++;
+
+       if (n_fields_em == n_fields)
+               return RTE_SWX_TABLE_MATCH_EXACT;
+
+       if ((n_fields_em == n_fields - 1) &&
+           (fields[max_offset_field_id].match_type == RTE_SWX_TABLE_MATCH_LPM))
+               return RTE_SWX_TABLE_MATCH_LPM;
+
+       return RTE_SWX_TABLE_MATCH_WILDCARD;
+}
+
+static int
+table_match_fields_check(struct rte_swx_pipeline *p,
+                        struct rte_swx_pipeline_table_params *params,
+                        struct header **header,
+                        uint32_t *min_offset_field_id,
+                        uint32_t *max_offset_field_id)
+{
+       struct header *h0 = NULL;
+       struct field *hf, *mf;
+       uint32_t *offset = NULL, min_offset, max_offset, min_offset_pos, max_offset_pos, i;
+       int status = 0;
+
+       /* Return if no match fields. */
+       if (!params->n_fields) {
+               if (params->fields) {
+                       status = -EINVAL;
+                       goto end;
+               }
+
+               return 0;
+       }
+
+       /* Memory allocation. */
+       offset = calloc(params->n_fields, sizeof(uint32_t));
+       if (!offset) {
+               status = -ENOMEM;
+               goto end;
+       }
+
+       /* Check that all the match fields belong to either the same header or
+        * to the meta-data.
+        */
+       hf = header_field_parse(p, params->fields[0].name, &h0);
+       mf = metadata_field_parse(p, params->fields[0].name);
+       if (!hf && !mf) {
+               status = -EINVAL;
+               goto end;
+       }
+
+       offset[0] = h0 ? hf->offset : mf->offset;
+
+       for (i = 1; i < params->n_fields; i++)
+               if (h0) {
+                       struct header *h;
+
+                       hf = header_field_parse(p, params->fields[i].name, &h);
+                       if (!hf || (h->id != h0->id)) {
+                               status = -EINVAL;
+                               goto end;
+                       }
+
+                       offset[i] = hf->offset;
+               } else {
+                       mf = metadata_field_parse(p, params->fields[i].name);
+                       if (!mf) {
+                               status = -EINVAL;
+                               goto end;
+                       }
+
+                       offset[i] = mf->offset;
+               }
+
+       /* Check that there are no duplicated match fields. */
+       for (i = 0; i < params->n_fields; i++) {
+               uint32_t j;
 
-       /* Node initialization. */
-       strcpy(elem->name, name);
-       elem->match_type = match_type;
-       memcpy(&elem->ops, ops, sizeof(*ops));
+               for (j = 0; j < i; j++)
+                       if (offset[j] == offset[i]) {
+                               status = -EINVAL;
+                               goto end;
+                       }
+       }
 
-       /* Node add to tailq. */
-       TAILQ_INSERT_TAIL(&p->table_types, elem, node);
+       /* Find the min and max offset fields. */
+       min_offset = offset[0];
+       max_offset = offset[0];
+       min_offset_pos = 0;
+       max_offset_pos = 0;
 
-       return 0;
-}
+       for (i = 1; i < params->n_fields; i++) {
+               if (offset[i] < min_offset) {
+                       min_offset = offset[i];
+                       min_offset_pos = i;
+               }
 
-static enum rte_swx_table_match_type
-table_match_type_resolve(struct rte_swx_match_field_params *fields,
-                        uint32_t n_fields)
-{
-       uint32_t i;
+               if (offset[i] > max_offset) {
+                       max_offset = offset[i];
+                       max_offset_pos = i;
+               }
+       }
 
-       for (i = 0; i < n_fields; i++)
-               if (fields[i].match_type != RTE_SWX_TABLE_MATCH_EXACT)
-                       break;
+       /* Return. */
+       if (header)
+               *header = h0;
 
-       if (i == n_fields)
-               return RTE_SWX_TABLE_MATCH_EXACT;
+       if (min_offset_field_id)
+               *min_offset_field_id = min_offset_pos;
 
