pipeline: add SWX instruction optimizer
[dpdk.git] / lib / librte_pipeline / rte_swx_pipeline.c
index 19bf276..77eae19 100644 (file)
@@ -6,9 +6,11 @@
 #include <stdio.h>
 #include <errno.h>
 #include <sys/queue.h>
+#include <arpa/inet.h>
 
 #include <rte_common.h>
 #include <rte_prefetch.h>
+#include <rte_byteorder.h>
 
 #include "rte_swx_pipeline.h"
 #include "rte_swx_ctl.h"
@@ -32,6 +34,9 @@ do {                                                                           \
 #define TRACE(...)
 #endif
 
+#define ntoh64(x) rte_be_to_cpu_64(x)
+#define hton64(x) rte_cpu_to_be_64(x)
+
 /*
  * Struct.
  */
@@ -236,6 +241,208 @@ enum instruction_type {
        INSTR_HDR_EMIT6_TX,
        INSTR_HDR_EMIT7_TX,
        INSTR_HDR_EMIT8_TX,
+
+       /* validate h.header */
+       INSTR_HDR_VALIDATE,
+
+       /* invalidate h.header */
+       INSTR_HDR_INVALIDATE,
+
+       /* mov dst src
+        * 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 */
+
+       /* dma h.header t.field
+        * memcpy(h.header, t.field, sizeof(h.header))
+        */
+       INSTR_DMA_HT,
+       INSTR_DMA_HT2,
+       INSTR_DMA_HT3,
+       INSTR_DMA_HT4,
+       INSTR_DMA_HT5,
+       INSTR_DMA_HT6,
+       INSTR_DMA_HT7,
+       INSTR_DMA_HT8,
+
+       /* add dst src
+        * dst += src
+        * dst = HMEF, src = HMEFTI
+        */
+       INSTR_ALU_ADD,    /* dst = MEF, src = MEF */
+       INSTR_ALU_ADD_MH, /* dst = MEF, src = H */
+       INSTR_ALU_ADD_HM, /* dst = H, src = MEF */
+       INSTR_ALU_ADD_HH, /* dst = H, src = H */
+       INSTR_ALU_ADD_MI, /* dst = MEF, src = I */
+       INSTR_ALU_ADD_HI, /* dst = H, src = I */
+
+       /* sub dst src
+        * dst -= src
+        * dst = HMEF, src = HMEFTI
+        */
+       INSTR_ALU_SUB,    /* dst = MEF, src = MEF */
+       INSTR_ALU_SUB_MH, /* dst = MEF, src = H */
+       INSTR_ALU_SUB_HM, /* dst = H, src = MEF */
+       INSTR_ALU_SUB_HH, /* dst = H, src = H */
+       INSTR_ALU_SUB_MI, /* dst = MEF, src = I */
+       INSTR_ALU_SUB_HI, /* dst = H, src = I */
+
+       /* ckadd dst src
+        * dst = dst '+ src[0:1] '+ src[2:3] + ...
+        * dst = H, src = {H, h.header}
+        */
+       INSTR_ALU_CKADD_FIELD,    /* src = H */
+       INSTR_ALU_CKADD_STRUCT20, /* src = h.header, with sizeof(header) = 20 */
+       INSTR_ALU_CKADD_STRUCT,   /* src = h.hdeader, with any sizeof(header) */
+
+       /* cksub dst src
+        * dst = dst '- src
+        * dst = H, src = H
+        */
+       INSTR_ALU_CKSUB_FIELD,
+
+       /* and dst src
+        * 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 */
+
+       /* 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 */
+
+       /* 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 */
+
+       /* shl dst src
+        * dst <<= src
+        * dst = HMEF, src = HMEFTI
+        */
+       INSTR_ALU_SHL,    /* dst = MEF, src = MEF */
+       INSTR_ALU_SHL_MH, /* dst = MEF, src = H */
+       INSTR_ALU_SHL_HM, /* dst = H, src = MEF */
+       INSTR_ALU_SHL_HH, /* dst = H, src = H */
+       INSTR_ALU_SHL_MI, /* dst = MEF, src = I */
+       INSTR_ALU_SHL_HI, /* dst = H, src = I */
+
+       /* shr dst src
+        * dst >>= src
+        * dst = HMEF, src = HMEFTI
+        */
+       INSTR_ALU_SHR,    /* dst = MEF, src = MEF */
+       INSTR_ALU_SHR_MH, /* dst = MEF, src = H */
+       INSTR_ALU_SHR_HM, /* dst = H, src = MEF */
+       INSTR_ALU_SHR_HH, /* dst = H, src = H */
+       INSTR_ALU_SHR_MI, /* dst = MEF, src = I */
+       INSTR_ALU_SHR_HI, /* dst = H, src = I */
+
+       /* table TABLE */
+       INSTR_TABLE,
+
+       /* extern e.obj.func */
+       INSTR_EXTERN_OBJ,
+
+       /* extern f.func */
+       INSTR_EXTERN_FUNC,
+
+       /* jmp LABEL
+        * Unconditional jump
+        */
+       INSTR_JMP,
+
+       /* jmpv LABEL h.header
+        * Jump if header is valid
+        */
+       INSTR_JMP_VALID,
+
+       /* jmpnv LABEL h.header
+        * Jump if header is invalid
+        */
+       INSTR_JMP_INVALID,
+
+       /* jmph LABEL
+        * Jump if table lookup hit
+        */
+       INSTR_JMP_HIT,
+
+       /* jmpnh LABEL
+        * Jump if table lookup miss
+        */
+       INSTR_JMP_MISS,
+
+       /* jmpa LABEL ACTION
+        * Jump if action run
+        */
+       INSTR_JMP_ACTION_HIT,
+
+       /* jmpna LABEL ACTION
+        * Jump if action not run
+        */
+       INSTR_JMP_ACTION_MISS,
+
+       /* jmpeq LABEL a b
+        * Jump is 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) */
+
+       /* jmpneq LABEL a b
+        * Jump is 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) */
+
+       /* 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_HH, /* a = H, b = H */
+       INSTR_JMP_LT_MI, /* a = MEF, 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_HH, /* a = H, b = H */
+       INSTR_JMP_GT_MI, /* a = MEF, b = I */
+       INSTR_JMP_GT_HI, /* a = H, b = I */
+
+       /* return
+        * Return from action
+        */
+       INSTR_RETURN,
+};
+
+struct instr_operand {
+       uint8_t struct_id;
+       uint8_t n_bits;
+       uint8_t offset;
+       uint8_t pad;
 };
 
 struct instr_io {
@@ -252,10 +459,71 @@ struct instr_io {
        } hdr;
 };
 
+struct instr_hdr_validity {
+       uint8_t header_id;
+};
+
+struct instr_table {
+       uint8_t table_id;
+};
+
+struct instr_extern_obj {
+       uint8_t ext_obj_id;
+       uint8_t func_id;
+};
+
+struct instr_extern_func {
+       uint8_t ext_func_id;
+};
+
+struct instr_dst_src {
+       struct instr_operand dst;
+       union {
+               struct instr_operand src;
+               uint32_t src_val;
+       };
+};
+
+struct instr_dma {
+       struct {
+               uint8_t header_id[8];
+               uint8_t struct_id[8];
+       } dst;
+
+       struct {
+               uint8_t offset[8];
+       } src;
+
+       uint16_t n_bytes[8];
+};
+
+struct instr_jmp {
+       struct instruction *ip;
+
+       union {
+               struct instr_operand a;
+               uint8_t header_id;
+               uint8_t action_id;
+       };
+
+       union {
+               struct instr_operand b;
+               uint32_t b_val;
+       };
+};
+
 struct instruction {
        enum instruction_type type;
        union {
                struct instr_io io;
+               struct instr_hdr_validity valid;
+               struct instr_dst_src mov;
+               struct instr_dma dma;
+               struct instr_dst_src alu;
+               struct instr_table table;
+               struct instr_extern_obj ext_obj;
+               struct instr_extern_func ext_func;
+               struct instr_jmp jmp;
        };
 };
 
@@ -370,6 +638,302 @@ struct thread {
 #define MASK64_BIT_SET(mask, pos) ((mask) | (1LLU << (pos)))
 #define MASK64_BIT_CLR(mask, pos) ((mask) & ~(1LLU << (pos)))
 
