Fix checkedInstructionOrNonGenerateAtUseSite.

Fix checkedInstructionOrNonGenerateAtUseSite.

BUG= http://dartbug.com/15720

R=sra@google.com

Committed: https://code.google.com/p/dart/source/detail?r=32074
Reverted: https://code.google.com/p/dart/source/detail?r=32080

Committed: https://code.google.com/p/dart/source/detail?r=34325
Reverted: https://code.google.com/p/dart/source/detail?r=34335

Committed: https://code.google.com/p/dart/source/detail?r=35966

[floitsch, 2014-01-07 10:49:05]

ping.

[floitsch, 2014-01-27 14:32:55]

ping.

[sra1, 2014-01-28 17:56:29]

lgtm

[floitsch, 2014-01-28 17:58:40]

Committed patchset #1 manually as r32074 (presubmit successful).

[floitsch, 2014-03-24 20:10:41]

Committed patchset #3 manually as r34325 (presubmit successful).

[floitsch, 2014-03-24 21:52:51]

Tried to commit again, and had to revert because of:

python tools/test.py -mrelease -cdart2js -rd8 --use-sdk --write-debug-log
--write-test-outcome-log --checked -t120
pkg/polymer_expressions/test/parser_test

Will debug tomorrow.

[floitsch, 2014-03-24 21:55:11]

On 2014/03/24 21:52:51, floitsch wrote:
> Tried to commit again, and had to revert because of:
> 
> python tools/test.py -mrelease -cdart2js -rd8 --use-sdk --write-debug-log
> --write-test-outcome-log --checked -t120
> pkg/polymer_expressions/test/parser_test
> 
> Will debug tomorrow.

Other failing tests:
python tools/test.py -mrelease -cdart2js -rd8 --use-sdk --write-debug-log
--write-test-outcome-log --checked -t120 language/type_conversion_ssa_test
python tools/test.py -mrelease -cdart2js -rd8 --use-sdk --write-debug-log
--write-test-outcome-log --checked -t120
pkg/collection/test/unmodifiable_collection_test

[floitsch, 2014-04-05 02:33:55]

PTAL.
I think I have finally cracked this.

[floitsch, 2014-04-05 03:09:30]

Never mind.
Just got two errors:
python tools/test.py -rd8 -cdart2js --use-sdk -mrelease -t60
pkg/crypto/test/hmac_sha256_test
python tools/test.py -rd8 -cdart2js --use-sdk -mrelease -t60
pkg/crypto/test/sha256_test

I will ping you again, once I have fixed it.

[floitsch, 2014-04-07 19:05:33]

Had to give up on the variable-allocator changes. The code-gen uses the
input-types to know what checks to generate. By changing the inputs the checks
weren't working anymore... :(

[floitsch, 2014-05-06 15:49:12]

Ping.

This CL adds some overhead to the rewriteToUseSelfAsInterceptor method (in the
interceptor_simplifier, but it keeps fixes the code so that our assumption, that
we always use the tightest type, holds.

[sra1, 2014-05-08 03:34:35]

LGTM,

I have a question, though...

The current simplification replaces HInterceptor.
It might make more sense simplify uses of HInterceptor individually, especially
for this simplification (use 'self').

Consider:

if (x is Foo) {
  ... x.length
}

The is-check might require the interceptor, so we would have

t1 = getInterceptor(x);
if (t1.$isFoo) {
  // x = HTypeKnown(x)
  ... t1.get$length(x);
}

We could simply the getter-receiver use independently, possibly reducing the
uses of the interceptor

if (getInterceptor(x).$isFoo) {
  // x = HTypeKnown(x)
  ... x.get$length(x)
}

If we used this more diffuse approach, the receiver 'x' would be the HCheck
(HTypeKnown), not the input to HInterceptor.  Would it also solve the problem?

https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
File sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart
(right):

https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart:265:
// receiver which was the input to the interceptor.
Please add a test/guard against 'use' being an argument rather than the
receiver.  It would be relatively simple to change the calls to getInterceptor
in the runtime library to an intrinsic that generates HInterceptor and maybe
benefit from GVN.

[floitsch, 2014-05-08 18:21:24]

On 2014/05/08 03:34:35, sra1 wrote:
> LGTM,
> 
> I have a question, though...
> 
> The current simplification replaces HInterceptor.
> It might make more sense simplify uses of HInterceptor individually,
especially
> for this simplification (use 'self').
> 
> Consider:
> 
> if (x is Foo) {
>   ... x.length
> }
> 
> The is-check might require the interceptor, so we would have
> 
> t1 = getInterceptor(x);
> if (t1.$isFoo) {
>   // x = HTypeKnown(x)
>   ... t1.get$length(x);
> }
> 
> We could simply the getter-receiver use independently, possibly reducing the
> uses of the interceptor
> 
> if (getInterceptor(x).$isFoo) {
>   // x = HTypeKnown(x)
>   ... x.get$length(x)
> }
> 
> If we used this more diffuse approach, the receiver 'x' would be the HCheck
> (HTypeKnown), not the input to HInterceptor.  Would it also solve the problem?
I think so.
> 
>
https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
> File sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart
> (right):
> 
>
https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
> sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart:265:
> // receiver which was the input to the interceptor.
> Please add a test/guard against 'use' being an argument rather than the
> receiver.  It would be relatively simple to change the calls to getInterceptor
> in the runtime library to an intrinsic that generates HInterceptor and maybe
> benefit from GVN.

[floitsch, 2014-05-08 18:21:56]

PTAL. (I'm not 100% sure I understood your comment).

https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
File sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart
(right):

https://codereview.chromium.org/123423002/diff/170001/sdk/lib/_internal/compi...
sdk/lib/_internal/compiler/implementation/ssa/interceptor_simplifier.dart:265:
// receiver which was the input to the interceptor.
On 2014/05/08 03:34:35, sra1 wrote:
> Please add a test/guard against 'use' being an argument rather than the
> receiver.  It would be relatively simple to change the calls to getInterceptor
> in the runtime library to an intrinsic that generates HInterceptor and maybe
> benefit from GVN.

Done.

[sra1, 2014-05-08 19:42:31]

lgtm

[floitsch, 2014-05-09 09:45:52]

Committed patchset #8 manually as r35966 (presubmit successful).