mu/030container.cc

701 lines
24 KiB
C++
Raw Normal View History

//: Containers contain a fixed number of elements of different types.
2015-04-17 18:22:59 +00:00
:(before "End Mu Types Initialization")
2016-09-07 05:11:03 +00:00
//: We'll use this container as a running example in scenarios below.
type_ordinal point = put(Type_ordinal, "point", Next_type_ordinal++);
2016-03-28 17:11:23 +00:00
get_or_insert(Type, point); // initialize
get(Type, point).kind = CONTAINER;
get(Type, point).name = "point";
get(Type, point).elements.push_back(reagent("x:number"));
get(Type, point).elements.push_back(reagent("y:number"));
2015-02-20 07:49:13 +00:00
//: Containers can be copied around with a single instruction just like
//: numbers, no matter how large they are.
2016-05-07 17:33:53 +00:00
//: Tests in this layer often explicitly set up memory before reading it as a
2018-06-16 05:16:09 +00:00
//: container. Don't do this in general. I'm tagging such cases with /unsafe;
//: they'll be exceptions to later checks.
2015-02-20 07:49:13 +00:00
:(scenario copy_multiple_locations)
def main [
1:num <- copy 34
2:num <- copy 35
2016-01-12 06:57:35 +00:00
3:point <- copy 1:point/unsafe
2015-02-20 07:49:13 +00:00
]
2015-03-24 06:59:59 +00:00
+mem: storing 34 in location 3
+mem: storing 35 in location 4
2015-02-20 08:03:47 +00:00
2015-10-01 20:43:32 +00:00
//: trying to copy to a differently-typed destination will fail
2015-08-07 20:01:49 +00:00
:(scenario copy_checks_size)
% Hide_errors = true;
def main [
2:point <- copy 1:num
2015-08-07 20:01:49 +00:00
]
+error: main: can't copy '1:num' to '2:point'; types don't match
2015-08-07 20:01:49 +00:00
2015-03-27 04:06:14 +00:00
:(before "End Mu Types Initialization")
2016-09-07 05:11:03 +00:00
// A more complex example container, containing another container as one of
// its elements.
type_ordinal point_number = put(Type_ordinal, "point-number", Next_type_ordinal++);
2016-03-28 17:11:23 +00:00
get_or_insert(Type, point_number); // initialize
get(Type, point_number).kind = CONTAINER;
get(Type, point_number).name = "point-number";
get(Type, point_number).elements.push_back(reagent("xy:point"));
get(Type, point_number).elements.push_back(reagent("z:number"));
2015-03-27 04:06:14 +00:00
2015-04-24 17:19:03 +00:00
:(scenario copy_handles_nested_container_elements)
def main [
12:num <- copy 34
13:num <- copy 35
14:num <- copy 36
2016-01-12 06:57:35 +00:00
15:point-number <- copy 12:point-number/unsafe
2015-03-27 04:06:14 +00:00
]
+mem: storing 36 in location 17
//: products of recipes can include containers
:(scenario return_container)
def main [
3:point <- f 2
]
def f [
12:num <- next-ingredient
13:num <- copy 35
return 12:point/raw
]
+run: result 0 is [2, 35]
+mem: storing 2 in location 3
+mem: storing 35 in location 4
//: Containers can be checked for equality with a single instruction just like
//: numbers, no matter how large they are.