+#define HEADER_VALID(thread, header_id) \
+       MASK64_BIT_GET((thread)->valid_headers, header_id)
+
+#define ALU(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 = src64 & src64_mask;                                     \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | (result & dst64_mask);            \
+}
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+#define ALU_S(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 = ntoh64(src64) >> (64 - (ip)->alu.src.n_bits);           \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+                                                                              \
+       *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];      \
+       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 = ntoh64(dst64) >> (64 - (ip)->alu.dst.n_bits);           \
+                                                                              \
+       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 = src64 & src64_mask;                                     \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+       result = hton64(result << (64 - (ip)->alu.dst.n_bits));                \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | result;                           \
+}
+
+#define ALU_HH(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 = ntoh64(dst64) >> (64 - (ip)->alu.dst.n_bits);           \
+                                                                              \
+       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 = ntoh64(src64) >> (64 - (ip)->alu.src.n_bits);           \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+       result = hton64(result << (64 - (ip)->alu.dst.n_bits));                \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | result;                           \
+}
+
+#else
+
+#define ALU_S ALU
+#define ALU_MH ALU
+#define ALU_HM ALU
+#define ALU_HH ALU
+
+#endif
+
+#define ALU_I(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;                                     \
+                                                                              \
+       uint64_t src = (ip)->alu.src_val;                                      \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | (result & dst64_mask);            \
+}
+
+#define ALU_MI ALU_I
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+#define ALU_HI(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 = ntoh64(dst64) >> (64 - (ip)->alu.dst.n_bits);           \
+                                                                              \
+       uint64_t src = (ip)->alu.src_val;                                      \
+                                                                              \
+       uint64_t result = dst operator src;                                    \
+       result = hton64(result << (64 - (ip)->alu.dst.n_bits));                \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | result;                           \
+}
+
+#else
+
+#define ALU_HI ALU_I
+
+#endif
+
+#define MOV(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;                                     \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | (src & dst64_mask);               \
+}
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+#define MOV_S(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 = ntoh64(src64) >> (64 - (ip)->mov.src.n_bits);           \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | (src & dst64_mask);               \
+}
+
+#else
+
+#define MOV_S MOV
+
+#endif
+
+#define MOV_I(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);       \
+                                                                              \
+       uint64_t src = (ip)->mov.src_val;                                      \
+                                                                              \
+       *dst64_ptr = (dst64 & ~dst64_mask) | (src & dst64_mask);               \
+}
+
+#define JMP_CMP(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 a64_mask = UINT64_MAX >> (64 - (ip)->jmp.a.n_bits);           \
+       uint64_t a = a64 & a64_mask;                                           \
+                                                                              \
+       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 b64_mask = UINT64_MAX >> (64 - (ip)->jmp.b.n_bits);           \
+       uint64_t b = b64 & b64_mask;                                           \
+                                                                              \
+       (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
+}
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+#define JMP_CMP_S(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 a64_mask = UINT64_MAX >> (64 - (ip)->jmp.a.n_bits);           \
+       uint64_t a = a64 & a64_mask;                                           \
+                                                                              \
+       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 = ntoh64(b64) >> (64 - (ip)->jmp.b.n_bits);                 \
+                                                                              \
+       (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];          \
+       uint64_t *a64_ptr = (uint64_t *)&a_struct[(ip)->jmp.a.offset];         \
+       uint64_t a64 = *a64_ptr;                                               \
+       uint64_t a = ntoh64(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 b64_mask = UINT64_MAX >> (64 - (ip)->jmp.b.n_bits);           \
+       uint64_t b = b64 & b64_mask;                                           \
+                                                                              \
+       (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
+}
+
+#define JMP_CMP_HH(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 = ntoh64(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 = ntoh64(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
+
+#endif
+
+#define JMP_CMP_I(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 a64_mask = UINT64_MAX >> (64 - (ip)->jmp.a.n_bits);           \
+       uint64_t a = a64 & a64_mask;                                           \
+                                                                              \
+       uint64_t b = (ip)->jmp.b_val;                                          \
+                                                                              \
+       (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
+}
+
+#define JMP_CMP_MI JMP_CMP_I
+
+#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+
+#define JMP_CMP_HI(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 = ntoh64(a64) >> (64 - (ip)->jmp.a.n_bits);                 \
+                                                                              \
+       uint64_t b = (ip)->jmp.b_val;                                          \
+                                                                              \
+       (thread)->ip = (a operator b) ? (ip)->jmp.ip : ((thread)->ip + 1);     \
+}
+
+#else
+
+#define JMP_CMP_HI JMP_CMP_I
+
+#endif
+
 #define METADATA_READ(thread, offset, n_bits)                                  \
 ({                                                                             \
        uint64_t *m64_ptr = (uint64_t *)&(thread)->metadata[offset];           \
@@ -933,6 +1497,94 @@ extern_obj_find(struct rte_swx_pipeline *p, const char *name)
        return NULL;
 }
 
+static struct extern_type_member_func *
+extern_obj_member_func_parse(struct rte_swx_pipeline *p,
+                            const char *name,
+                            struct extern_obj **obj)
+{
+       struct extern_obj *object;
+       struct extern_type_member_func *func;
+       char *object_name, *func_name;
+
+       if (name[0] != 'e' || name[1] != '.')
+               return NULL;
+
+       object_name = strdup(&name[2]);
+       if (!object_name)
+               return NULL;
+
+       func_name = strchr(object_name, '.');
+       if (!func_name) {
+               free(object_name);
+               return NULL;
+       }
+
+       *func_name = 0;
+       func_name++;
+
+       object = extern_obj_find(p, object_name);
+       if (!object) {
+               free(object_name);
+               return NULL;
+       }
+
+       func = extern_type_member_func_find(object->type, func_name);
+       if (!func) {
+               free(object_name);
+               return NULL;
+       }
+
+       if (obj)
+               *obj = object;
+
+       free(object_name);
+       return func;
+}
+
+static struct field *
+extern_obj_mailbox_field_parse(struct rte_swx_pipeline *p,
+                              const char *name,
+                              struct extern_obj **object)
+{
+       struct extern_obj *obj;
+       struct field *f;
+       char *obj_name, *field_name;
+
+       if ((name[0] != 'e') || (name[1] != '.'))
+               return NULL;
+
+       obj_name = strdup(&name[2]);
+       if (!obj_name)
+               return NULL;
+
+       field_name = strchr(obj_name, '.');
+       if (!field_name) {
+               free(obj_name);
+               return NULL;
+       }
+
+       *field_name = 0;
+       field_name++;
+
+       obj = extern_obj_find(p, obj_name);
+       if (!obj) {
+               free(obj_name);
+               return NULL;
+       }
+
+       f = struct_type_field_find(obj->type->mailbox_struct_type, field_name);
+       if (!f) {
+               free(obj_name);
+               return NULL;
+       }
+
+       if (object)
+               *object = obj;
+
+       free(obj_name);
+       return f;
+}
+
 int
 rte_swx_pipeline_extern_type_register(struct rte_swx_pipeline *p,
        const char *name,
@@ -1171,21 +1823,75 @@ extern_func_find(struct rte_swx_pipeline *p, const char *name)
        return NULL;
 }
 
-int
-rte_swx_pipeline_extern_func_register(struct rte_swx_pipeline *p,
-                                     const char *name,
-                                     const char *mailbox_struct_type_name,
-                                     rte_swx_extern_func_t func)
+static struct extern_func *
+extern_func_parse(struct rte_swx_pipeline *p,
+                 const char *name)
 {
-       struct extern_func *f;
-       struct struct_type *mailbox_struct_type;
-
-       CHECK(p, EINVAL);
+       if (name[0] != 'f' || name[1] != '.')
+               return NULL;
 
-       CHECK_NAME(name, EINVAL);
-       CHECK(!extern_func_find(p, name), EEXIST);
+       return extern_func_find(p, &name[2]);
+}
 
-       CHECK_NAME(mailbox_struct_type_name, EINVAL);
+static struct field *
+extern_func_mailbox_field_parse(struct rte_swx_pipeline *p,
+                               const char *name,
+                               struct extern_func **function)
+{
+       struct extern_func *func;
+       struct field *f;
+       char *func_name, *field_name;
+
+       if ((name[0] != 'f') || (name[1] != '.'))
+               return NULL;
+
+       func_name = strdup(&name[2]);
+       if (!func_name)
+               return NULL;
+
+       field_name = strchr(func_name, '.');
+       if (!field_name) {
+               free(func_name);
+               return NULL;
+       }
+
+       *field_name = 0;
+       field_name++;
+
+       func = extern_func_find(p, func_name);
+       if (!func) {
+               free(func_name);
+               return NULL;
+       }
+
+       f = struct_type_field_find(func->mailbox_struct_type, field_name);
+       if (!f) {
+               free(func_name);
+               return NULL;
+       }
+
+       if (function)
+               *function = func;
+
+       free(func_name);
+       return f;
+}
+
+int
+rte_swx_pipeline_extern_func_register(struct rte_swx_pipeline *p,
+                                     const char *name,
+                                     const char *mailbox_struct_type_name,
+                                     rte_swx_extern_func_t func)
+{
+       struct extern_func *f;
+       struct struct_type *mailbox_struct_type;
+
+       CHECK(p, EINVAL);
+
+       CHECK_NAME(name, EINVAL);
+       CHECK(!extern_func_find(p, name), EEXIST);
+
+       CHECK_NAME(mailbox_struct_type_name, EINVAL);
        mailbox_struct_type = struct_type_find(p, mailbox_struct_type_name);
        CHECK(mailbox_struct_type, EINVAL);
 
@@ -1551,6 +2257,118 @@ metadata_free(struct rte_swx_pipeline *p)
 /*
  * Instruction.
  */
+static int
+instruction_is_jmp(struct instruction *instr)
+{
+       switch (instr->type) {
+       case INSTR_JMP:
+       case INSTR_JMP_VALID:
+       case INSTR_JMP_INVALID:
+       case INSTR_JMP_HIT:
+       case INSTR_JMP_MISS:
+       case INSTR_JMP_ACTION_HIT:
+       case INSTR_JMP_ACTION_MISS:
+       case INSTR_JMP_EQ:
+       case INSTR_JMP_EQ_S:
+       case INSTR_JMP_EQ_I:
+       case INSTR_JMP_NEQ:
+       case INSTR_JMP_NEQ_S:
+       case INSTR_JMP_NEQ_I:
+       case INSTR_JMP_LT:
+       case INSTR_JMP_LT_MH:
+       case INSTR_JMP_LT_HM:
+       case INSTR_JMP_LT_HH:
+       case INSTR_JMP_LT_MI:
+       case INSTR_JMP_LT_HI:
+       case INSTR_JMP_GT:
+       case INSTR_JMP_GT_MH:
+       case INSTR_JMP_GT_HM:
+       case INSTR_JMP_GT_HH:
+       case INSTR_JMP_GT_MI:
+       case INSTR_JMP_GT_HI:
+               return 1;
+
+       default:
+               return 0;
+       }
+}
+
+static struct field *
+action_field_parse(struct action *action, const char *name);
+
+static struct field *
+struct_field_parse(struct rte_swx_pipeline *p,
+                  struct action *action,
+                  const char *name,
+                  uint32_t *struct_id)
+{
+       struct field *f;
+
+       switch (name[0]) {
+       case 'h':
+       {
+               struct header *header;
+
+               f = header_field_parse(p, name, &header);
+               if (!f)
+                       return NULL;
+
+               *struct_id = header->struct_id;
+               return f;
+       }
+
+       case 'm':
+       {
+               f = metadata_field_parse(p, name);
+               if (!f)
+                       return NULL;
+
+               *struct_id = p->metadata_struct_id;
+               return f;
+       }
+
+       case 't':
+       {
+               if (!action)
+                       return NULL;
+
+               f = action_field_parse(action, name);
+               if (!f)
+                       return NULL;
+
+               *struct_id = 0;
+               return f;
+       }
+
+       case 'e':
+       {
+               struct extern_obj *obj;
+
+               f = extern_obj_mailbox_field_parse(p, name, &obj);
+               if (!f)
+                       return NULL;
+
+               *struct_id = obj->struct_id;
+               return f;
+       }
+
+       case 'f':
+       {
+               struct extern_func *func;
+
+               f = extern_func_mailbox_field_parse(p, name, &func);
+               if (!f)
+                       return NULL;
+
+               *struct_id = func->struct_id;
+               return f;
+       }
+
+       default:
+               return NULL;
+       }
+}
+
 static inline void
 pipeline_port_inc(struct rte_swx_pipeline *p)
 {
@@ -1563,6 +2381,21 @@ thread_ip_reset(struct rte_swx_pipeline *p, struct thread *t)
        t->ip = p->instructions;
 }
 
