csum: add csum_replace4 for 4-byte value

Test using tcp/udp, modify both 16-bit port and 32-bit address,
see xdp7.p4 for example.

p4include/xdp_model.p4

@@ -59,6 +59,10 @@ extern bit<16> csum_replace2(in bit<16> csum,  // current csum
                              in bit<16> old,   // old value of the field
                              in bit<16> new);
 
+extern bit<16> csum_replace4(in bit<16> csum,
+                             in bit<32> old,
+                             in bit<32> new);
+
 extern bit<32> BPF_PERF_EVENT_OUTPUT();
 // FIXME: use 64 bit
 extern bit<32> BPF_KTIME_GET_NS();

tests/xdp7.p4

@@ -16,14 +16,14 @@ header IPv4 {
     bit<13> fragOffset;
     bit<8>  ttl;
     bit<8>  protocol;
-    bit<16> hdrChecksum;
+    bit<16> checksum;
     bit<32> srcAddr;
     bit<32> dstAddr;
 }
 
 header icmp_t {
     bit<16> typeCode;
-    bit<16> hdrChecksum;
+    bit<16> checksum;
 }
 
 header tcp_t {
@@ -92,21 +92,38 @@ control Ingress(inout Headers hd, in xdp_input xin, out xdp_output xout) {
     // from, to are host byte order
     bit<16> from;
     bit<16> to;
+    bit<32> from_addr;
+    bit<32> to_addr;
 
     action Fallback_action()
     {
-        // UDP
+#if 0
+        // UDP port
         from = hd.udp.dstPort;
         to = 16w0x400;
         hd.udp.dstPort = to;
         hd.udp.checksum = csum_replace2(hd.udp.checksum, from, to);
 
+        // UDP IP addr
+        from_addr = hd.ipv4.dstAddr;
+        to_addr = 32w0x01020304;
+        hd.ipv4.dstAddr = to_addr;
+        hd.ipv4.checksum = csum_replace4(hd.ipv4.checksum, from_addr, to_addr);
+        hd.udp.checksum = csum_replace4(hd.udp.checksum, from_addr, to_addr);
+#endif
         // TCP
         from = hd.tcp.srcPort;
         to = 16w0x841;
         hd.tcp.srcPort = to;
         hd.tcp.checksum = csum_replace2(hd.tcp.checksum, from, to);
 
+        // TCP IP addr
+        from_addr = hd.ipv4.srcAddr;
+        to_addr = 32w0x05060708;
+        hd.ipv4.srcAddr = to_addr;
+        hd.ipv4.checksum = csum_replace4(hd.ipv4.checksum, from_addr, to_addr);
+        hd.tcp.checksum = csum_replace4(hd.tcp.checksum, from_addr, to_addr);
+
         xoutdrop = false;
     }
 

xdpModel.h

@@ -65,6 +65,7 @@ class XDPModel : public EBPF::EBPFModel {
     XDPModel() : EBPF::EBPFModel(), xdp(), inputMetadataModel(), outputMetadataModel(),
                  ipv4_checksum("ebpf_ipv4_checksum"), bpf_event_output("bpf_event_output"),
                  bpf_ktime_get_ns("bpf_ktime_get_ns"), csum_replace2("csum_replace2"),
+                 csum_replace4("csum_replace4"),
                  action_enum()
     {}
 
@@ -75,6 +76,7 @@ class XDPModel : public EBPF::EBPFModel {
     OutputMetadataModel outputMetadataModel;
     ::Model::Extern_Model ipv4_checksum;
     ::Model::Extern_Model csum_replace2;
+    ::Model::Extern_Model csum_replace4;
     ::Model::Extern_Model bpf_event_output;
     ::Model::Extern_Model bpf_ktime_get_ns;
     XDP_Action_Model      action_enum;

xdpProgram.cpp

@@ -150,6 +150,29 @@ void XDPProgram::emitC(EBPF::CodeBuilder* builder, cstring headerFile) {
 		"    return (~csum16_add(csum16_sub(~csum, old), new));\n"
 	"}\n");
 
+    builder->appendLine(
+        "inline u16 csum_fold(u32 csum) {\n"
+        "    u32 r = csum << 16 | csum >> 16;\n"
+        "    csum = ~csum;\n"
+        "    csum -= r;\n"
+        "    return (u16)(csum >> 16);\n"
+        "}\n"
+        "inline u32 csum_unfold(u16 csum) {\n"
+        "    return (u32)csum;\n"
+        "}\n"
+		"inline u32 csum32_add(u32 csum, u32 addend) {\n"
+		"    u32 res = csum;\n"
+		"    res += addend;\n"
+		"    return (res + (res < addend));\n"
+		"}\n"
+		"inline u32 csum32_sub(u32 csum, u32 addend) {\n"
+		"    return csum32_add(csum, ~addend);\n"
+		"}\n"
+		"inline u16 csum_replace4(u16 csum, u32 from, u32 to) {\n"
+        "    u32 tmp = csum32_sub(~csum_unfold(csum), from);\n"
+		"    return csum_fold(csum32_add(tmp, to));\n"
+	"}\n");
+
     builder->appendLine(
         "struct bpf_map_def SEC(\"maps\") perf_event = {\n"
         "   .type = BPF_MAP_TYPE_PERF_EVENT_ARRAY,\n"