-       if ((i == n_fields - 1) &&
-           (fields[i].match_type == RTE_SWX_TABLE_MATCH_LPM))
-               return RTE_SWX_TABLE_MATCH_LPM;
+       if (max_offset_field_id)
+               *max_offset_field_id = max_offset_pos;
 
-       return RTE_SWX_TABLE_MATCH_WILDCARD;
+end:
+       free(offset);
+       return status;
 }
 
 int
@@ -7517,8 +9246,8 @@ rte_swx_pipeline_table_config(struct rte_swx_pipeline *p,
        struct table *t;
        struct action *default_action;
        struct header *header = NULL;
-       int is_header = 0;
-       uint32_t offset_prev = 0, action_data_size_max = 0, i;
+       uint32_t action_data_size_max = 0, min_offset_field_id = 0, max_offset_field_id = 0, i;
+       int status = 0;
 
        CHECK(p, EINVAL);
 
@@ -7528,35 +9257,13 @@ rte_swx_pipeline_table_config(struct rte_swx_pipeline *p,
        CHECK(params, EINVAL);
 
        /* Match checks. */
-       CHECK(!params->n_fields || params->fields, EINVAL);
-       for (i = 0; i < params->n_fields; i++) {
-               struct rte_swx_match_field_params *field = &params->fields[i];
-               struct header *h;
-               struct field *hf, *mf;
-               uint32_t offset;
-
-               CHECK_NAME(field->name, EINVAL);
-
-               hf = header_field_parse(p, field->name, &h);
-               mf = metadata_field_parse(p, field->name);
-               CHECK(hf || mf, EINVAL);
-
-               offset = hf ? hf->offset : mf->offset;
-
-               if (i == 0) {
-                       is_header = hf ? 1 : 0;
-                       header = hf ? h : NULL;
-                       offset_prev = offset;
-
-                       continue;
-               }
-
-               CHECK((is_header && hf && (h->id == header->id)) ||
-                     (!is_header && mf), EINVAL);
-
-               CHECK(offset > offset_prev, EINVAL);
-               offset_prev = offset;
-       }
+       status = table_match_fields_check(p,
+                                         params,
+                                         &header,
+                                         &min_offset_field_id,
+                                         &max_offset_field_id);
+       if (status)
+               return status;
 
        /* Action checks. */
        CHECK(params->n_actions, EINVAL);
@@ -7594,7 +9301,8 @@ rte_swx_pipeline_table_config(struct rte_swx_pipeline *p,
                enum rte_swx_table_match_type match_type;
 
                match_type = table_match_type_resolve(params->fields,
-                                                     params->n_fields);
+                                                     params->n_fields,
+                                                     max_offset_field_id);
                type = table_type_resolve(p,
                                          recommended_table_type_name,
                                          match_type);
@@ -7641,12 +9349,11 @@ rte_swx_pipeline_table_config(struct rte_swx_pipeline *p,
                struct match_field *f = &t->fields[i];
 
                f->match_type = field->match_type;
-               f->field = is_header ?
+               f->field = header ?
                        header_field_parse(p, field->name, NULL) :
                        metadata_field_parse(p, field->name);
        }
        t->n_fields = params->n_fields;
-       t->is_header = is_header;
        t->header = header;
 
        for (i = 0; i < params->n_actions; i++)
@@ -7683,9 +9390,21 @@ table_params_get(struct table *table)
        if (!params)
                return NULL;
 
-       /* Key offset and size. */
+       /* Find first (smallest offset) and last (biggest offset) match fields. */
        first = table->fields[0].field;
-       last = table->fields[table->n_fields - 1].field;
+       last = table->fields[0].field;
+
+       for (i = 0; i < table->n_fields; i++) {
+               struct field *f = table->fields[i].field;
+
+               if (f->offset < first->offset)
+                       first = f;
+
+               if (f->offset > last->offset)
+                       last = f;
+       }
+
+       /* Key offset and size. */
        key_offset = first->offset / 8;
        key_size = (last->offset + last->n_bits - first->offset) / 8;
 