+static inline void
+thread_ip_set(struct thread *t, struct instruction *ip)
+{
+       t->ip = ip;
+}
+
+static inline void
+thread_ip_action_call(struct rte_swx_pipeline *p,
+                     struct thread *t,
+                     uint32_t action_id)
+{
+       t->ret = t->ip + 1;
+       t->ip = p->action_instructions[action_id];
+}
+
 static inline void
 thread_ip_inc(struct rte_swx_pipeline *p);
 
@@ -1586,6 +2419,12 @@ thread_yield(struct rte_swx_pipeline *p)
        p->thread_id = (p->thread_id + 1) & (RTE_SWX_PIPELINE_THREADS_MAX - 1);
 }
 
+static inline void
+thread_yield_cond(struct rte_swx_pipeline *p, int cond)
+{
+       p->thread_id = (p->thread_id + cond) & (RTE_SWX_PIPELINE_THREADS_MAX - 1);
+}
+
 /*
  * rx.
  */
@@ -2098,150 +2937,3018 @@ instr_hdr_emit8_tx_exec(struct rte_swx_pipeline *p)
        instr_tx_exec(p);
 }
 
-#define RTE_SWX_INSTRUCTION_TOKENS_MAX 16
-
+/*
+ * validate.
+ */
 static int
-instr_translate(struct rte_swx_pipeline *p,
-               struct action *action,
-               char *string,
-               struct instruction *instr,
-               struct instruction_data *data)
+instr_hdr_validate_translate(struct rte_swx_pipeline *p,
+                            struct action *action __rte_unused,
+                            char **tokens,
+                            int n_tokens,
+                            struct instruction *instr,
+                            struct instruction_data *data __rte_unused)
 {
-       char *tokens[RTE_SWX_INSTRUCTION_TOKENS_MAX];
-       int n_tokens = 0, tpos = 0;
-
-       /* Parse the instruction string into tokens. */
-       for ( ; ; ) {
-               char *token;
+       struct header *h;
 
-               token = strtok_r(string, " \t\v", &string);
-               if (!token)
-                       break;
+       CHECK(n_tokens == 2, EINVAL);
 
-               CHECK(n_tokens < RTE_SWX_INSTRUCTION_TOKENS_MAX, EINVAL);
+       h = header_parse(p, tokens[1]);
+       CHECK(h, EINVAL);
 
-               tokens[n_tokens] = token;
-               n_tokens++;
-       }
+       instr->type = INSTR_HDR_VALIDATE;
+       instr->valid.header_id = h->id;
+       return 0;
+}
 
-       CHECK(n_tokens, EINVAL);
+static inline void
+instr_hdr_validate_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t header_id = ip->valid.header_id;
 
-       /* Handle the optional instruction label. */
-       if ((n_tokens >= 2) && !strcmp(tokens[1], ":")) {
-               strcpy(data->label, tokens[0]);
+       TRACE("[Thread %2u] validate header %u\n", p->thread_id, header_id);
 
-               tpos += 2;
-               CHECK(n_tokens - tpos, EINVAL);
-       }
+       /* Headers. */
+       t->valid_headers = MASK64_BIT_SET(t->valid_headers, header_id);
 
-       /* Identify the instruction type. */
-       if (!strcmp(tokens[tpos], "rx"))
-               return instr_rx_translate(p,
-                                         action,
-                                         &tokens[tpos],
-                                         n_tokens - tpos,
-                                         instr,
-                                         data);
+       /* Thread. */
+       thread_ip_inc(p);
+}
 
-       if (!strcmp(tokens[tpos], "tx"))
-               return instr_tx_translate(p,
-                                         action,
-                                         &tokens[tpos],
-                                         n_tokens - tpos,
-                                         instr,
-                                         data);
+/*
+ * invalidate.
+ */
+static int
+instr_hdr_invalidate_translate(struct rte_swx_pipeline *p,
+                              struct action *action __rte_unused,
+                              char **tokens,
+                              int n_tokens,
+                              struct instruction *instr,
+                              struct instruction_data *data __rte_unused)
+{
+       struct header *h;
 
-       if (!strcmp(tokens[tpos], "extract"))
-               return instr_hdr_extract_translate(p,
-                                                  action,
-                                                  &tokens[tpos],
-                                                  n_tokens - tpos,
-                                                  instr,
-                                                  data);
+       CHECK(n_tokens == 2, EINVAL);
 
-       if (!strcmp(tokens[tpos], "emit"))
-               return instr_hdr_emit_translate(p,
-                                               action,
-                                               &tokens[tpos],
-                                               n_tokens - tpos,
-                                               instr,
-                                               data);
+       h = header_parse(p, tokens[1]);
+       CHECK(h, EINVAL);
 
-       CHECK(0, EINVAL);
+       instr->type = INSTR_HDR_INVALIDATE;
+       instr->valid.header_id = h->id;
+       return 0;
 }
 
-static uint32_t
-label_is_used(struct instruction_data *data, uint32_t n, const char *label)
+static inline void
+instr_hdr_invalidate_exec(struct rte_swx_pipeline *p)
 {
-       uint32_t count = 0, i;
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t header_id = ip->valid.header_id;
 
-       if (!label[0])
-               return 0;
+       TRACE("[Thread %2u] invalidate header %u\n", p->thread_id, header_id);
 
-       for (i = 0; i < n; i++)
-               if (!strcmp(label, data[i].jmp_label))
-                       count++;
+       /* Headers. */
+       t->valid_headers = MASK64_BIT_CLR(t->valid_headers, header_id);
 
-       return count;
+       /* Thread. */
+       thread_ip_inc(p);
 }
 
+/*
+ * table.
+ */
+static struct table *
+table_find(struct rte_swx_pipeline *p, const char *name);
+
 static int
-instr_label_check(struct instruction_data *instruction_data,
-                 uint32_t n_instructions)
+instr_table_translate(struct rte_swx_pipeline *p,
+                     struct action *action,
+                     char **tokens,
+                     int n_tokens,
+                     struct instruction *instr,
+                     struct instruction_data *data __rte_unused)
 {
-       uint32_t i;
+       struct table *t;
 
-       /* Check that all instruction labels are unique. */
-       for (i = 0; i < n_instructions; i++) {
-               struct instruction_data *data = &instruction_data[i];
-               char *label = data->label;
-               uint32_t j;
+       CHECK(!action, EINVAL);
+       CHECK(n_tokens == 2, EINVAL);
 
-               if (!label[0])
-                       continue;
+       t = table_find(p, tokens[1]);
+       CHECK(t, EINVAL);
 
-               for (j = i + 1; j < n_instructions; j++)
-                       CHECK(strcmp(label, data[j].label), EINVAL);
-       }
+       instr->type = INSTR_TABLE;
+       instr->table.table_id = t->id;
+       return 0;
+}
 
-       /* Get users for each instruction label. */
-       for (i = 0; i < n_instructions; i++) {
-               struct instruction_data *data = &instruction_data[i];
-               char *label = data->label;
+static inline void
+instr_table_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       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;
+       uint8_t *action_data;
+       int done, hit;
+
+       /* Table. */
+       done = table->func(ts->obj,
+                          table->mailbox,
+                          table->key,
+                          &action_id,
+                          &action_data,
+                          &hit);
+       if (!done) {
+               /* Thread. */
+               TRACE("[Thread %2u] table %u (not finalized)\n",
+                     p->thread_id,
+                     table_id);
 
-               data->n_users = label_is_used(instruction_data,
-                                             n_instructions,
-                                             label);
+               thread_yield(p);
+               return;
        }
 
-       return 0;
+       action_id = hit ? action_id : ts->default_action_id;
+       action_data = hit ? action_data : ts->default_action_data;
+
+       TRACE("[Thread %2u] table %u (%s, action %u)\n",
+             p->thread_id,
+             table_id,
+             hit ? "hit" : "miss",
+             (uint32_t)action_id);
+
+       t->action_id = action_id;
+       t->structs[0] = action_data;
+       t->hit = hit;
+
+       /* Thread. */
+       thread_ip_action_call(p, t, action_id);
 }
 
+/*
+ * extern.
+ */
 static int
-instruction_config(struct rte_swx_pipeline *p,
-                  struct action *a,
-                  const char **instructions,
-                  uint32_t n_instructions)
+instr_extern_translate(struct rte_swx_pipeline *p,
+                      struct action *action __rte_unused,
+                      char **tokens,
+                      int n_tokens,
+                      struct instruction *instr,
+                      struct instruction_data *data __rte_unused)
 {
-       struct instruction *instr = NULL;
-       struct instruction_data *data = NULL;
-       char *string = NULL;
-       int err = 0;
-       uint32_t i;
+       char *token = tokens[1];
 
-       CHECK(n_instructions, EINVAL);
-       CHECK(instructions, EINVAL);
-       for (i = 0; i < n_instructions; i++)
-               CHECK(instructions[i], EINVAL);
+       CHECK(n_tokens == 2, EINVAL);
 