:(scenario compare_multiple_locations)
def main [
1:num <- copy 34 # first
2:num <- copy 35
3:num <- copy 36
4:num <- copy 34 # second
5:num <- copy 35
6:num <- copy 36
2016-09-17 07:46:03 +00:00
7:bool <- equal 1:point-number/raw, 4:point-number/unsafe
]
+mem: storing 1 in location 7
2015-08-09 19:26:31 +00:00
:(scenario compare_multiple_locations_2)
def main [
1:num <- copy 34 # first
2:num <- copy 35
3:num <- copy 36
4:num <- copy 34 # second
5:num <- copy 35
6:num <- copy 37 # different
2016-09-17 07:46:03 +00:00
7:bool <- equal 1:point-number/raw, 4:point-number/unsafe
]
+mem: storing 0 in location 7
:(before "End size_of(type) Special-cases")
if (type->value == -1) return 1; // error value, but we'll raise it elsewhere
if (type->value == 0) return 1;
if (!contains_key(Type, type->value)) {
raise << "no such type " << type->value << '\n' << end();
return 0;
}
type_info t = get(Type, type->value);
2015-10-27 03:06:51 +00:00
if (t.kind == CONTAINER) {
// size of a container is the sum of the sizes of its elements
int result = 0;
for (int i = 0; i < SIZE(t.elements); ++i) {
// todo: strengthen assertion to disallow mutual type recursion
if (t.elements.at(i).type->value == type->value) {
raise << "container " << t.name << " can't include itself as a member\n" << end();
return 0;
}
result += size_of(element_type(type, i));
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
}
3663 - fix a refcounting bug: '(type)' != 'type' This was a large commit, and most of it is a follow-up to commit 3309, undoing what is probably the final ill-considered optimization I added to s-expressions in Mu: I was always representing (a b c) as (a b . c), etc. That is now gone. Why did I need to take it out? The key problem was the error silently ignored in layer 30. That was causing size_of("(type)") to silently return garbage rather than loudly complain (assuming 'type' was a simple type). But to take it out I had to modify types_strictly_match (layer 21) to actually strictly match and not just do a prefix match. In the process of removing the prefix match, I had to make extracting recipe types from recipe headers more robust. So far it only matched the first element of each ingredient's type; these matched: (recipe address:number -> address:number) (recipe address -> address) I didn't notice because the dotted notation optimization was actually representing this as: (recipe address:number -> address number) --- One final little thing in this commit: I added an alias for 'assert' called 'assert_for_now', to indicate that I'm not sure something's really an invariant, that it might be triggered by (invalid) user programs, and so require more thought on error handling down the road. But this may well be an ill-posed distinction. It may be overwhelmingly uneconomic to continually distinguish between model invariants and error states for input. I'm starting to grow sympathetic to Google Analytics's recent approach of just banning assertions altogether. We'll see..
2016-11-11 05:39:02 +00:00
return result;
}
:(scenario stash_container)
def main [
1:num <- copy 34 # first
2:num <- copy 35
3:num <- copy 36
2015-08-24 20:40:21 +00:00
stash [foo:], 1:point-number/raw
]
+app: foo: 34 35 36
//:: To access elements of a container, use 'get'
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
//: 'get' takes a 'base' container and an 'offset' into it and returns the
//: appropriate element of the container value.
2015-04-24 17:19:03 +00:00
:(scenario get)
def main [
12:num <- copy 34
13:num <- copy 35
15:num <- get 12:point/raw, 1:offset # unsafe
2015-03-27 04:06:14 +00:00
]
+mem: storing 35 in location 15
:(before "End Primitive Recipe Declarations")
GET,
2015-02-20 08:03:47 +00:00
:(before "End Primitive Recipe Numbers")
put(Recipe_ordinal, "get", GET);
2015-10-01 20:43:32 +00:00
:(before "End Primitive Recipe Checks")
2015-02-20 08:03:47 +00:00
case GET: {
2015-10-01 20:43:32 +00:00
if (SIZE(inst.ingredients) != 2) {
2017-05-26 23:43:18 +00:00
raise << maybe(get(Recipe, r).name) << "'get' expects exactly 2 ingredients in '" << to_original_string(inst) << "'\n" << end();
2015-10-01 20:43:32 +00:00
break;
}
2016-05-06 07:46:39 +00:00
reagent/*copy*/ base = inst.ingredients.at(0); // new copy for every invocation
// Update GET base in Check
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
if (!base.type) {
raise << maybe(get(Recipe, r).name) << "first ingredient of 'get' should be a container, but got '" << inst.ingredients.at(0).original_string << "'\n" << end();
break;
}
const type_tree* base_type = base.type;
// Update GET base_type in Check
if (!base_type->atom || base_type->value == 0 || !contains_key(Type, base_type->value) || get(Type, base_type->value).kind != CONTAINER) {
raise << maybe(get(Recipe, r).name) << "first ingredient of 'get' should be a container, but got '" << inst.ingredients.at(0).original_string << "'\n" << end();
