After extruding an edge 3 times with E, I could reproduce the issue on 5300ba0ba50 and:
Windows 10 Pro 64 bit (1803)
AMD Radeon RX480 4GB with Radeon Software 18.7.1

I used a debug build, here's the backtrace.

In summary the problem is in the update_mesh_edit_mode_pointers and discard_mesh_edit_mode_pointers functions of the deg_eval_copy_on_write.cc file.

A copy of the edit_btmesh from mesh_orig is done in the mesh_cow. So the same looptris pointer is used on both meshes.

But when doing an extrusion, the number of looptris changes and the looptris are reallocated in the original mesh causing the mesh_cow pointer to be invalid.

The looptris are reallocated in the BKE_editmesh_tessface_calc function.

Issue is in fact exact same one as with undo/redo: in macros, when next operator needs access to evaluated data, we have to ensure depsgraph is properly evaluated before calling that next operator:

diff --git a/source/blender/editors/transform/transform_ops.c b/source/blender/editors/transform/transform_ops.c
index 1d66616a038..4ff98ced9e5 100644
--- a/source/blender/editors/transform/transform_ops.c
+++ b/source/blender/editors/transform/transform_ops.c
@@ -640,7 +640,7 @@ static void TRANSFORM_OT_translate(struct wmOperatorType *ot)
 	ot->name   = "Move";
 	ot->description = "Move selected items";
 	ot->idname = OP_TRANSLATION;
-	ot->flag = OPTYPE_REGISTER | OPTYPE_UNDO | OPTYPE_BLOCKING;
+	ot->flag = OPTYPE_REGISTER | OPTYPE_UNDO | OPTYPE_BLOCKING | OPTYPE_USE_EVAL_DATA;
 
 	/* api callbacks */
 	ot->invoke = transform_invoke;
diff --git a/source/blender/windowmanager/intern/wm_operator_type.c b/source/blender/windowmanager/intern/wm_operator_type.c
index 89421ac0ab0..9f468825ae6 100644
--- a/source/blender/windowmanager/intern/wm_operator_type.c
+++ b/source/blender/windowmanager/intern/wm_operator_type.c
@@ -43,6 +43,8 @@
 #include "BKE_context.h"
 #include "BKE_idprop.h"
 #include "BKE_library.h"
+#include "BKE_main.h"
+#include "BKE_scene.h"
 
 #include "RNA_access.h"
 #include "RNA_define.h"
@@ -357,7 +359,7 @@ static int wm_macro_invoke_internal(bContext *C, wmOperator *op, const wmEvent *
 	int retval = OPERATOR_FINISHED;
 
 	/* start from operator received as argument */
-	for (; opm; opm = opm->next) {
+	while (opm) {
 		if (opm->type->invoke)
 			retval = opm->type->invoke(C, opm, event);
 		else if (opm->type->exec)
@@ -374,6 +376,16 @@ static int wm_macro_invoke_internal(bContext *C, wmOperator *op, const wmEvent *
 		else {
 			break; /* operator didn't finish, end macro */
 		}
+
+		opm = opm->next;
+		if (opm != NULL && opm->type->flag & OPTYPE_USE_EVAL_DATA) {
+			/* We need to force refresh of depsgraph after first operator,
+			 * since next one *may* rely on some valid evaluated data. */
+			Main *bmain = CTX_data_main(C);
+			Scene *scene = CTX_data_scene(C);
+			ViewLayer *view_layer = CTX_data_view_layer(C);
+			BKE_scene_view_layer_graph_evaluated_ensure(bmain, scene, view_layer);
+		}
 	}
 
 	return wm_macro_end(op, retval);

@Campbell Barton (campbellbarton) issue with patch above is that it needs to set whole transform operator as 'needing evaluated data', which is not a good idea in redo context (when tweaking a transform value in redo panel e.g.). I think we'll actually need to add a new 'flag_poll' callback to operatortype? That way, we could (optionally) return flags finely tuned to actual needs… E.g. in that case, only set OPTYPE_USE_EVAL_DATA when we actually do snapping on geometry, etc.

I don't really like the idea of keeping a looptris pointer for each object evaluated.
Since the bmesh does not change, these looptris would be exact copies and it would be necessary to make sure to recalculate the looptris in each access.

This is one of the reasons for the reported inefficiency of the transform operator in edit mode.
Since it is not easy to determine when the looptris pointer has been deallocated, the BKE_editmesh_tessface_calc function is always called to do a render_data with looptris in the draw_manager: draw_cache_impl_mesh.c 414.

Therefore the looptris are recalculated at least 3 times at each interaction. One in the operator transform to the original mesh, another in DRW_mesh_batch_cache_get_surface_shaded and one in DRW_mesh_batch_cache_get_overlay_triangles.

Since bmesh does not change, the looptris are the same and should be recalculated only once.

@Germano Cavalcante (mano-wii) that is kind of separated issue, not directly related to that bug imho.