-       /* Memory allocation. */
-       instr = calloc(n_instructions, sizeof(struct instruction));
-       if (!instr) {
-               err = ENOMEM;
-               goto error;
-       }
+       if (token[0] == 'e') {
+               struct extern_obj *obj;
+               struct extern_type_member_func *func;
 
-       data = calloc(n_instructions, sizeof(struct instruction_data));
+               func = extern_obj_member_func_parse(p, token, &obj);
+               CHECK(func, EINVAL);
+
+               instr->type = INSTR_EXTERN_OBJ;
+               instr->ext_obj.ext_obj_id = obj->id;
+               instr->ext_obj.func_id = func->id;
+
+               return 0;
+       }
+
+       if (token[0] == 'f') {
+               struct extern_func *func;
+
+               func = extern_func_parse(p, token);
+               CHECK(func, EINVAL);
+
+               instr->type = INSTR_EXTERN_FUNC;
+               instr->ext_func.ext_func_id = func->id;
+
+               return 0;
+       }
+
+       CHECK(0, EINVAL);
+}
+
+static inline void
+instr_extern_obj_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t obj_id = ip->ext_obj.ext_obj_id;
+       uint32_t func_id = ip->ext_obj.func_id;
+       struct extern_obj_runtime *obj = &t->extern_objs[obj_id];
+       rte_swx_extern_type_member_func_t func = obj->funcs[func_id];
+
+       TRACE("[Thread %2u] extern obj %u member func %u\n",
+             p->thread_id,
+             obj_id,
+             func_id);
+
+       /* Extern object member function execute. */
+       uint32_t done = func(obj->obj, obj->mailbox);
+
+       /* Thread. */
+       thread_ip_inc_cond(t, done);
+       thread_yield_cond(p, done ^ 1);
+}
+
+static inline void
+instr_extern_func_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t ext_func_id = ip->ext_func.ext_func_id;
+       struct extern_func_runtime *ext_func = &t->extern_funcs[ext_func_id];
+       rte_swx_extern_func_t func = ext_func->func;
+
+       TRACE("[Thread %2u] extern func %u\n",
+             p->thread_id,
+             ext_func_id);
+
+       /* Extern function execute. */
+       uint32_t done = func(ext_func->mailbox);
+
+       /* Thread. */
+       thread_ip_inc_cond(t, done);
+       thread_yield_cond(p, done ^ 1);
+}
+
+/*
+ * mov.
+ */
+static int
+instr_mov_translate(struct rte_swx_pipeline *p,
+                   struct action *action,
+                   char **tokens,
+                   int n_tokens,
+                   struct instruction *instr,
+                   struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* MOV or MOV_S. */
+       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;
+
+               instr->mov.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->mov.dst.n_bits = fdst->n_bits;
+               instr->mov.dst.offset = fdst->offset / 8;
+               instr->mov.src.struct_id = (uint8_t)src_struct_id;
+               instr->mov.src.n_bits = fsrc->n_bits;
+               instr->mov.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* MOV_I. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       if (dst[0] == 'h')
+               src_val = htonl(src_val);
+
+       instr->type = INSTR_MOV_I;
+       instr->mov.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->mov.dst.n_bits = fdst->n_bits;
+       instr->mov.dst.offset = fdst->offset / 8;
+       instr->mov.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static inline void
+instr_mov_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] mov\n",
+             p->thread_id);
+
+       MOV(t, ip);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_mov_s_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",
+             p->thread_id);
+
+       MOV_S(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;
+
+       TRACE("[Thread %2u] mov m.f %x\n",
+             p->thread_id,
+             ip->mov.src_val);
+
+       MOV_I(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)
+{
+       char *dst = tokens[1];
+       char *src = tokens[2];
+       struct header *h;
+       struct field *tf;
+
+       CHECK(action, EINVAL);
+       CHECK(n_tokens == 3, EINVAL);
+
+       h = header_parse(p, dst);
+       CHECK(h, EINVAL);
+
+       tf = action_field_parse(action, src);
+       CHECK(tf, EINVAL);
+
+       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;
+
+       return 0;
+}
+
+static inline void
+__instr_dma_ht_exec(struct rte_swx_pipeline *p, uint32_t n_dma);
+
+static inline void
+__instr_dma_ht_exec(struct rte_swx_pipeline *p, uint32_t n_dma)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint8_t *action_data = t->structs[0];
+       uint64_t valid_headers = t->valid_headers;
+       uint32_t i;
+
+       for (i = 0; i < n_dma; i++) {
+               uint32_t header_id = ip->dma.dst.header_id[i];
+               uint32_t struct_id = ip->dma.dst.struct_id[i];
+               uint32_t offset = ip->dma.src.offset[i];
+               uint32_t n_bytes = ip->dma.n_bytes[i];
+
+               struct header_runtime *h = &t->headers[header_id];
+               uint8_t *h_ptr0 = h->ptr0;
+               uint8_t *h_ptr = t->structs[struct_id];
+
+               void *dst = MASK64_BIT_GET(valid_headers, header_id) ?
+                       h_ptr : h_ptr0;
+               void *src = &action_data[offset];
+
+               TRACE("[Thread %2u] dma h.s t.f\n", p->thread_id);
+
+               /* Headers. */
+               memcpy(dst, src, n_bytes);
+               t->structs[struct_id] = dst;
+               valid_headers = MASK64_BIT_SET(valid_headers, header_id);
+       }
+
+       t->valid_headers = valid_headers;
+}
+
+static inline void
+instr_dma_ht_exec(struct rte_swx_pipeline *p)
+{
+       __instr_dma_ht_exec(p, 1);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht2_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 2 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 2);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht3_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 3 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 3);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht4_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 4 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 4);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht5_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 5 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 5);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht6_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 6 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 6);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht7_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 7 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 7);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_dma_ht8_exec(struct rte_swx_pipeline *p)
+{
+       TRACE("[Thread %2u] *** The next 8 instructions are fused. ***\n",
+             p->thread_id);
+
+       __instr_dma_ht_exec(p, 8);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+/*
+ * alu.
+ */
+static int
+instr_alu_add_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* ADD, ADD_HM, ADD_MH, ADD_HH. */
+       fsrc = struct_field_parse(p, action, src, &src_struct_id);
+       if (fsrc) {
+               instr->type = INSTR_ALU_ADD;
+               if (dst[0] == 'h' && src[0] == 'm')
+                       instr->type = INSTR_ALU_ADD_HM;
+               if (dst[0] == 'm' && src[0] == 'h')
+                       instr->type = INSTR_ALU_ADD_MH;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_ADD_HH;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* ADD_MI, ADD_HI. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       instr->type = INSTR_ALU_ADD_MI;
+       if (dst[0] == 'h')
+               instr->type = INSTR_ALU_ADD_HI;
+
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_sub_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* SUB, SUB_HM, SUB_MH, SUB_HH. */
+       fsrc = struct_field_parse(p, action, src, &src_struct_id);
+       if (fsrc) {
+               instr->type = INSTR_ALU_SUB;
+               if (dst[0] == 'h' && src[0] == 'm')
+                       instr->type = INSTR_ALU_SUB_HM;
+               if (dst[0] == 'm' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SUB_MH;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SUB_HH;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* SUB_MI, SUB_HI. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       instr->type = INSTR_ALU_SUB_MI;
+       if (dst[0] == 'h')
+               instr->type = INSTR_ALU_SUB_HI;
+
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_ckadd_translate(struct rte_swx_pipeline *p,
+                         struct action *action __rte_unused,
+                         char **tokens,
+                         int n_tokens,
+                         struct instruction *instr,
+                         struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct header *hdst, *hsrc;
+       struct field *fdst, *fsrc;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = header_field_parse(p, dst, &hdst);
+       CHECK(fdst && (fdst->n_bits == 16), EINVAL);
+
+       /* CKADD_FIELD. */
+       fsrc = header_field_parse(p, src, &hsrc);
+       if (fsrc) {
+               instr->type = INSTR_ALU_CKADD_FIELD;
+               instr->alu.dst.struct_id = (uint8_t)hdst->struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)hsrc->struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* CKADD_STRUCT, CKADD_STRUCT20. */
+       hsrc = header_parse(p, src);
+       CHECK(hsrc, EINVAL);
+
+       instr->type = INSTR_ALU_CKADD_STRUCT;
+       if ((hsrc->st->n_bits / 8) == 20)
+               instr->type = INSTR_ALU_CKADD_STRUCT20;
+
+       instr->alu.dst.struct_id = (uint8_t)hdst->struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src.struct_id = (uint8_t)hsrc->struct_id;
+       instr->alu.src.n_bits = hsrc->st->n_bits;
+       instr->alu.src.offset = 0; /* Unused. */
+       return 0;
+}
+
+static int
+instr_alu_cksub_translate(struct rte_swx_pipeline *p,
+                         struct action *action __rte_unused,
+                         char **tokens,
+                         int n_tokens,
+                         struct instruction *instr,
+                         struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct header *hdst, *hsrc;
+       struct field *fdst, *fsrc;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = header_field_parse(p, dst, &hdst);
+       CHECK(fdst && (fdst->n_bits == 16), EINVAL);
+
+       fsrc = header_field_parse(p, src, &hsrc);
+       CHECK(fsrc, EINVAL);
+
+       instr->type = INSTR_ALU_CKSUB_FIELD;
+       instr->alu.dst.struct_id = (uint8_t)hdst->struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src.struct_id = (uint8_t)hsrc->struct_id;
+       instr->alu.src.n_bits = fsrc->n_bits;
+       instr->alu.src.offset = fsrc->offset / 8;
+       return 0;
+}
+
+static int
+instr_alu_shl_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* SHL, SHL_HM, SHL_MH, SHL_HH. */
+       fsrc = struct_field_parse(p, action, src, &src_struct_id);
+       if (fsrc) {
+               instr->type = INSTR_ALU_SHL;
+               if (dst[0] == 'h' && src[0] == 'm')
+                       instr->type = INSTR_ALU_SHL_HM;
+               if (dst[0] == 'm' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SHL_MH;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SHL_HH;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* SHL_MI, SHL_HI. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       instr->type = INSTR_ALU_SHL_MI;
+       if (dst[0] == 'h')
+               instr->type = INSTR_ALU_SHL_HI;
+
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_shr_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* SHR, SHR_HM, SHR_MH, SHR_HH. */
+       fsrc = struct_field_parse(p, action, src, &src_struct_id);
+       if (fsrc) {
+               instr->type = INSTR_ALU_SHR;
+               if (dst[0] == 'h' && src[0] == 'm')
+                       instr->type = INSTR_ALU_SHR_HM;
+               if (dst[0] == 'm' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SHR_MH;
+               if (dst[0] == 'h' && src[0] == 'h')
+                       instr->type = INSTR_ALU_SHR_HH;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* SHR_MI, SHR_HI. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       instr->type = INSTR_ALU_SHR_MI;
+       if (dst[0] == 'h')
+               instr->type = INSTR_ALU_SHR_HI;
+
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_and_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* AND or AND_S. */
+       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;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* AND_I. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       if (dst[0] == 'h')
+               src_val = htonl(src_val);
+
+       instr->type = INSTR_ALU_AND_I;
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_or_translate(struct rte_swx_pipeline *p,
+                      struct action *action,
