Continue footprint building even if an error occur in another API com…

…mand (#1256)

* Continue footprint building even if an error occur in another API command

* Remove error log for faults in a trace, make sure don't crash due to out-of-bound in the behavior list, fallback to dce disabled mode if dependency tracking info is broken.

gapis/api/transform/dce.go

@@ -68,6 +68,7 @@ func (s *commandIndicesSet) contains(fci api.SubCmdIdx) bool {
 type DCE struct {
 	footprint        *dependencygraph.Footprint
 	endBehaviorIndex uint64
+	endCmdIndex      api.CmdID
 	requests         *commandIndicesSet
 	requestCount     uint64
 }
@@ -90,6 +91,9 @@ func (t *DCE) Request(ctx context.Context, fci api.SubCmdIdx) {
 	if bi > t.endBehaviorIndex {
 		t.endBehaviorIndex = bi
 	}
+	if api.CmdID(fci[0]) > t.endCmdIndex {
+		t.endCmdIndex = api.CmdID(fci[0])
+	}
 }
 
 // Transform is to comform the interface of Transformer, but does not accept
@@ -106,6 +110,15 @@ func (t *DCE) Transform(ctx context.Context, id api.CmdID, c api.Cmd,
 // the alive commands to the following transforms to mutate them and write them
 // to build instructions for replay.
 func (t *DCE) Flush(ctx context.Context, out Writer) {
+	if t.endBehaviorIndex >= uint64(len(t.footprint.Behaviors)) {
+		log.E(ctx, "DCE: Cannot backpropagate through def-use chain from behavior index: %v, "+
+			"with length of behavior list: %v.", t.endBehaviorIndex, len(t.footprint.Behaviors))
+		log.W(ctx, "DCE: Fallback to disable DCE.")
+		for i := api.CmdID(0); i <= t.endCmdIndex; i++ {
+			out.MutateAndWrite(ctx, i, t.footprint.Commands[int(i)])
+		}
+		return
+	}
 	t0 := dCECounter.Start()
 	livenessBoard, aliveCmds := t.backPropagate(ctx)
 	dCECounter.Stop(t0)

gapis/api/vulkan/footprint_builder.go

@@ -1279,7 +1279,8 @@ func (vb *FootprintBuilder) BuildFootprint(ctx context.Context,
 
 	// Mutate
 	if err := cmd.Mutate(ctx, id, s, nil); err != nil {
-		log.E(ctx, "Command %v %v: %v", id, cmd, err)
+		// Continue the footprint building without emitting errors here. It is the
+		// following mutate() calls' responsibility to catch the error.
 		return
 	}
 

gapis/resolve/dependencygraph/footprint.go

@@ -161,7 +161,7 @@ func (f *Footprint) BehaviorIndex(ctx context.Context,
 		if u, ok := v.(uint64); ok {
 			return u
 		}
-		log.E(ctx, "Invalid behavior index: %v is not a uint64", v)
+		log.E(ctx, "Invalid behavior index: %v is not a uint64. Request command index: %v", v, fci)
 		return uint64(0)
 	}
 	log.E(ctx, "Cannot get behavior index for command indexed with: %v", fci)
@@ -229,8 +229,10 @@ func (r *FootprintResolvable) Resolve(ctx context.Context) (interface{}, error)
 				// side effect of the this command.
 				if err := cmd.Mutate(ctx, id, s, nil); err != nil {
 					bh.Aborted = true
+					// Continue the footprint building even if errors are found. It is
+					// following mutate calls, which are to build the replay
+					// instructions, that are responsible to catch the error.
 					// TODO: This error should be moved to report view.
-					return fmt.Errorf("Command %v %v: %v", id, cmd, err)
 				}
 				ft.AddBehavior(ctx, bh)
 				return nil