@@ -7829,6 +9548,16 @@ table_build(struct rte_swx_pipeline *p)
 {
        uint32_t i;
 
+       /* Per pipeline: table statistics. */
+       p->table_stats = calloc(p->n_tables, sizeof(struct table_statistics));
+       CHECK(p->table_stats, ENOMEM);
+
+       for (i = 0; i < p->n_tables; i++) {
+               p->table_stats[i].n_pkts_action = calloc(p->n_actions, sizeof(uint64_t));
+               CHECK(p->table_stats[i].n_pkts_action, ENOMEM);
+       }
+
+       /* Per thread: table runt-time. */
        for (i = 0; i < RTE_SWX_PIPELINE_THREADS_MAX; i++) {
                struct thread *t = &p->threads[i];
                struct table *table;
@@ -7854,7 +9583,7 @@ table_build(struct rte_swx_pipeline *p)
                                }
 
                                /* r->key. */
-                               r->key = table->is_header ?
+                               r->key = table->header ?
                                        &t->structs[table->header->struct_id] :
                                        &t->structs[p->metadata_struct_id];
                        } else {
@@ -7887,6 +9616,13 @@ table_build_free(struct rte_swx_pipeline *p)
                free(t->tables);
                t->tables = NULL;
        }
+
+       if (p->table_stats) {
+               for (i = 0; i < p->n_tables; i++)
+                       free(p->table_stats[i].n_pkts_action);
+
+               free(p->table_stats);
+       }
 }
 
 static void
@@ -8048,6 +9784,182 @@ regarray_free(struct rte_swx_pipeline *p)
        }
 }
 