+                      char **tokens,
+                      int n_tokens,
+                      struct instruction *instr,
+                      struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* OR or OR_S. */
+       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;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* OR_I. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       if (dst[0] == 'h')
+               src_val = htonl(src_val);
+
+       instr->type = INSTR_ALU_OR_I;
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static int
+instr_alu_xor_translate(struct rte_swx_pipeline *p,
+                       struct action *action,
+                       char **tokens,
+                       int n_tokens,
+                       struct instruction *instr,
+                       struct instruction_data *data __rte_unused)
+{
+       char *dst = tokens[1], *src = tokens[2];
+       struct field *fdst, *fsrc;
+       uint32_t dst_struct_id, src_struct_id, src_val;
+
+       CHECK(n_tokens == 3, EINVAL);
+
+       fdst = struct_field_parse(p, NULL, dst, &dst_struct_id);
+       CHECK(fdst, EINVAL);
+
+       /* XOR or XOR_S. */
+       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;
+
+               instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+               instr->alu.dst.n_bits = fdst->n_bits;
+               instr->alu.dst.offset = fdst->offset / 8;
+               instr->alu.src.struct_id = (uint8_t)src_struct_id;
+               instr->alu.src.n_bits = fsrc->n_bits;
+               instr->alu.src.offset = fsrc->offset / 8;
+               return 0;
+       }
+
+       /* XOR_I. */
+       src_val = strtoul(src, &src, 0);
+       CHECK(!src[0], EINVAL);
+
+       if (dst[0] == 'h')
+               src_val = htonl(src_val);
+
+       instr->type = INSTR_ALU_XOR_I;
+       instr->alu.dst.struct_id = (uint8_t)dst_struct_id;
+       instr->alu.dst.n_bits = fdst->n_bits;
+       instr->alu.dst.offset = fdst->offset / 8;
+       instr->alu.src_val = (uint32_t)src_val;
+       return 0;
+}
+
+static inline void
+instr_alu_add_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_add_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_add_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_add_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add (hh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HH(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_add_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add (mi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MI(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_add_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] add (hi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HI(t, ip, +);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub (hh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HH(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub (mi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MI(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_sub_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] sub (hi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HI(t, ip, -);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl (hh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HH(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl (mi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MI(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shl_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shl (hi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HI(t, ip, <<);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr (mh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MH(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr (hm)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HM(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr (hh)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HH(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr (mi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_MI(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_shr_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] shr (hi)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_HI(t, ip, >>);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_and_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] and\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, &);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_and_s_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);
+
+       /* Structs. */
+       ALU_S(t, ip, &);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_and_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] and (i)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_I(t, ip, &);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_or_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] or\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, |);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_or_s_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);
+
+       /* Structs. */
+       ALU_S(t, ip, |);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_or_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] or (i)\n", p->thread_id);
+
+       /* Structs. */
+       ALU_I(t, ip, |);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_xor_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] xor\n", p->thread_id);
+
+       /* Structs. */
+       ALU(t, ip, ^);
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_xor_s_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);
+
+       /* Structs. */
+       ALU_S(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];
+       src64_ptr = (uint64_t *)&src_struct[ip->alu.src.offset];
+       src64 = *src64_ptr;
+       src64_mask = UINT64_MAX >> (64 - ip->alu.src.n_bits);
+       src = src64 & src64_mask;
+
+       r = dst;
+       r = ~r & 0xFFFF;
+
+       /* The first input (r) is a 16-bit number. The second and the third
+        * inputs are 32-bit numbers. In the worst case scenario, the sum of the
+        * three numbers (output r) is a 34-bit number.
+        */
+       r += (src >> 32) + (src & 0xFFFFFFFF);
+
+       /* The first input is a 16-bit number. The second input is an 18-bit
+        * number. In the worst case scenario, the sum of the two numbers is a
+        * 19-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* The first input is a 16-bit number (0 .. 0xFFFF). The second input is
+        * a 3-bit number (0 .. 7). Their sum is a 17-bit number (0 .. 0x10006).
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* When the input r is (0 .. 0xFFFF), the output r is equal to the input
+        * r, so the output is (0 .. 0xFFFF). When the input r is (0x10000 ..
+        * 0x10006), the output r is (0 .. 7). So no carry bit can be generated,
+        * therefore the output r is always a 16-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       r = ~r & 0xFFFF;
+       r = r ? r : 0xFFFF;
+
+       *dst16_ptr = (uint16_t)r;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_cksub_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] cksub (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];
+       src64_ptr = (uint64_t *)&src_struct[ip->alu.src.offset];
+       src64 = *src64_ptr;
+       src64_mask = UINT64_MAX >> (64 - ip->alu.src.n_bits);
+       src = src64 & src64_mask;
+
+       r = dst;
+       r = ~r & 0xFFFF;
+
+       /* Subtraction in 1's complement arithmetic (i.e. a '- b) is the same as
+        * the following sequence of operations in 2's complement arithmetic:
+        *    a '- b = (a - b) % 0xFFFF.
+        *
+        * In order to prevent an underflow for the below subtraction, in which
+        * a 33-bit number (the subtrahend) is taken out of a 16-bit number (the
+        * minuend), we first add a multiple of the 0xFFFF modulus to the
+        * minuend. The number we add to the minuend needs to be a 34-bit number
+        * or higher, so for readability reasons we picked the 36-bit multiple.
+        * We are effectively turning the 16-bit minuend into a 36-bit number:
+        *    (a - b) % 0xFFFF = (a + 0xFFFF00000 - b) % 0xFFFF.
+        */
+       r += 0xFFFF00000ULL; /* The output r is a 36-bit number. */
+
+       /* A 33-bit number is subtracted from a 36-bit number (the input r). The
+        * result (the output r) is a 36-bit number.
+        */
+       r -= (src >> 32) + (src & 0xFFFFFFFF);
+
+       /* The first input is a 16-bit number. The second input is a 20-bit
+        * number. Their sum is a 21-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* The first input is a 16-bit number (0 .. 0xFFFF). The second input is
+        * a 5-bit number (0 .. 31). The sum is a 17-bit number (0 .. 0x1001E).
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* When the input r is (0 .. 0xFFFF), the output r is equal to the input
+        * r, so the output is (0 .. 0xFFFF). When the input r is (0x10000 ..
+        * 0x1001E), the output r is (0 .. 31). So no carry bit can be
+        * generated, therefore the output r is always a 16-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       r = ~r & 0xFFFF;
+       r = r ? r : 0xFFFF;
+
+       *dst16_ptr = (uint16_t)r;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_ckadd_struct20_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;
+       uint32_t *src32_ptr;
+       uint64_t r0, r1;
+
+       TRACE("[Thread %2u] ckadd (struct of 20 bytes)\n", p->thread_id);
+
+       /* Structs. */
+       dst_struct = t->structs[ip->alu.dst.struct_id];
+       dst16_ptr = (uint16_t *)&dst_struct[ip->alu.dst.offset];
+
+       src_struct = t->structs[ip->alu.src.struct_id];
+       src32_ptr = (uint32_t *)&src_struct[0];
+
+       r0 = src32_ptr[0]; /* r0 is a 32-bit number. */
+       r1 = src32_ptr[1]; /* r1 is a 32-bit number. */
+       r0 += src32_ptr[2]; /* The output r0 is a 33-bit number. */
+       r1 += src32_ptr[3]; /* The output r1 is a 33-bit number. */
+       r0 += r1 + src32_ptr[4]; /* The output r0 is a 35-bit number. */
+
+       /* The first input is a 16-bit number. The second input is a 19-bit
+        * number. Their sum is a 20-bit number.
+        */
+       r0 = (r0 & 0xFFFF) + (r0 >> 16);
+
+       /* The first input is a 16-bit number (0 .. 0xFFFF). The second input is
+        * a 4-bit number (0 .. 15). The sum is a 17-bit number (0 .. 0x1000E).
+        */
+       r0 = (r0 & 0xFFFF) + (r0 >> 16);
+
+       /* When the input r is (0 .. 0xFFFF), the output r is equal to the input
+        * r, so the output is (0 .. 0xFFFF). When the input r is (0x10000 ..
+        * 0x1000E), the output r is (0 .. 15). So no carry bit can be
+        * generated, therefore the output r is always a 16-bit number.
+        */
+       r0 = (r0 & 0xFFFF) + (r0 >> 16);
+
+       r0 = ~r0 & 0xFFFF;
+       r0 = r0 ? r0 : 0xFFFF;
+
+       *dst16_ptr = (uint16_t)r0;
+
+       /* Thread. */
+       thread_ip_inc(p);
+}
+
+static inline void
+instr_alu_ckadd_struct_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;
+       uint32_t *src32_ptr;
+       uint64_t r = 0;
+       uint32_t i;
+
+       TRACE("[Thread %2u] ckadd (struct)\n", p->thread_id);
+
+       /* Structs. */
+       dst_struct = t->structs[ip->alu.dst.struct_id];
+       dst16_ptr = (uint16_t *)&dst_struct[ip->alu.dst.offset];
+
+       src_struct = t->structs[ip->alu.src.struct_id];
+       src32_ptr = (uint32_t *)&src_struct[0];
+
+       /* The max number of 32-bit words in a 256-byte header is 8 = 2^3.
+        * Therefore, in the worst case scenario, a 35-bit number is added to a
+        * 16-bit number (the input r), so the output r is 36-bit number.
+        */
+       for (i = 0; i < ip->alu.src.n_bits / 32; i++, src32_ptr++)
+               r += *src32_ptr;
+
+       /* The first input is a 16-bit number. The second input is a 20-bit
+        * number. Their sum is a 21-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* The first input is a 16-bit number (0 .. 0xFFFF). The second input is
+        * a 5-bit number (0 .. 31). The sum is a 17-bit number (0 .. 0x1000E).
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       /* When the input r is (0 .. 0xFFFF), the output r is equal to the input
+        * r, so the output is (0 .. 0xFFFF). When the input r is (0x10000 ..
+        * 0x1001E), the output r is (0 .. 31). So no carry bit can be
+        * generated, therefore the output r is always a 16-bit number.
+        */
+       r = (r & 0xFFFF) + (r >> 16);
+
+       r = ~r & 0xFFFF;
+       r = r ? r : 0xFFFF;
+
+       *dst16_ptr = (uint16_t)r;
+
+       /* 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 == 2, 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;
+       uint32_t a_struct_id, b_struct_id, b_val;
+
+       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 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->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 = strtoul(b, &b, 0);
+       CHECK(!b[0], EINVAL);
+
+       if (a[0] == 'h')
