add an additional field (packet) to LCII_context_t

lci/experimental/coll/coll.h

@@ -144,7 +144,7 @@ static inline void LCIXC_mcoll_complete(LCI_endpoint_t ep, LCI_mbuffer_t buffer,
   LCII_comp_attr_set_comp_type(ctx->comp_attr, ep->msg_comp_type);
   ctx->data_type = LCI_MEDIUM;
   ctx->user_context = user_context;
-  ctx->data = (LCI_data_t){.mbuffer = buffer};
+  ctx->data.mbuffer = buffer;
   ctx->rank = -1; /* this doesn't make much sense for collectives */
   ctx->tag = tag;
   ctx->completion = completion;
@@ -160,7 +160,7 @@ static inline void LCIXC_lcoll_complete(LCI_endpoint_t ep, LCI_lbuffer_t buffer,
   LCII_comp_attr_set_comp_type(ctx->comp_attr, ep->msg_comp_type);
   ctx->data_type = LCI_LONG;
   ctx->user_context = user_context;
-  ctx->data = (LCI_data_t){.lbuffer = buffer};
+  ctx->data.lbuffer = buffer;
   ctx->rank = -1; /* this doesn't make much sense for collectives */
   ctx->tag = tag;
   ctx->completion = completion;

lci/runtime/completion/sync_flag.c

@@ -57,9 +57,8 @@ LCI_error_t LCI_sync_signal(LCI_comp_t completion, LCI_request_t request)
   ctx->rank = request.rank;
   ctx->tag = request.tag;
   ctx->data_type = request.type;
-  ctx->data = request.data;
+  memcpy(&ctx->data, &request.data, sizeof(ctx->data));
   ctx->user_context = request.user_context;
-
   LCII_sync_signal(completion, ctx);
   return LCI_OK;
 }

lci/runtime/lcii.h

@@ -137,7 +137,15 @@ typedef struct __attribute__((aligned(LCI_CACHE_LINE))) {
   // LCI_request_t fields, 52 bytes
   LCI_data_type_t data_type;  // 4 bytes
   void* user_context;         // 8 bytes
-  LCI_data_t data;            // 32 bytes
+  union {
+    LCI_short_t immediate;    // 32 bytes
+    struct {                  // 24 bytes
+      LCI_mbuffer_t mbuffer;
+      LCII_packet_t *packet;
+    };
+    LCI_lbuffer_t lbuffer;    // 24 bytes
+    LCI_iovec_t iovec;        // 28 bytes
+  } data;                     // 32 bytes
   uint32_t rank;              // 4 bytes
   LCI_tag_t tag;              // 4 bytes
   // used by LCI internally
@@ -200,8 +208,11 @@ static inline LCI_request_t LCII_ctx2req(LCII_context_t* ctx)
                            .rank = ctx->rank,
                            .tag = ctx->tag,
                            .type = ctx->data_type,
-                           .data = ctx->data,
                            .user_context = ctx->user_context};
+  LCI_DBG_Assert(sizeof(request.data) == sizeof(ctx->data), "Unexpected size!\n");
+  memcpy(&request.data, &ctx->data, sizeof(request.data));
+  LCI_DBG_Assert(request.data.mbuffer.address == ctx->data.mbuffer.address, "Invalid conversion!");
+  LCI_DBG_Assert(request.data.mbuffer.length == ctx->data.mbuffer.length, "Invalid conversion!");
   LCIU_free(ctx);
   return request;
 }