Difference between revisions of "DIP83"

From D Wiki
Jump to: navigation, search
(References)
 
Line 224: Line 224:
 
</syntaxhighlight>
 
</syntaxhighlight>
  
This in order for comparison to be more visually helpful to developers.
+
This makes comparison more visually helpful to developers.
  
 
== References ==
 
== References ==

Latest revision as of 21:46, 24 December 2016

Title: Configurable Assert Diagnostics
DIP: 83
Version: 1
Status: Draft
Created: 2015-10-01
Last Modified: 2016-12-24
Author: Per Nordlöw
Links:

Abstract

Allow for assert to do pretty printing of its failing expression when flagged for in call to compiler. Printing is configurable via specific sets of (template) function overloads.

Rationale

A failing assert (in a unittest), currently, give no hint about why it failed. To aid the developer in debugging, the failing assert of a

  • binary expression, such as assert(x == y), should print the values of x and y and
  • unary expression, such as assert(!x), should print the value of x.

This extra, so called, pretty printing can be enabled by changing the dmd flag -unittest to, say, -unittest=verbose for a specific failing module. This will be more convenient than explicitly adding the prints of the left-hand-side expression lhs and right-hand-side expression rhs directly before the call to the failing assert.

Description

This DIP proposes to add library-level-configurable diagnostics to failing calls to assert(expr) typically called from within unittest-blocks.

This diagnostics is activated only when DMD is called with a specific command line flag, say -unittest=verbose or perhaps -diagnose=assert.

Modifications needed in Compiler Frontend

Binary Operator Lowering

If DMD is called with this flag it will rewrite (lower) all assert expressions (AssertExpr in dmd source) such as

assert(a BINOP b)

into

(auto ref a, auto ref b) {
    if (a BINOP b) return;
    onAssertFailed!"BINOP"(a, b, __FILE__, __LINE__, __FUNCTION__, __MODULE__);
} (e1, e2)

where onAssertFailed, in this binary operator case, is declared as

void onAssertFailed(string op, E1, E2)(E1 e1, E2 e2, string file, uint line, string function, string module);

Unary Operator Lowering

Similarly, for unary expressions, rewrite (lower)

assert(UNOP e)

with

(auto ref e) {
    if (UNOP e) return;
    onAssertFailed!"UNOP"(e, __FILE__, __LINE__, __FUNCTION__, __MODULE__);
} (e)

where on onAssertFailed, in this unary operator case, is declared as

void onAssertFailed(string op, E)(E e, string file, uint line, string function, string module);

Non-Operator Lowering

For the case when no unary or binary operator is present in the top assert expression, unary overload of onAssertFailed is called with op being empty string. In other words the assert expression

assert(e)

is rewritten (lowered) into

(auto ref e) {
    if (e) return;
    onAssertFailed!""(e, __FILE__, __LINE__, __FUNCTION__, __MODULE__);
} (e)

If the assert lowering must also be sensitive to expressions like

Non-Equality Operator Lowering

Further note that operator != may need special care because D only supports overloading via opUnary!"!" and opBinary!"==". This because the D compiler rewrites

x != y

into

!(x == y)

To make assert diagnostics as configurable as possible it may become relevant to enable the possibility for the developer to to define the overload onAssertFailed!"!=" aswell. If this is desirable the lowering logic in DMD must be sensitive to expressions

assert(!(a == b))

which, preferrably, will be rewritten to

(auto ref a, auto ref b) {
    if (!(a == b)) return;
    onAssertFailed!"!="(a, b, __FILE__, __LINE__, __FUNCTION__, __MODULE__);
} (e1, e2)

Modifications needed in D Runtime

Configuring Diagnostics

The default implementations (in druntime) of all the onAssertFailed-overloads should, as a first step, by defined to exactly mimic the current behaviour of assert(). This because, memory requirements of the call to DMD will increase when the compiler must generate terminal printing-code of all the arguments for all the unittests found in the compilation unit of interest. And it is currently unclear if this, current increase in system requirements, will be compatible with system resources currently available in build servers connected to GitHub.

Specific printing behaviour of assert() diagnostics can then be extendable by adding (typically templated) overloads of onAssertFailed for specific sets of types (concepts).

For instance, diagnostics specifically when comparing arithmetic types, could be realized through the overload

import std.traits : isArithmetic;
void onAssertFailed(string op, E1, E2)(E1 e1, E2 e2, string file, uint line, string function, string module)
    if (isArithmetic!L && 
        isArithmetic!R)
{
    version(assert)
    {
        import core.exception : AssertError;
        throw AssertError("Failed arithmetic assert: " ~ e1 ~ " " ~ binOp ~ " " ~ e2);
    }
}

In this way D would get the extendability we want in testing-frameworks such as std.experimental.testing (https://github.com/D-Programming-Language/phobos/pull/3207) without adding a new assert-overload-set and without sacrifycing default memory usage in DMD/Phobos unittests.

Further, this solution enables the possibility to provide fancy diagnostics behaviour in onAssertFailed for failing array/range or aggregate (struct or class) comparisons. This diagnostics could also have different pretty printing backends such as HTML.

Examples of Possible Diagnostics Outputs

Short Array Assert Arguments

For example a failing

assert([1,2,3] == [1,2,4]);

could pretty-print

([1,2,3][2] is 3) != ([1,2,4][2] is 4)

Aggregate Assert Arguments

or, for aggregates, a failing

struct A { int x, y; }
auto a = A(1,2);
auto b = A(1,3);
assert(a == b);

could pretty-print

(a.y is 2) != (b.y is 3)

Large Array Assert Arguments

or, for very large arrays, a failing

const n = 1_000_000;
auto a = iota(0, n).array;
auto b = a.dup;
a[1_000 .. 1_002] = 0;
assert(a == b);

could pretty-print

(a[1_000 .. 1_002] is [0, 0]) != (b[1_000 .. 1_002] is [1_000, 1_001])

An optional multiLine flag could in this case be motivated. When set this flag changes output to

(a[1_000 .. 1_002] is [0, 0]) !=
(b[1_000 .. 1_002] is [1_000, 1_001])

This makes comparison more visually helpful to developers.

References

Parts of the solution list at

https://issues.dlang.org/show_bug.cgi?id=5547#c3

including rewriting/expansion of AssertExpr could probably reused.

https://issues.dlang.org/show_bug.cgi?id=15889 - similar idea for bound checking

Discussions on topic

Copyright

This document has been placed in the Public Domain.