+               b_val = htonl(b_val);
+
+       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 = (uint32_t)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;
+       uint32_t a_struct_id, b_struct_id, b_val;
+
+       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 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->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 = strtoul(b, &b, 0);
+       CHECK(!b[0], EINVAL);
+
+       if (a[0] == 'h')
+               b_val = htonl(b_val);
+
+       instr->type = INSTR_JMP_NEQ_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 = (uint32_t)b_val;
+       return 0;
+}
+
+static int
+instr_jmp_lt_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;
+       uint32_t a_struct_id, b_struct_id, b_val;
+
+       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_LT, JMP_LT_MH, JMP_LT_HM, JMP_LT_HH. */
+       fb = struct_field_parse(p, action, b, &b_struct_id);
+       if (fb) {
+               instr->type = INSTR_JMP_LT;
+               if (a[0] == 'h' && b[0] == 'm')
+                       instr->type = INSTR_JMP_LT_HM;
+               if (a[0] == 'm' && b[0] == 'h')
+                       instr->type = INSTR_JMP_LT_MH;
+               if (a[0] == 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_LT_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_LT_MI, JMP_LT_HI. */
+       b_val = strtoul(b, &b, 0);
+       CHECK(!b[0], EINVAL);
+
+       instr->type = INSTR_JMP_LT_MI;
+       if (a[0] == 'h')
+               instr->type = INSTR_JMP_LT_HI;
+       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 = (uint32_t)b_val;
+       return 0;
+}
+
+static int
+instr_jmp_gt_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;
+       uint32_t a_struct_id, b_struct_id, b_val;
+
+       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_GT, JMP_GT_MH, JMP_GT_HM, JMP_GT_HH. */
+       fb = struct_field_parse(p, action, b, &b_struct_id);
+       if (fb) {
+               instr->type = INSTR_JMP_GT;
+               if (a[0] == 'h' && b[0] == 'm')
+                       instr->type = INSTR_JMP_GT_HM;
+               if (a[0] == 'm' && b[0] == 'h')
+                       instr->type = INSTR_JMP_GT_MH;
+               if (a[0] == 'h' && b[0] == 'h')
+                       instr->type = INSTR_JMP_GT_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_GT_MI, JMP_GT_HI. */
+       b_val = strtoul(b, &b, 0);
+       CHECK(!b[0], EINVAL);
+
+       instr->type = INSTR_JMP_GT_MI;
+       if (a[0] == 'h')
+               instr->type = INSTR_JMP_GT_HI;
+       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 = (uint32_t)b_val;
+       return 0;
+}
+
+static inline void
+instr_jmp_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmp\n", p->thread_id);
+
+       thread_ip_set(t, ip->jmp.ip);
+}
+
+static inline void
+instr_jmp_valid_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t header_id = ip->jmp.header_id;
+
+       TRACE("[Thread %2u] jmpv\n", p->thread_id);
+
+       t->ip = HEADER_VALID(t, header_id) ? ip->jmp.ip : (t->ip + 1);
+}
+
+static inline void
+instr_jmp_invalid_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       uint32_t header_id = ip->jmp.header_id;
+
+       TRACE("[Thread %2u] jmpnv\n", p->thread_id);
+
+       t->ip = HEADER_VALID(t, header_id) ? (t->ip + 1) : ip->jmp.ip;
+}
+
+static inline void
+instr_jmp_hit_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct instruction *ip_next[] = {t->ip + 1, ip->jmp.ip};
+
+       TRACE("[Thread %2u] jmph\n", p->thread_id);
+
+       t->ip = ip_next[t->hit];
+}
+
+static inline void
+instr_jmp_miss_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+       struct instruction *ip_next[] = {ip->jmp.ip, t->ip + 1};
+
+       TRACE("[Thread %2u] jmpnh\n", p->thread_id);
+
+       t->ip = ip_next[t->hit];
+}
+
+static inline void
+instr_jmp_action_hit_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpa\n", p->thread_id);
+
+       t->ip = (ip->jmp.action_id == t->action_id) ? ip->jmp.ip : (t->ip + 1);
+}
+
+static inline void
+instr_jmp_action_miss_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpna\n", p->thread_id);
+
+       t->ip = (ip->jmp.action_id == t->action_id) ? (t->ip + 1) : ip->jmp.ip;
+}
+
+static inline void
+instr_jmp_eq_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpeq\n", p->thread_id);
+
+       JMP_CMP(t, ip, ==);
+}
+
+static inline void
+instr_jmp_eq_s_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);
+
+       JMP_CMP_S(t, ip, ==);
+}
+
+static inline void
+instr_jmp_eq_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpeq (i)\n", p->thread_id);
+
+       JMP_CMP_I(t, ip, ==);
+}
+
+static inline void
+instr_jmp_neq_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpneq\n", p->thread_id);
+
+       JMP_CMP(t, ip, !=);
+}
+
+static inline void
+instr_jmp_neq_s_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);
+
+       JMP_CMP_S(t, ip, !=);
+}
+
+static inline void
+instr_jmp_neq_i_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpneq (i)\n", p->thread_id);
+
+       JMP_CMP_I(t, ip, !=);
+}
+
+static inline void
+instr_jmp_lt_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt\n", p->thread_id);
+
+       JMP_CMP(t, ip, <);
+}
+
+static inline void
+instr_jmp_lt_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt (mh)\n", p->thread_id);
+
+       JMP_CMP_MH(t, ip, <);
+}
+
+static inline void
+instr_jmp_lt_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt (hm)\n", p->thread_id);
+
+       JMP_CMP_HM(t, ip, <);
+}
+
+static inline void
+instr_jmp_lt_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt (hh)\n", p->thread_id);
+
+       JMP_CMP_HH(t, ip, <);
+}
+
+static inline void
+instr_jmp_lt_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt (mi)\n", p->thread_id);
+
+       JMP_CMP_MI(t, ip, <);
+}
+
+static inline void
+instr_jmp_lt_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmplt (hi)\n", p->thread_id);
+
+       JMP_CMP_HI(t, ip, <);
+}
+
+static inline void
+instr_jmp_gt_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt\n", p->thread_id);
+
+       JMP_CMP(t, ip, >);
+}
+
+static inline void
+instr_jmp_gt_mh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt (mh)\n", p->thread_id);
+
+       JMP_CMP_MH(t, ip, >);
+}
+
+static inline void
+instr_jmp_gt_hm_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt (hm)\n", p->thread_id);
+
+       JMP_CMP_HM(t, ip, >);
+}
+
+static inline void
+instr_jmp_gt_hh_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt (hh)\n", p->thread_id);
+
+       JMP_CMP_HH(t, ip, >);
+}
+
+static inline void
+instr_jmp_gt_mi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt (mi)\n", p->thread_id);
+
+       JMP_CMP_MI(t, ip, >);
+}
+
+static inline void
+instr_jmp_gt_hi_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+       struct instruction *ip = t->ip;
+
+       TRACE("[Thread %2u] jmpgt (hi)\n", p->thread_id);
+
+       JMP_CMP_HI(t, ip, >);
+}
+
+/*
+ * return.
+ */
+static int
+instr_return_translate(struct rte_swx_pipeline *p __rte_unused,
+                      struct action *action,
+                      char **tokens __rte_unused,
+                      int n_tokens,
+                      struct instruction *instr,
+                      struct instruction_data *data __rte_unused)
+{
+       CHECK(action, EINVAL);
+       CHECK(n_tokens == 1, EINVAL);
+
+       instr->type = INSTR_RETURN;
+       return 0;
+}
+
+static inline void
+instr_return_exec(struct rte_swx_pipeline *p)
+{
+       struct thread *t = &p->threads[p->thread_id];
+
+       TRACE("[Thread %2u] return\n", p->thread_id);
+
+       t->ip = t->ret;
+}
+
+#define RTE_SWX_INSTRUCTION_TOKENS_MAX 16
+
+static int
+instr_translate(struct rte_swx_pipeline *p,
+               struct action *action,
+               char *string,
+               struct instruction *instr,
+               struct instruction_data *data)
+{
+       char *tokens[RTE_SWX_INSTRUCTION_TOKENS_MAX];
+       int n_tokens = 0, tpos = 0;
+
+       /* Parse the instruction string into tokens. */
+       for ( ; ; ) {
+               char *token;
+
+               token = strtok_r(string, " \t\v", &string);
+               if (!token)
+                       break;
+
+               CHECK(n_tokens < RTE_SWX_INSTRUCTION_TOKENS_MAX, EINVAL);
+
+               tokens[n_tokens] = token;
+               n_tokens++;
+       }
+
+       CHECK(n_tokens, EINVAL);
+
+       /* Handle the optional instruction label. */
+       if ((n_tokens >= 2) && !strcmp(tokens[1], ":")) {
+               strcpy(data->label, tokens[0]);
+
+               tpos += 2;
+               CHECK(n_tokens - tpos, EINVAL);
+       }
+
+       /* Identify the instruction type. */
+       if (!strcmp(tokens[tpos], "rx"))
+               return instr_rx_translate(p,
+                                         action,
+                                         &tokens[tpos],
+                                         n_tokens - tpos,
+                                         instr,
+                                         data);
+
+       if (!strcmp(tokens[tpos], "tx"))
+               return instr_tx_translate(p,
+                                         action,
+                                         &tokens[tpos],
+                                         n_tokens - tpos,
+                                         instr,
+                                         data);
+
+       if (!strcmp(tokens[tpos], "extract"))
+               return instr_hdr_extract_translate(p,
+                                                  action,
+                                                  &tokens[tpos],
+                                                  n_tokens - tpos,
+                                                  instr,
+                                                  data);
+
+       if (!strcmp(tokens[tpos], "emit"))
+               return instr_hdr_emit_translate(p,
+                                               action,
+                                               &tokens[tpos],
+                                               n_tokens - tpos,
+                                               instr,
+                                               data);
+
+       if (!strcmp(tokens[tpos], "validate"))
+               return instr_hdr_validate_translate(p,
+                                                   action,
+                                                   &tokens[tpos],
+                                                   n_tokens - tpos,
+                                                   instr,
+                                                   data);
+
+       if (!strcmp(tokens[tpos], "invalidate"))
+               return instr_hdr_invalidate_translate(p,
+                                                     action,
+                                                     &tokens[tpos],
+                                                     n_tokens - tpos,
+                                                     instr,
+                                                     data);
+
+       if (!strcmp(tokens[tpos], "mov"))
+               return instr_mov_translate(p,
+                                          action,
+                                          &tokens[tpos],
+                                          n_tokens - tpos,
+                                          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,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "sub"))
+               return instr_alu_sub_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "ckadd"))
+               return instr_alu_ckadd_translate(p,
+                                                action,
+                                                &tokens[tpos],
+                                                n_tokens - tpos,
+                                                instr,
+                                                data);
+
+       if (!strcmp(tokens[tpos], "cksub"))
+               return instr_alu_cksub_translate(p,
+                                                action,
+                                                &tokens[tpos],
+                                                n_tokens - tpos,
+                                                instr,
+                                                data);
+
+       if (!strcmp(tokens[tpos], "and"))
+               return instr_alu_and_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "or"))
+               return instr_alu_or_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       if (!strcmp(tokens[tpos], "xor"))
+               return instr_alu_xor_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "shl"))
+               return instr_alu_shl_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "shr"))
+               return instr_alu_shr_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "table"))
+               return instr_table_translate(p,
+                                            action,
+                                            &tokens[tpos],