+/*
+ * Meter array.
+ */
+static struct meter_profile *
+meter_profile_find(struct rte_swx_pipeline *p, const char *name)
+{
+       struct meter_profile *elem;
+
+       TAILQ_FOREACH(elem, &p->meter_profiles, node)
+               if (!strcmp(elem->name, name))
+                       return elem;
+
+       return NULL;
+}
+
+static struct metarray *
+metarray_find(struct rte_swx_pipeline *p, const char *name)
+{
+       struct metarray *elem;
+
+       TAILQ_FOREACH(elem, &p->metarrays, node)
+               if (!strcmp(elem->name, name))
+                       return elem;
+
+       return NULL;
+}
+
+static struct metarray *
+metarray_find_by_id(struct rte_swx_pipeline *p, uint32_t id)
+{
+       struct metarray *elem = NULL;
+
+       TAILQ_FOREACH(elem, &p->metarrays, node)
+               if (elem->id == id)
+                       return elem;
+
+       return NULL;
+}
+
+int
+rte_swx_pipeline_metarray_config(struct rte_swx_pipeline *p,
+                                const char *name,
+                                uint32_t size)
+{
+       struct metarray *m;
+
+       CHECK(p, EINVAL);
+
+       CHECK_NAME(name, EINVAL);
+       CHECK(!metarray_find(p, name), EEXIST);
+
+       CHECK(size, EINVAL);
+       size = rte_align32pow2(size);
+
+       /* Memory allocation. */
+       m = calloc(1, sizeof(struct metarray));
+       CHECK(m, ENOMEM);
+
+       /* Node initialization. */
+       strcpy(m->name, name);
+       m->size = size;
+       m->id = p->n_metarrays;
+
+       /* Node add to tailq. */
+       TAILQ_INSERT_TAIL(&p->metarrays, m, node);
+       p->n_metarrays++;
+
+       return 0;
+}
+
+struct meter_profile meter_profile_default = {
+       .node = {0},
+       .name = "",
+       .params = {0},
+
+       .profile = {
+               .cbs = 10000,
+               .pbs = 10000,
+               .cir_period = 1,
+               .cir_bytes_per_period = 1,
+               .pir_period = 1,
+               .pir_bytes_per_period = 1,
+       },
+
+       .n_users = 0,
+};
+
+static void
+meter_init(struct meter *m)
+{
+       memset(m, 0, sizeof(struct meter));
+       rte_meter_trtcm_config(&m->m, &meter_profile_default.profile);
+       m->profile = &meter_profile_default;
+       m->color_mask = RTE_COLOR_GREEN;
+
+       meter_profile_default.n_users++;
+}
+
+static int
+metarray_build(struct rte_swx_pipeline *p)
+{
+       struct metarray *m;
+
+       if (!p->n_metarrays)
+               return 0;
+
+       p->metarray_runtime = calloc(p->n_metarrays, sizeof(struct metarray_runtime));
+       CHECK(p->metarray_runtime, ENOMEM);
+
+       TAILQ_FOREACH(m, &p->metarrays, node) {
+               struct metarray_runtime *r = &p->metarray_runtime[m->id];
+               uint32_t i;
+
+               r->metarray = env_malloc(m->size * sizeof(struct meter),
+                                        RTE_CACHE_LINE_SIZE,
+                                        p->numa_node);
+               CHECK(r->metarray, ENOMEM);
+
+               for (i = 0; i < m->size; i++)
+                       meter_init(&r->metarray[i]);
+
+               r->size_mask = m->size - 1;
+       }
+
+       return 0;
+}
+
+static void
+metarray_build_free(struct rte_swx_pipeline *p)
+{
+       uint32_t i;
+
+       if (!p->metarray_runtime)
+               return;
+
+       for (i = 0; i < p->n_metarrays; i++) {
+               struct metarray *m = metarray_find_by_id(p, i);
+               struct metarray_runtime *r = &p->metarray_runtime[i];
+
+               env_free(r->metarray, m->size * sizeof(struct meter));
+       }
+
+       free(p->metarray_runtime);
+       p->metarray_runtime = NULL;
+}
+
+static void
+metarray_free(struct rte_swx_pipeline *p)
+{
+       metarray_build_free(p);
+
+       /* Meter arrays. */
+       for ( ; ; ) {
+               struct metarray *elem;
+
+               elem = TAILQ_FIRST(&p->metarrays);
+               if (!elem)
+                       break;
+
+               TAILQ_REMOVE(&p->metarrays, elem, node);
+               free(elem);
+       }
+
+       /* Meter profiles. */
+       for ( ; ; ) {
+               struct meter_profile *elem;
+
+               elem = TAILQ_FIRST(&p->meter_profiles);
+               if (!elem)
+                       break;
+
+               TAILQ_REMOVE(&p->meter_profiles, elem, node);
+               free(elem);
+       }
+}
+
 /*
  * Pipeline.
  */
@@ -8077,6 +9989,8 @@ rte_swx_pipeline_config(struct rte_swx_pipeline **p, int numa_node)
        TAILQ_INIT(&pipeline->table_types);
        TAILQ_INIT(&pipeline->tables);
        TAILQ_INIT(&pipeline->regarrays);
+       TAILQ_INIT(&pipeline->meter_profiles);
+       TAILQ_INIT(&pipeline->metarrays);
 
        pipeline->n_structs = 1; /* Struct 0 is reserved for action_data. */
        pipeline->numa_node = numa_node;
@@ -8093,6 +10007,7 @@ rte_swx_pipeline_free(struct rte_swx_pipeline *p)
 
        free(p->instructions);
 
+       metarray_free(p);
        regarray_free(p);
        table_state_free(p);
        table_free(p);
@@ -8182,10 +10097,15 @@ rte_swx_pipeline_build(struct rte_swx_pipeline *p)
        if (status)
                goto error;
 