2015-10-01 20:43:32 +00:00
break;
}
2016-05-06 07:46:39 +00:00
const reagent& offset = inst.ingredients.at(1);
2015-10-01 20:43:32 +00:00
if (!is_literal(offset) || !is_mu_scalar(offset)) {
raise << maybe(get(Recipe, r).name) << "second ingredient of 'get' should have type 'offset', but got '" << inst.ingredients.at(1).original_string << "'\n" << end();
2015-07-24 08:05:59 +00:00
break;
}
int offset_value = 0;
if (is_integer(offset.name)) {
offset_value = to_integer(offset.name);
}
// End update GET offset_value in Check
if (offset_value < 0 || offset_value >= SIZE(get(Type, base_type->value).elements)) {
raise << maybe(get(Recipe, r).name) << "invalid offset '" << offset_value << "' for '" << get(Type, base_type->value).name << "'\n" << end();
2015-11-02 02:24:17 +00:00
break;
}
2015-11-16 02:36:34 +00:00
if (inst.products.empty()) break;
2016-05-06 07:46:39 +00:00
reagent/*copy*/ product = inst.products.at(0);
// Update GET product in Check
//: use base.type rather than base_type because later layers will introduce compound types
const reagent/*copy*/ element = element_type(base.type, offset_value);
if (!types_coercible(product, element)) {
raise << maybe(get(Recipe, r).name) << "'get " << base.original_string << ", " << offset.original_string << "' should write to " << names_to_string_without_quotes(element.type) << " but '" << product.name << "' has type " << names_to_string_without_quotes(product.type) << '\n' << end();
2015-10-06 00:02:32 +00:00
break;
}
2015-10-01 20:43:32 +00:00
break;
}
:(before "End Primitive Recipe Implementations")
case GET: {
2016-05-06 07:46:39 +00:00
reagent/*copy*/ base = current_instruction().ingredients.at(0);
// Update GET base in Run
int base_address = base.value;
if (base_address == 0) {
2017-05-26 23:43:18 +00:00
raise << maybe(current_recipe_name()) << "tried to access location 0 in '" << to_original_string(current_instruction()) << "'\n" << end();
break;
}
const type_tree* base_type = base.type;
// Update GET base_type in Run
int offset = ingredients.at(1).at(0);
if (offset < 0 || offset >= SIZE(get(Type, base_type->value).elements)) break; // copied from Check above
2018-06-16 05:16:09 +00:00
int src = base_address;
for (int i = 0; i < offset; ++i)
src += size_of(element_type(base.type, i));
2015-10-29 19:09:23 +00:00
trace(9998, "run") << "address to copy is " << src << end();
//: use base.type rather than base_type because later layers will introduce compound types
reagent/*copy*/ element = element_type(base.type, offset);
2016-04-20 16:58:07 +00:00
element.set_value(src);
trace(9998, "run") << "its type is " << names_to_string(element.type) << end();
// Read element
products.push_back(read_memory(element));
2015-02-20 08:03:47 +00:00
break;
}
:(code)
2016-04-30 17:09:38 +00:00
const reagent element_type(const type_tree* type, int offset_value) {
assert(offset_value >= 0);
const type_tree* base_type = type;
// Update base_type in element_type
assert(contains_key(Type, base_type->value));
assert(!get(Type, base_type->value).name.empty());
const type_info& info = get(Type, base_type->value);
assert(info.kind == CONTAINER);
2016-06-18 00:22:15 +00:00
if (offset_value >= SIZE(info.elements)) return reagent(); // error handled elsewhere
2016-05-06 07:46:39 +00:00
reagent/*copy*/ element = info.elements.at(offset_value);
// End element_type Special-cases
return element;
}
2015-04-24 17:19:03 +00:00
:(scenario get_handles_nested_container_elements)
def main [
12:num <- copy 34
13:num <- copy 35
14:num <- copy 36
15:num <- get 12:point-number/raw, 1:offset # unsafe
2015-02-22 08:15:14 +00:00
]
2015-03-24 06:59:59 +00:00
+mem: storing 36 in location 15
2015-02-22 08:15:14 +00:00
:(scenario get_out_of_bounds)
% Hide_errors = true;
def main [
12:num <- copy 34
13:num <- copy 35
14:num <- copy 36
get 12:point-number/raw, 2:offset # point-number occupies 3 locations but has only 2 fields; out of bounds
]
+error: main: invalid offset '2' for 'point-number'
2015-08-09 19:26:31 +00:00
:(scenario get_out_of_bounds_2)
% Hide_errors = true;
def main [
12:num <- copy 34
13:num <- copy 35
14:num <- copy 36
get 12:point-number/raw, -1:offset
]
+error: main: invalid offset '-1' for 'point-number'
:(scenario get_product_type_mismatch)
% Hide_errors = true;
def main [
12:num <- copy 34
13:num <- copy 35
14:num <- copy 36
15:&:num <- get 12:point-number/raw, 1:offset
]
+error: main: 'get 12:point-number/raw, 1:offset' should write to number but '15' has type (address number)
2015-11-16 02:36:34 +00:00
//: we might want to call 'get' without saving the results, say in a sandbox
:(scenario get_without_product)
def main [
12:num <- copy 34
13:num <- copy 35
2015-11-16 02:36:34 +00:00
get 12:point/raw, 1:offset # unsafe
]
# just don't die
//:: To write to elements of containers, use 'put'.
