#4066 tests/gcc-torture-execute-20011109-1.c fals for -mmcs51 --model-small --stack-auto --opt-code-speed --nolospre --nolabelopt --nosidechannels --max-allocs-per-node 512

[Philipp Klaus Krause]

To reproduce:

philipp@notebook7:~/sdcc-trunk/sdcc/support/regression$ make random-test-1594146079855823397 
Creating random-1594146079855823397 based on mcs51-small-stack-auto with additional options: --opt-code-speed --nolospre --nolabelopt --nosidechannels --max-allocs-per-node 512 
Running random-1594146079855823397 regression tests
Summary for 'random-1594146079855823397': 1 abnormal stops ( ), 0 failures, 27913 tests, 6279 test cases, 10437537 bytes, 8537655576 ticks
   abnormal stop: gcc-torture-execute-20011109-1.c

I see the bug on two Debian GNU/Linux systems: Debian testing on amd64, Debian unstable on powerpc64.

[Maarten Brock]

My first guess is that this is a stack overflow.

[Li-Hsin Chien]

I did some investigation on this issue.
I removed the additional flags one by one, and the problem can be reproduced with only --opt-code-speed and --nosidechannels.

The test mainly contains two switch statements, so I checked the generated assembly.
Both switches use jump tables, and I noticed that peephole rules 260.i and 257.b are applied.

Rule 260.i changes the jump table entries to 2-byte sjmp instructions.
However, rule 257.b later changes some of them back to 3-byte ljmp instructions.
I think this breaks the entry size of the jump table.

I added notInJumpTable() to rule 257.b, and the original test passes with this change.

I submitted [patches:#514] for review: https://sourceforge.net/p/sdcc/patches/514/
Please let me know if this makes sense.

[Maarten Brock]

Hi @janet-chien
Can you please try to use [patches:#nnn] instead of a pasted URL next time? It gives the benefit of linking tracker items together.

[Li-Hsin Chien]

Thanks, I’ll use [patches:#nnn] next time.

I also just noticed that you had already fixed the format for me on my previous patch.
Sorry I didn’t notice that earlier, and thank you!

[Maarten Brock]

To be able to create a regression test for this bug it would help if there was a pragma to set no_side_channels.

[Philipp Klaus Krause]

There is.

#pragma nosidechannels

[Li-Hsin Chien]

Thanks, I will add a regression test using #pragma nosidechannels.

[Li-Hsin Chien]

I have added the regression test in [patches:#514].

It is based on the same switch statements from
gcc-torture-execute-20011109-1.c, with

#pragma opt_code_speed
#pragma nosidechannels

I tested it without the notInJumpTable() change, and the test failed.
With the fix applied, the test passed.
Thanks!

[Maarten Brock]

Fixed in [r16836] by applying [patches:#514].