+                                            n_tokens - tpos,
+                                            instr,
+                                            data);
+
+       if (!strcmp(tokens[tpos], "extern"))
+               return instr_extern_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       if (!strcmp(tokens[tpos], "jmp"))
+               return instr_jmp_translate(p,
+                                          action,
+                                          &tokens[tpos],
+                                          n_tokens - tpos,
+                                          instr,
+                                          data);
+
+       if (!strcmp(tokens[tpos], "jmpv"))
+               return instr_jmp_valid_translate(p,
+                                                action,
+                                                &tokens[tpos],
+                                                n_tokens - tpos,
+                                                instr,
+                                                data);
+
+       if (!strcmp(tokens[tpos], "jmpnv"))
+               return instr_jmp_invalid_translate(p,
+                                                  action,
+                                                  &tokens[tpos],
+                                                  n_tokens - tpos,
+                                                  instr,
+                                                  data);
+
+       if (!strcmp(tokens[tpos], "jmph"))
+               return instr_jmp_hit_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "jmpnh"))
+               return instr_jmp_miss_translate(p,
+                                               action,
+                                               &tokens[tpos],
+                                               n_tokens - tpos,
+                                               instr,
+                                               data);
+
+       if (!strcmp(tokens[tpos], "jmpa"))
+               return instr_jmp_action_hit_translate(p,
+                                                     action,
+                                                     &tokens[tpos],
+                                                     n_tokens - tpos,
+                                                     instr,
+                                                     data);
+
+       if (!strcmp(tokens[tpos], "jmpna"))
+               return instr_jmp_action_miss_translate(p,
+                                                      action,
+                                                      &tokens[tpos],
+                                                      n_tokens - tpos,
+                                                      instr,
+                                                      data);
+
+       if (!strcmp(tokens[tpos], "jmpeq"))
+               return instr_jmp_eq_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       if (!strcmp(tokens[tpos], "jmpneq"))
+               return instr_jmp_neq_translate(p,
+                                              action,
+                                              &tokens[tpos],
+                                              n_tokens - tpos,
+                                              instr,
+                                              data);
+
+       if (!strcmp(tokens[tpos], "jmplt"))
+               return instr_jmp_lt_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       if (!strcmp(tokens[tpos], "jmpgt"))
+               return instr_jmp_gt_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       if (!strcmp(tokens[tpos], "return"))
+               return instr_return_translate(p,
+                                             action,
+                                             &tokens[tpos],
+                                             n_tokens - tpos,
+                                             instr,
+                                             data);
+
+       CHECK(0, EINVAL);
+}
+
+static struct instruction_data *
+label_find(struct instruction_data *data, uint32_t n, const char *label)
+{
+       uint32_t i;
+
+       for (i = 0; i < n; i++)
+               if (!strcmp(label, data[i].label))
+                       return &data[i];
+
+       return NULL;
+}
+
+static uint32_t
+label_is_used(struct instruction_data *data, uint32_t n, const char *label)
+{
+       uint32_t count = 0, i;
+
+       if (!label[0])
+               return 0;
+
+       for (i = 0; i < n; i++)
+               if (!strcmp(label, data[i].jmp_label))
+                       count++;
+
+       return count;
+}
+
+static int
+instr_label_check(struct instruction_data *instruction_data,
+                 uint32_t n_instructions)
+{
+       uint32_t i;
+
+       /* Check that all instruction labels are unique. */
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction_data *data = &instruction_data[i];
+               char *label = data->label;
+               uint32_t j;
+
+               if (!label[0])
+                       continue;
+
+               for (j = i + 1; j < n_instructions; j++)
+                       CHECK(strcmp(label, data[j].label), EINVAL);
+       }
+
+       /* Get users for each instruction label. */
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction_data *data = &instruction_data[i];
+               char *label = data->label;
+
+               data->n_users = label_is_used(instruction_data,
+                                             n_instructions,
+                                             label);
+       }
+
+       return 0;
+}
+
+static int
+instr_jmp_resolve(struct instruction *instructions,
+                 struct instruction_data *instruction_data,
+                 uint32_t n_instructions)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+               struct instruction_data *found;
+
+               if (!instruction_is_jmp(instr))
+                       continue;
+
+               found = label_find(instruction_data,
+                                  n_instructions,
+                                  data->jmp_label);
+               CHECK(found, EINVAL);
+
+               instr->jmp.ip = &instr[found - instruction_data];
+       }
+
+       return 0;
+}
+
+static int
+instr_verify(struct rte_swx_pipeline *p __rte_unused,
+            struct action *a,
+            struct instruction *instr,
+            struct instruction_data *data __rte_unused,
+            uint32_t n_instructions)
+{
+       if (!a) {
+               enum instruction_type type;
+               uint32_t i;
+
+               /* Check that the first instruction is rx. */
+               CHECK(instr[0].type == INSTR_RX, EINVAL);
+
+               /* Check that there is at least one tx instruction. */
+               for (i = 0; i < n_instructions; i++) {
+                       type = instr[i].type;
+
+                       if (instr[i].type == INSTR_TX)
+                               break;
+               }
+               CHECK(i < n_instructions, EINVAL);
+
+               /* Check that the last instruction is either tx or unconditional
+                * jump.
+                */
+               type = instr[n_instructions - 1].type;
+               CHECK((type == INSTR_TX) || (type == INSTR_JMP), EINVAL);
+       }
+
+       if (a) {
+               enum instruction_type type;
+               uint32_t i;
+
+               /* Check that there is at least one return or tx instruction. */
+               for (i = 0; i < n_instructions; i++) {
+                       type = instr[i].type;
+
+                       if ((type == INSTR_RETURN) || (type == INSTR_TX))
+                               break;
+               }
+               CHECK(i < n_instructions, EINVAL);
+       }
+
+       return 0;
+}
+
+static int
+instr_pattern_extract_many_detect(struct instruction *instr,
+                                 struct instruction_data *data,
+                                 uint32_t n_instr,
+                                 uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_HDR_EXTRACT)
+                       break;
+
+               if (i == RTE_DIM(instr->io.hdr.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (i < 2)
+               return 0;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_extract_many_optimize(struct instruction *instr,
+                                   struct instruction_data *data,
+                                   uint32_t n_instr)
+{
+       uint32_t i;
+
+       for (i = 1; i < n_instr; i++) {
+               instr[0].type++;
+               instr[0].io.hdr.header_id[i] = instr[i].io.hdr.header_id[0];
+               instr[0].io.hdr.struct_id[i] = instr[i].io.hdr.struct_id[0];
+               instr[0].io.hdr.n_bytes[i] = instr[i].io.hdr.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+}
+
+static int
+instr_pattern_emit_many_tx_detect(struct instruction *instr,
+                                 struct instruction_data *data,
+                                 uint32_t n_instr,
+                                 uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_HDR_EMIT)
+                       break;
+
+               if (i == RTE_DIM(instr->io.hdr.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (!i)
+               return 0;
+
+       if (instr[i].type != INSTR_TX)
+               return 0;
+
+       i++;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_emit_many_tx_optimize(struct instruction *instr,
+                                   struct instruction_data *data,
+                                   uint32_t n_instr)
+{
+       uint32_t i;
+
+       /* Any emit instruction in addition to the first one. */
+       for (i = 1; i < n_instr - 1; i++) {
+               instr[0].type++;
+               instr[0].io.hdr.header_id[i] = instr[i].io.hdr.header_id[0];
+               instr[0].io.hdr.struct_id[i] = instr[i].io.hdr.struct_id[0];
+               instr[0].io.hdr.n_bytes[i] = instr[i].io.hdr.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+
+       /* The TX instruction is the last one in the pattern. */
+       instr[0].type++;
+       instr[0].io.io.offset = instr[i].io.io.offset;
+       instr[0].io.io.n_bits = instr[i].io.io.n_bits;
+       data[i].invalid = 1;
+}
+
+static int
+instr_pattern_dma_many_detect(struct instruction *instr,
+                             struct instruction_data *data,
+                             uint32_t n_instr,
+                             uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_DMA_HT)
+                       break;
+
+               if (i == RTE_DIM(instr->dma.dst.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (i < 2)
+               return 0;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_dma_many_optimize(struct instruction *instr,
+                               struct instruction_data *data,
+                               uint32_t n_instr)
+{
+       uint32_t i;
+
+       for (i = 1; i < n_instr; i++) {
+               instr[0].type++;
+               instr[0].dma.dst.header_id[i] = instr[i].dma.dst.header_id[0];
+               instr[0].dma.dst.struct_id[i] = instr[i].dma.dst.struct_id[0];
+               instr[0].dma.src.offset[i] = instr[i].dma.src.offset[0];
+               instr[0].dma.n_bytes[i] = instr[i].dma.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+}
+
+static uint32_t
+instr_optimize(struct instruction *instructions,
+              struct instruction_data *instruction_data,
+              uint32_t n_instructions)
+{
+       uint32_t i, pos = 0;
+
+       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_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,
+                                                        data,
+                                                        n_instructions - i,
+                                                        &n_instr);
+               if (detected) {
+                       instr_pattern_dma_many_optimize(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. */
+       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
+instruction_config(struct rte_swx_pipeline *p,
+                  struct action *a,
+                  const char **instructions,
+                  uint32_t n_instructions)
+{
+       struct instruction *instr = NULL;
+       struct instruction_data *data = NULL;
+       char *string = NULL;
+       int err = 0;
+       uint32_t i;
+
+       CHECK(n_instructions, EINVAL);
+       CHECK(instructions, EINVAL);
+       for (i = 0; i < n_instructions; i++)
+               CHECK(instructions[i], EINVAL);
+
+       /* Memory allocation. */
+       instr = calloc(n_instructions, sizeof(struct instruction));
+       if (!instr) {
+               err = ENOMEM;
+               goto error;
+       }
+
+       data = calloc(n_instructions, sizeof(struct instruction_data));
        if (!data) {
                err = ENOMEM;
                goto error;
@@ -2265,6 +5972,16 @@ instruction_config(struct rte_swx_pipeline *p,
        if (err)
                goto error;
 