:(scenario put)
def main [
12:num <- copy 34
13:num <- copy 35
$clear-trace
12:point <- put 12:point, 1:offset, 36
]
+mem: storing 36 in location 13
-mem: storing 34 in location 12
:(before "End Primitive Recipe Declarations")
PUT,
:(before "End Primitive Recipe Numbers")
put(Recipe_ordinal, "put", PUT);
:(before "End Primitive Recipe Checks")
case PUT: {
if (SIZE(inst.ingredients) != 3) {
2017-05-26 23:43:18 +00:00
raise << maybe(get(Recipe, r).name) << "'put' expects exactly 3 ingredients in '" << to_original_string(inst) << "'\n" << end();
break;
}
2016-05-06 07:46:39 +00:00
reagent/*copy*/ base = inst.ingredients.at(0);
// Update PUT base in Check
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
if (!base.type) {
raise << maybe(get(Recipe, r).name) << "first ingredient of 'put' should be a container, but got '" << inst.ingredients.at(0).original_string << "'\n" << end();
break;
}
const type_tree* base_type = base.type;
// Update PUT base_type in Check
if (!base_type->atom || base_type->value == 0 || !contains_key(Type, base_type->value) || get(Type, base_type->value).kind != CONTAINER) {
raise << maybe(get(Recipe, r).name) << "first ingredient of 'put' should be a container, but got '" << inst.ingredients.at(0).original_string << "'\n" << end();
break;
}
2016-05-06 07:46:39 +00:00
reagent/*copy*/ offset = inst.ingredients.at(1);
// Update PUT offset in Check
if (!is_literal(offset) || !is_mu_scalar(offset)) {
raise << maybe(get(Recipe, r).name) << "second ingredient of 'put' should have type 'offset', but got '" << inst.ingredients.at(1).original_string << "'\n" << end();
break;
}
int offset_value = 0;
//: later layers will permit non-integer offsets
if (is_integer(offset.name)) {
offset_value = to_integer(offset.name);
if (offset_value < 0 || offset_value >= SIZE(get(Type, base_type->value).elements)) {
raise << maybe(get(Recipe, r).name) << "invalid offset '" << offset_value << "' for '" << get(Type, base_type->value).name << "'\n" << end();
break;
}
}
else {
offset_value = offset.value;
}
2016-05-06 07:46:39 +00:00
const reagent& value = inst.ingredients.at(2);
//: use base.type rather than base_type because later layers will introduce compound types
const reagent& element = element_type(base.type, offset_value);
if (!types_coercible(element, value)) {
raise << maybe(get(Recipe, r).name) << "'put " << base.original_string << ", " << offset.original_string << "' should write to " << names_to_string_without_quotes(element.type) << " but '" << value.name << "' has type " << names_to_string_without_quotes(value.type) << '\n' << end();
break;
}
if (inst.products.empty()) break; // no more checks necessary
if (inst.products.at(0).name != inst.ingredients.at(0).name) {
raise << maybe(get(Recipe, r).name) << "product of 'put' must be first ingredient '" << inst.ingredients.at(0).original_string << "', but got '" << inst.products.at(0).original_string << "'\n" << end();
break;
}
// End PUT Product Checks
break;
}
:(before "End Primitive Recipe Implementations")
case PUT: {
2016-05-06 07:46:39 +00:00
reagent/*copy*/ base = current_instruction().ingredients.at(0);
// Update PUT base in Run
int base_address = base.value;
if (base_address == 0) {
2017-05-26 23:43:18 +00:00
raise << maybe(current_recipe_name()) << "tried to access location 0 in '" << to_original_string(current_instruction()) << "'\n" << end();
break;
}
const type_tree* base_type = base.type;
// Update PUT base_type in Run
int offset = ingredients.at(1).at(0);
if (offset < 0 || offset >= SIZE(get(Type, base_type->value).elements)) break; // copied from Check above
2018-06-16 05:16:09 +00:00
int address = base_address;
for (int i = 0; i < offset; ++i)
address += size_of(element_type(base.type, i));
trace(9998, "run") << "address to copy to is " << address << end();
// optimization: directly write the element rather than updating 'product'
// and writing the entire container
// Write Memory in PUT in Run
write_products = false;
2016-10-20 05:10:35 +00:00
for (int i = 0; i < SIZE(ingredients.at(2)); ++i) {
trace("mem") << "storing " << no_scientific(ingredients.at(2).at(i)) << " in location " << address+i << end();
put(Memory, address+i, ingredients.at(2).at(i));
}
break;
}
:(scenario put_product_error)
% Hide_errors = true;
def main [
local-scope
load-ingredients
1:point <- merge 34, 35
3:point <- put 1:point, x:offset, 36
]
+error: main: product of 'put' must be first ingredient '1:point', but got '3:point'
2016-10-22 23:56:07 +00:00
//:: Allow containers to be defined in Mu code.
