Hi Christoph,
I've got no experience with the decompiler at all. I had a look at your changes. It does what I expect it to do in the case I mentioned but it breaks when the variable is used more than once:
| x | x := 1. (x := 1) caseOf: { [ 2 ] -> [ 2 ]. [ 3 ] -> [ 3 ] } otherwise: [ x + 1 ]
will be decompiled as
x := 1. 1 caseOf: { [2] -> [2]. [3] -> [3]} otherwise: [:x | x + 1]]
Notice that x is not defined when it first assigned to in the decompiled code.
The same problem is present with #ifNil:#ifNotNil: too:
| x | x := 1. (x := 1) ifNil: [ 2 ] ifNotNil: [ x + 1 ]
is decompiled as
x := 1. 1 ifNil: [2] ifNotNil: [:x | x + 1]
with the Compiler-mt.462 as well.
I understand that the decompiler cannot be perfect without the compiler storing additional information in the bytecodes but I think it should be possible - to always generate the compact version with the block arguments unless the same variable is used outside of the expression's scope, and - to never generate invalid code.
Levente
On Mon, 22 Nov 2021, christoph.thiede@student.hpi.uni-potsdam.de wrote:
Hi Levente,
please see Compiler-ct.464 (inbox). To you as a more experienced decompiler developer, does this implementation look convincing? If yes, I can merge it myself now. :-)
Best, Christoph
Sent from Squeak Inbox Talk
On 2021-11-18T14:04:16+01:00, marcel.taeumel@hpi.de wrote:
Hi Levente --
Thanks. Decompiler only adds the block-arg for not-inlined caseOfs. See ObjectTest >> #testCaseOfOtherwise.
Hmm... Christoph? =) Otherwise (ha!) I will take a look at it soon-ish.
Best, Marcel Am 18.11.2021 13:51:02 schrieb Levente Uzonyi <leves at caesar.elte.hu>: Hi Marcel,
I just checked how this works, and noticed that the decompiler cannot recreate the original code if the otherwise: block has an argument.
Levente
On Thu, 18 Nov 2021, Marcel Taeumel wrote:
Merged.
Am 21.02.2020 17:33:03 schrieb Thiede, Christoph :
This changeset makes it possible to accept an argument in the otherwise block of a #caseOf:otherwise: call.
Diff of the #caseOf:otherwise: implementation:
caseOf: aBlockAssociationCollection otherwise: aBlock
"The elements of aBlockAssociationCollection are associations between blocks. Answer the evaluated value of the first association in aBlockAssociationCollection whose evaluated key equals the receiver. If no match is found, answer the result of evaluating aBlock."
aBlockAssociationCollection associationsDo:
[:assoc | (assoc key value = self) ifTrue: [^assoc value value]].
- ^ aBlock value
- ^ aBlock cull: self
"| z | z := {[#a]->[1+1]. ['b' asSymbol]->[2+2]. [#c]->[3+3]}. #b caseOf: z otherwise: [0]"
"| z | z := {[#a]->[1+1]. ['d' asSymbol]->[2+2]. [#c]->[3+3]}. #b caseOf: z otherwise: [0]"
"The following are compiled in-line:"
"#b caseOf: {[#a]->[1+1]. ['b' asSymbol]->[2+2]. [#c]->[3+3]} otherwise: [0]"
"#b caseOf: {[#a]->[1+1]. ['d' asSymbol]->[2+2]. [#c]->[3+3]} otherwise: [0]"
+ "#b caseOf: {[#a]->[1+1]. ['d' asSymbol]->[2+2]. [#c]->[3+3]} otherwise: [:x | x halt]"
Furthermore, the changeset includes a necessary modification of MessageNode >> #transformCase: so that the otherwise argument can be compiled in-line.
Last but not least, I wrote some tests for #caseOf:[otherwise:].
Please review!
(In a later change, it would be possible to allow arguments for the association key blocks as well. But I love short feedback loops, so let's assess this one first :-))
Best,
Christoph