+       err = instr_verify(p, a, instr, data, n_instructions);
+       if (err)
+               goto error;
+
+       n_instructions = instr_optimize(instr, data, n_instructions);
+
+       err = instr_jmp_resolve(instr, data, n_instructions);
+       if (err)
+               goto error;
+
        free(data);
 
        if (a) {
@@ -2308,6 +6025,103 @@ static instr_exec_t instruction_table[] = {
        [INSTR_HDR_EMIT6_TX] = instr_hdr_emit6_tx_exec,
        [INSTR_HDR_EMIT7_TX] = instr_hdr_emit7_tx_exec,
        [INSTR_HDR_EMIT8_TX] = instr_hdr_emit8_tx_exec,
+
+       [INSTR_HDR_VALIDATE] = instr_hdr_validate_exec,
+       [INSTR_HDR_INVALIDATE] = instr_hdr_invalidate_exec,
+
+       [INSTR_MOV] = instr_mov_exec,
+       [INSTR_MOV_S] = instr_mov_s_exec,
+       [INSTR_MOV_I] = instr_mov_i_exec,
+
+       [INSTR_DMA_HT] = instr_dma_ht_exec,
+       [INSTR_DMA_HT2] = instr_dma_ht2_exec,
+       [INSTR_DMA_HT3] = instr_dma_ht3_exec,
+       [INSTR_DMA_HT4] = instr_dma_ht4_exec,
+       [INSTR_DMA_HT5] = instr_dma_ht5_exec,
+       [INSTR_DMA_HT6] = instr_dma_ht6_exec,
+       [INSTR_DMA_HT7] = instr_dma_ht7_exec,
+       [INSTR_DMA_HT8] = instr_dma_ht8_exec,
+
+       [INSTR_ALU_ADD] = instr_alu_add_exec,
+       [INSTR_ALU_ADD_MH] = instr_alu_add_mh_exec,
+       [INSTR_ALU_ADD_HM] = instr_alu_add_hm_exec,
+       [INSTR_ALU_ADD_HH] = instr_alu_add_hh_exec,
+       [INSTR_ALU_ADD_MI] = instr_alu_add_mi_exec,
+       [INSTR_ALU_ADD_HI] = instr_alu_add_hi_exec,
+
+       [INSTR_ALU_SUB] = instr_alu_sub_exec,
+       [INSTR_ALU_SUB_MH] = instr_alu_sub_mh_exec,
+       [INSTR_ALU_SUB_HM] = instr_alu_sub_hm_exec,
+       [INSTR_ALU_SUB_HH] = instr_alu_sub_hh_exec,
+       [INSTR_ALU_SUB_MI] = instr_alu_sub_mi_exec,
+       [INSTR_ALU_SUB_HI] = instr_alu_sub_hi_exec,
+
+       [INSTR_ALU_CKADD_FIELD] = instr_alu_ckadd_field_exec,
+       [INSTR_ALU_CKADD_STRUCT] = instr_alu_ckadd_struct_exec,
+       [INSTR_ALU_CKADD_STRUCT20] = instr_alu_ckadd_struct20_exec,
+       [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_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_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_I] = instr_alu_xor_i_exec,
+
+       [INSTR_ALU_SHL] = instr_alu_shl_exec,
+       [INSTR_ALU_SHL_MH] = instr_alu_shl_mh_exec,
+       [INSTR_ALU_SHL_HM] = instr_alu_shl_hm_exec,
+       [INSTR_ALU_SHL_HH] = instr_alu_shl_hh_exec,
+       [INSTR_ALU_SHL_MI] = instr_alu_shl_mi_exec,
+       [INSTR_ALU_SHL_HI] = instr_alu_shl_hi_exec,
+
+       [INSTR_ALU_SHR] = instr_alu_shr_exec,
+       [INSTR_ALU_SHR_MH] = instr_alu_shr_mh_exec,
+       [INSTR_ALU_SHR_HM] = instr_alu_shr_hm_exec,
+       [INSTR_ALU_SHR_HH] = instr_alu_shr_hh_exec,
+       [INSTR_ALU_SHR_MI] = instr_alu_shr_mi_exec,
+       [INSTR_ALU_SHR_HI] = instr_alu_shr_hi_exec,
+
+       [INSTR_TABLE] = instr_table_exec,
+       [INSTR_EXTERN_OBJ] = instr_extern_obj_exec,
+       [INSTR_EXTERN_FUNC] = instr_extern_func_exec,
+
+       [INSTR_JMP] = instr_jmp_exec,
+       [INSTR_JMP_VALID] = instr_jmp_valid_exec,
+       [INSTR_JMP_INVALID] = instr_jmp_invalid_exec,
+       [INSTR_JMP_HIT] = instr_jmp_hit_exec,
+       [INSTR_JMP_MISS] = instr_jmp_miss_exec,
+       [INSTR_JMP_ACTION_HIT] = instr_jmp_action_hit_exec,
+       [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_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_I] = instr_jmp_neq_i_exec,
+
+       [INSTR_JMP_LT] = instr_jmp_lt_exec,
+       [INSTR_JMP_LT_MH] = instr_jmp_lt_mh_exec,
+       [INSTR_JMP_LT_HM] = instr_jmp_lt_hm_exec,
+       [INSTR_JMP_LT_HH] = instr_jmp_lt_hh_exec,
+       [INSTR_JMP_LT_MI] = instr_jmp_lt_mi_exec,
+       [INSTR_JMP_LT_HI] = instr_jmp_lt_hi_exec,
+
+       [INSTR_JMP_GT] = instr_jmp_gt_exec,
+       [INSTR_JMP_GT_MH] = instr_jmp_gt_mh_exec,
+       [INSTR_JMP_GT_HM] = instr_jmp_gt_hm_exec,
+       [INSTR_JMP_GT_HH] = instr_jmp_gt_hh_exec,
+       [INSTR_JMP_GT_MI] = instr_jmp_gt_mi_exec,
+       [INSTR_JMP_GT_HI] = instr_jmp_gt_hi_exec,
+
+       [INSTR_RETURN] = instr_return_exec,
 };
 
 static inline void
@@ -2338,6 +6152,21 @@ action_find(struct rte_swx_pipeline *p, const char *name)
        return NULL;
 }
 
+static struct field *
+action_field_find(struct action *a, const char *name)
+{
+       return a->st ? struct_type_field_find(a->st, name) : NULL;
+}
+
+static struct field *
+action_field_parse(struct action *action, const char *name)
+{
+       if (name[0] != 't' || name[1] != '.')
+               return NULL;
+
+       return action_field_find(action, &name[2]);
+}
+
 int
 rte_swx_pipeline_action_config(struct rte_swx_pipeline *p,
                               const char *name,