+       status = metarray_build(p);
+       if (status)
+               goto error;
+
        p->build_done = 1;
        return 0;
 
 error:
+       metarray_build_free(p);
        regarray_build_free(p);
        table_state_build_free(p);
        table_build_free(p);
@@ -8248,6 +10168,7 @@ rte_swx_ctl_pipeline_info_get(struct rte_swx_pipeline *p,
        pipeline->n_actions = n_actions;
        pipeline->n_tables = n_tables;
        pipeline->n_regarrays = p->n_regarrays;
+       pipeline->n_metarrays = p->n_metarrays;
 
        return 0;
 }
@@ -8300,6 +10221,7 @@ rte_swx_ctl_action_arg_info_get(struct rte_swx_pipeline *p,
        arg = &a->st->fields[action_arg_id];
        strcpy(action_arg->name, arg->name);
        action_arg->n_bits = arg->n_bits;
+       action_arg->is_network_byte_order = a->args_endianness[action_arg_id];
 
        return 0;
 }
@@ -8345,7 +10267,7 @@ rte_swx_ctl_table_match_field_info_get(struct rte_swx_pipeline *p,
 
        f = &t->fields[match_field_id];
        match_field->match_type = f->match_type;
-       match_field->is_header = t->is_header;
+       match_field->is_header = t->header ? 1 : 0;
        match_field->n_bits = f->field->n_bits;
        match_field->offset = f->field->offset;
 
@@ -8456,6 +10378,33 @@ rte_swx_ctl_pipeline_port_out_stats_read(struct rte_swx_pipeline *p,
        return 0;
 }
 