2015-05-14 17:30:01 +00:00
:(scenarios load)
:(scenario container)
container foo [
x:num
y:num
2015-05-14 17:30:01 +00:00
]
2015-10-29 18:56:10 +00:00
+parse: --- defining container foo
+parse: element: {x: "number"}
+parse: element: {y: "number"}
2015-05-14 17:30:01 +00:00
2015-05-18 23:09:09 +00:00
:(scenario container_use_before_definition)
container foo [
x:num
2015-05-18 23:09:09 +00:00
y:bar
]
container bar [
x:num
y:num
2015-05-18 23:09:09 +00:00
]
2015-10-29 18:56:10 +00:00
+parse: --- defining container foo
2015-05-18 23:09:09 +00:00
+parse: type number: 1000
+parse: element: {x: "number"}
# todo: brittle
# type bar is unknown at this point, but we assign it a number
+parse: element: {y: "bar"}
# later type bar geon
2015-10-29 18:56:10 +00:00
+parse: --- defining container bar
2015-05-18 23:09:09 +00:00
+parse: type number: 1001
+parse: element: {x: "number"}
+parse: element: {y: "number"}
2015-05-18 23:09:09 +00:00
//: if a container is defined again, the new fields add to the original definition
:(scenarios run)
:(scenario container_extend)
container foo [
x:num
]
# add to previous definition
container foo [
y:num
]
def main [
1:num <- copy 34
2:num <- copy 35
3:num <- get 1:foo, 0:offset
4:num <- get 1:foo, 1:offset
]
+mem: storing 34 in location 3
+mem: storing 35 in location 4
2015-05-14 17:30:01 +00:00
:(before "End Command Handlers")
else if (command == "container") {
2015-10-27 03:06:51 +00:00
insert_container(command, CONTAINER, in);
2015-05-14 17:30:01 +00:00
}
//: Even though we allow containers to be extended, we don't allow this after
//: a call to transform_all. But we do want to detect this situation and raise
//: an error. This field will help us raise such errors.
:(before "End type_info Fields")
int Num_calls_to_transform_all_at_first_definition;
:(before "End type_info Constructor")
Num_calls_to_transform_all_at_first_definition = -1;
2015-05-14 17:30:01 +00:00
:(code)
void insert_container(const string& command, kind_of_type kind, istream& in) {
skip_whitespace_but_not_newline(in);
2015-05-14 17:30:01 +00:00
string name = next_word(in);
if (name.empty()) {
assert(!has_data(in));
raise << "incomplete container definition at end of file (0)\n" << end();
return;
}
// End container Name Refinements
2015-10-29 18:56:10 +00:00
trace(9991, "parse") << "--- defining " << command << ' ' << name << end();
if (!contains_key(Type_ordinal, name)
|| get(Type_ordinal, name) == 0) {
put(Type_ordinal, name, Next_type_ordinal++);
2015-05-14 17:30:01 +00:00
}
trace("parse") << "type number: " << get(Type_ordinal, name) << end();
2016-06-18 00:22:15 +00:00
skip_bracket(in, "'"+command+"' must begin with '['");
type_info& info = get_or_insert(Type, get(Type_ordinal, name));
if (info.Num_calls_to_transform_all_at_first_definition == -1) {
// initial definition of this container
info.Num_calls_to_transform_all_at_first_definition = Num_calls_to_transform_all;
}
else if (info.Num_calls_to_transform_all_at_first_definition != Num_calls_to_transform_all) {
// extension after transform_all
raise << "there was a call to transform_all() between the definition of container '" << name << "' and a subsequent extension. This is not supported, since any recipes that used '" << name << "' values have already been transformed and \"frozen\".\n" << end();
return;
}
info.name = name;
info.kind = kind;
while (has_data(in)) {
2015-05-14 17:30:01 +00:00
skip_whitespace_and_comments(in);
string element = next_word(in);
if (element.empty()) {
assert(!has_data(in));
raise << "incomplete container definition at end of file (1)\n" << end();
return;
}
2015-05-14 17:30:01 +00:00
if (element == "]") break;
2016-06-18 00:22:15 +00:00
if (in.peek() != '\n') {
raise << command << " '" << name << "' contains multiple elements on a single line. Containers and exclusive containers must only contain elements, one to a line, no code.\n" << end();
// skip rest of container declaration
while (has_data(in)) {
skip_whitespace_and_comments(in);
if (next_word(in) == "]") break;
}
break;
}
info.elements.push_back(reagent(element));
expand_type_abbreviations(info.elements.back().type); // todo: use abbreviation before declaration
replace_unknown_types_with_unique_ordinals(info.elements.back().type, info);
trace(9993, "parse") << " element: " << to_string(info.elements.back()) << end();
// End Load Container Element Definition
2015-05-14 17:30:01 +00:00
}
}
void replace_unknown_types_with_unique_ordinals(type_tree* type, const type_info& info) {
if (!type) return;
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
if (!type->atom) {
replace_unknown_types_with_unique_ordinals(type->left, info);
replace_unknown_types_with_unique_ordinals(type->right, info);
return;
}
assert(!type->name.empty());
if (contains_key(Type_ordinal, type->name)) {
type->value = get(Type_ordinal, type->name);
}
// End insert_container Special-cases
else if (type->name != "->") { // used in recipe types
put(Type_ordinal, type->name, Next_type_ordinal++);
type->value = get(Type_ordinal, type->name);
}
}
2015-10-06 01:49:21 +00:00
void skip_bracket(istream& in, string message) {
skip_whitespace_and_comments(in);
if (in.get() != '[')
2016-02-26 21:04:55 +00:00
raise << message << '\n' << end();
2015-10-06 01:49:21 +00:00
}
2016-06-18 00:22:15 +00:00
:(scenario multi_word_line_in_container_declaration)
% Hide_errors = true;
container foo [
x:num y:num
2016-06-18 00:22:15 +00:00
]
+error: container 'foo' contains multiple elements on a single line. Containers and exclusive containers must only contain elements, one to a line, no code.
//: support type abbreviations in container definitions
:(scenario type_abbreviations_in_containers)
type foo = number
container bar [
x:foo
]
def main [
1:num <- copy 34
2:foo <- get 1:bar/unsafe, 0:offset
]
+mem: storing 34 in location 2
:(after "Transform.push_back(expand_type_abbreviations)")
Transform.push_back(expand_type_abbreviations_in_containers); // idempotent
:(code)
// extremely inefficient; we process all types over and over again, once for every single recipe
// but it doesn't seem to cause any noticeable slowdown
void expand_type_abbreviations_in_containers(const recipe_ordinal /*unused*/) {
2016-10-20 05:10:35 +00:00
for (map<type_ordinal, type_info>::iterator p = Type.begin(); p != Type.end(); ++p) {
for (int i = 0; i < SIZE(p->second.elements); ++i)
expand_type_abbreviations(p->second.elements.at(i).type);
}
}
2016-09-12 02:55:28 +00:00
//: ensure scenarios are consistent by always starting new container
//: declarations at the same type number
2017-07-09 21:34:17 +00:00
:(before "End Reset") //: for tests
Next_type_ordinal = 1000;
2015-05-14 17:30:01 +00:00
:(before "End Test Run Initialization")
assert(Next_type_ordinal < 1000);
2015-05-14 17:30:01 +00:00
:(code)
void test_error_on_transform_all_between_container_definition_and_extension() {
// define a container
run("container foo [\n"
" a:num\n"
"]\n");
// try to extend the container after transform
transform_all();
2016-11-16 05:55:56 +00:00
CHECK_TRACE_DOESNT_CONTAIN_ERRORS();
Hide_errors = true;
run("container foo [\n"
" b:num\n"
"]\n");
2016-11-16 05:55:56 +00:00
CHECK_TRACE_CONTAINS_ERRORS();
}
//:: Allow container definitions anywhere in the codebase, but complain if you
//:: can't find a definition at the end.