+int
+rte_swx_ctl_pipeline_table_stats_read(struct rte_swx_pipeline *p,
+                                     const char *table_name,
+                                     struct rte_swx_table_stats *stats)
+{
+       struct table *table;
+       struct table_statistics *table_stats;
+
+       if (!p || !table_name || !table_name[0] || !stats || !stats->n_pkts_action)
+               return -EINVAL;
+
+       table = table_find(p, table_name);
+       if (!table)
+               return -EINVAL;
+
+       table_stats = &p->table_stats[table->id];
+
+       memcpy(&stats->n_pkts_action,
+              &table_stats->n_pkts_action,
+              p->n_actions * sizeof(uint64_t));
+
+       stats->n_pkts_hit = table_stats->n_pkts_hit[1];
+       stats->n_pkts_miss = table_stats->n_pkts_hit[0];
+
+       return 0;
+}
+
 int
 rte_swx_ctl_regarray_info_get(struct rte_swx_pipeline *p,
                              uint32_t regarray_id,
@@ -8516,3 +10465,166 @@ rte_swx_ctl_pipeline_regarray_write(struct rte_swx_pipeline *p,
        r->regarray[regarray_index] = value;
        return 0;
 }
+
+int
+rte_swx_ctl_metarray_info_get(struct rte_swx_pipeline *p,
+                             uint32_t metarray_id,
+                             struct rte_swx_ctl_metarray_info *metarray)
+{
+       struct metarray *m;
+
+       if (!p || !metarray)
+               return -EINVAL;
+
+       m = metarray_find_by_id(p, metarray_id);
+       if (!m)
+               return -EINVAL;
+
+       strcpy(metarray->name, m->name);
+       metarray->size = m->size;
+       return 0;
+}
+
+int
+rte_swx_ctl_meter_profile_add(struct rte_swx_pipeline *p,
+                             const char *name,
+                             struct rte_meter_trtcm_params *params)
+{
+       struct meter_profile *mp;
+       int status;
+
+       CHECK(p, EINVAL);
+       CHECK_NAME(name, EINVAL);
+       CHECK(params, EINVAL);
+       CHECK(!meter_profile_find(p, name), EEXIST);
+
+       /* Node allocation. */
+       mp = calloc(1, sizeof(struct meter_profile));
+       CHECK(mp, ENOMEM);
+
+       /* Node initialization. */
+       strcpy(mp->name, name);
+       memcpy(&mp->params, params, sizeof(struct rte_meter_trtcm_params));
+       status = rte_meter_trtcm_profile_config(&mp->profile, params);
+       if (status) {
+               free(mp);
+               CHECK(0, EINVAL);
+       }
+
+       /* Node add to tailq. */
+       TAILQ_INSERT_TAIL(&p->meter_profiles, mp, node);
+
+       return 0;
+}
+
+int
+rte_swx_ctl_meter_profile_delete(struct rte_swx_pipeline *p,
+                                const char *name)
+{
+       struct meter_profile *mp;
+
+       CHECK(p, EINVAL);
+       CHECK_NAME(name, EINVAL);
+
+       mp = meter_profile_find(p, name);
+       CHECK(mp, EINVAL);
+       CHECK(!mp->n_users, EBUSY);
+
+       /* Remove node from tailq. */
+       TAILQ_REMOVE(&p->meter_profiles, mp, node);
+       free(mp);
+
+       return 0;
+}
+
+int
+rte_swx_ctl_meter_reset(struct rte_swx_pipeline *p,
+                       const char *metarray_name,
+                       uint32_t metarray_index)
+{
+       struct meter_profile *mp_old;
+       struct metarray *metarray;
+       struct metarray_runtime *metarray_runtime;
+       struct meter *m;
+
+       CHECK(p, EINVAL);
+       CHECK_NAME(metarray_name, EINVAL);
+
+       metarray = metarray_find(p, metarray_name);
+       CHECK(metarray, EINVAL);
+       CHECK(metarray_index < metarray->size, EINVAL);
+
+       metarray_runtime = &p->metarray_runtime[metarray->id];
+       m = &metarray_runtime->metarray[metarray_index];
+       mp_old = m->profile;
+
+       meter_init(m);
+
+       mp_old->n_users--;
+
+       return 0;
+}
+
+int
+rte_swx_ctl_meter_set(struct rte_swx_pipeline *p,
+                     const char *metarray_name,
+                     uint32_t metarray_index,
+                     const char *profile_name)
+{
+       struct meter_profile *mp, *mp_old;
+       struct metarray *metarray;
+       struct metarray_runtime *metarray_runtime;
+       struct meter *m;
+
+       CHECK(p, EINVAL);
+       CHECK_NAME(metarray_name, EINVAL);
+
+       metarray = metarray_find(p, metarray_name);
+       CHECK(metarray, EINVAL);
+       CHECK(metarray_index < metarray->size, EINVAL);
+
+       mp = meter_profile_find(p, profile_name);
+       CHECK(mp, EINVAL);
+
+       metarray_runtime = &p->metarray_runtime[metarray->id];
+       m = &metarray_runtime->metarray[metarray_index];
+       mp_old = m->profile;
+
+       memset(m, 0, sizeof(struct meter));
+       rte_meter_trtcm_config(&m->m, &mp->profile);
+       m->profile = mp;
+       m->color_mask = RTE_COLORS;
+
+       mp->n_users++;
+       mp_old->n_users--;
+
+       return 0;
+}
+
+int
+rte_swx_ctl_meter_stats_read(struct rte_swx_pipeline *p,
+                            const char *metarray_name,
+                            uint32_t metarray_index,
+                            struct rte_swx_ctl_meter_stats *stats)
+{
+       struct metarray *metarray;
+       struct metarray_runtime *metarray_runtime;
+       struct meter *m;
+
+       CHECK(p, EINVAL);
+       CHECK_NAME(metarray_name, EINVAL);
+
+       metarray = metarray_find(p, metarray_name);
+       CHECK(metarray, EINVAL);
+       CHECK(metarray_index < metarray->size, EINVAL);
+
+       CHECK(stats, EINVAL);
+
+       metarray_runtime = &p->metarray_runtime[metarray->id];
+       m = &metarray_runtime->metarray[metarray_index];
+
+       memcpy(stats->n_pkts, m->n_pkts, sizeof(m->n_pkts));
+       memcpy(stats->n_bytes, m->n_bytes, sizeof(m->n_bytes));
+
+       return 0;
+}