:(scenario run_complains_on_unknown_types)
% Hide_errors = true;
def main [
# integer is not a type
1:integer <- copy 0
]
2015-11-17 19:28:19 +00:00
+error: main: unknown type integer in '1:integer <- copy 0'
:(scenario run_allows_type_definition_after_use)
def main [
1:bar <- copy 0/unsafe
]
container bar [
x:num
]
$error: 0
:(before "End Type Modifying Transforms")
Transform.push_back(check_or_set_invalid_types); // idempotent
:(code)
void check_or_set_invalid_types(const recipe_ordinal r) {
2015-11-19 18:29:55 +00:00
recipe& caller = get(Recipe, r);
2016-02-15 07:18:33 +00:00
trace(9991, "transform") << "--- check for invalid types in recipe " << caller.name << end();
2016-10-20 05:10:35 +00:00
for (int index = 0; index < SIZE(caller.steps); ++index) {
2015-11-19 18:29:55 +00:00
instruction& inst = caller.steps.at(index);
2016-10-20 05:10:35 +00:00
for (int i = 0; i < SIZE(inst.ingredients); ++i)
check_or_set_invalid_types(inst.ingredients.at(i), caller, inst);
2016-10-20 05:10:35 +00:00
for (int i = 0; i < SIZE(inst.products); ++i)
check_or_set_invalid_types(inst.products.at(i), caller, inst);
}
2015-11-19 18:29:55 +00:00
// End check_or_set_invalid_types
}
void check_or_set_invalid_types(reagent& r, const recipe& caller, const instruction& inst) {
// Begin check_or_set_invalid_types(r)
2017-05-26 23:43:18 +00:00
check_or_set_invalid_types(r.type, maybe(caller.name), "'"+to_original_string(inst)+"'");
}
void check_or_set_invalid_types(type_tree* type, const string& location_for_error_messages, const string& name_for_error_messages) {
if (!type) return;
// End Container Type Checks
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
if (!type->atom) {
check_or_set_invalid_types(type->left, location_for_error_messages, name_for_error_messages);
check_or_set_invalid_types(type->right, location_for_error_messages, name_for_error_messages);
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
return;
}
2016-01-19 00:46:07 +00:00
if (type->value == 0) return;
if (!contains_key(Type, type->value)) {
assert(!type->name.empty());
if (contains_key(Type_ordinal, type->name))
type->value = get(Type_ordinal, type->name);
else
raise << location_for_error_messages << "unknown type " << type->name << " in " << name_for_error_messages << '\n' << end();
}
}
:(scenario container_unknown_field)
% Hide_errors = true;
container foo [
x:num
y:bar
]
+error: foo: unknown type in y
:(scenario read_container_with_bracket_in_comment)
container foo [
x:num
# ']' in comment
y:num
]
2015-10-29 18:56:10 +00:00
+parse: --- defining container foo
+parse: element: {x: "number"}
+parse: element: {y: "number"}
2016-07-24 09:19:40 +00:00
:(scenario container_with_compound_field_type)
container foo [
2016-07-24 09:19:40 +00:00
{x: (address array (address array character))}
]
$error: 0
2016-03-28 05:03:13 +00:00
:(before "End transform_all")
check_container_field_types();
:(code)
void check_container_field_types() {
2016-10-20 05:10:35 +00:00
for (map<type_ordinal, type_info>::iterator p = Type.begin(); p != Type.end(); ++p) {
const type_info& info = p->second;
// Check Container Field Types(info)
2016-10-20 05:10:35 +00:00
for (int i = 0; i < SIZE(info.elements); ++i)
check_invalid_types(info.elements.at(i).type, maybe(info.name), info.elements.at(i).name);
}
}
void check_invalid_types(const type_tree* type, const string& location_for_error_messages, const string& name_for_error_messages) {
if (!type) return; // will throw a more precise error elsewhere
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
if (!type->atom) {
check_invalid_types(type->left, location_for_error_messages, name_for_error_messages);
check_invalid_types(type->right, location_for_error_messages, name_for_error_messages);
3309 Rip out everything to fix one failing unit test (commit 3290; type abbreviations). This commit does several things at once that I couldn't come up with a clean way to unpack: A. It moves to a new representation for type trees without changing the actual definition of the `type_tree` struct. B. It adds unit tests for our type metadata precomputation, so that errors there show up early and in a simpler setting rather than dying when we try to load Mu code. C. It fixes a bug, guarding against infinite loops when precomputing metadata for recursive shape-shifting containers. To do this it uses a dumb way of comparing type_trees, comparing their string representations instead. That is likely incredibly inefficient. Perhaps due to C, this commit has made Mu incredibly slow. Running all tests for the core and the edit/ app now takes 6.5 minutes rather than 3.5 minutes. == more notes and details I've been struggling for the past week now to back out of a bad design decision, a premature optimization from the early days: storing atoms directly in the 'value' slot of a cons cell rather than creating a special 'atom' cons cell and storing it on the 'left' slot. In other words, if a cons cell looks like this: o / | \ left val right ..then the type_tree (a b c) used to look like this (before this commit): o | \ a o | \ b o | \ c null ..rather than like this 'classic' approach to s-expressions which never mixes val and right (which is what we now have): o / \ o o | / \ a o o | / \ b o null | c The old approach made several operations more complicated, most recently the act of replacing a (possibly atom/leaf) sub-tree with another. That was the final straw that got me to realize the contortions I was going through to save a few type_tree nodes (cons cells). Switching to the new approach was hard partly because I've been using the old approach for so long and type_tree manipulations had pervaded everything. Another issue I ran into was the realization that my layers were not cleanly separated. Key parts of early layers (precomputing type metadata) existed purely for far later ones (shape-shifting types). Layers I got repeatedly stuck at: 1. the transform for precomputing type sizes (layer 30) 2. type-checks on merge instructions (layer 31) 3. the transform for precomputing address offsets in types (layer 36) 4. replace operations in supporting shape-shifting recipes (layer 55) After much thrashing I finally noticed that it wasn't the entirety of these layers that was giving me trouble, but just the type metadata precomputation, which had bugs that weren't manifesting until 30 layers later. Or, worse, when loading .mu files before any tests had had a chance to run. A common failure mode was running into types at run time that I hadn't precomputed metadata for at transform time. Digging into these bugs got me to realize that what I had before wasn't really very good, but a half-assed heuristic approach that did a whole lot of extra work precomputing metadata for utterly meaningless types like `((address number) 3)` which just happened to be part of a larger type like `(array (address number) 3)`. So, I redid it all. I switched the representation of types (because the old representation made unit tests difficult to retrofit) and added unit tests to the metadata precomputation. I also made layer 30 only do the minimal metadata precomputation it needs for the concepts introduced until then. In the process, I also made the precomputation more correct than before, and added hooks in the right place so that I could augment the logic when I introduced shape-shifting containers. == lessons learned There's several levels of hygiene when it comes to layers: 1. Every layer introduces precisely what it needs and in the simplest way possible. If I was building an app until just that layer, nothing would seem over-engineered. 2. Some layers are fore-shadowing features in future layers. Sometimes this is ok. For example, layer 10 foreshadows containers and arrays and so on without actually supporting them. That is a net win because it lets me lay out the core of Mu's data structures out in one place. But if the fore-shadowing gets too complex things get nasty. Not least because it can be hard to write unit tests for features before you provide the plumbing to visualize and manipulate them. 3. A layer is introducing features that are tested only in later layers. 4. A layer is introducing features with tests that are invalidated in later layers. (This I knew from early on to be an obviously horrendous idea.) Summary: avoid Level 2 (foreshadowing layers) as much as possible. Tolerate it indefinitely for small things where the code stays simple over time, but become strict again when things start to get more complex. Level 3 is mostly a net lose, but sometimes it can be expedient (a real case of the usually grossly over-applied term "technical debt"), and it's better than the conventional baseline of no layers and no scenarios. Just clean it up as soon as possible. Definitely avoid layer 4 at any time. == minor lessons Avoid unit tests for trivial things, write scenarios in context as much as possible. But within those margins unit tests are fine. Just introduce them before any scenarios (commit 3297). Reorganizing layers can be easy. Just merge layers for starters! Punt on resplitting them in some new way until you've gotten them to work. This is the wisdom of Refactoring: small steps. What made it hard was not wanting to merge *everything* between layer 30 and 55. The eventual insight was realizing I just need to move those two full-strength transforms and nothing else.
2016-09-10 01:32:52 +00:00
return;
}
if (type->value != 0) { // value 0 = compound types (layer parse_tree) or type ingredients (layer shape_shifting_container)
if (!contains_key(Type, type->value))
raise << location_for_error_messages << "unknown type in " << name_for_error_messages << '\n' << end();
}
}
string to_original_string(const type_ordinal t) {
ostringstream out;
if (!contains_key(Type, t)) return out.str();
const type_info& info = get(Type, t);
if (info.kind == PRIMITIVE) return out.str();
out << (info.kind == CONTAINER ? "container" : "exclusive-container") << " " << info.name << " [\n";
for (int i = 0; i < SIZE(info.elements); ++i) {
out << " " << info.elements.at(i).original_string << "\n";
}
out << "]\n";
return out.str();
}