[regression] lit/exec/table64.wast test fails on x86 32-bit

[rathann]

When running the tests of binaryen 126 on x86 32-bit (Fedora Rawhide), I'm getting a failure in lit/exec/table64.wast:

FAIL: Binaryen lit tests :: exec/table64.wast (179 of 825)
******************** TEST 'Binaryen lit tests :: exec/table64.wast' FAILED ********************
Exit Code: 1
Command Output (stdout):
--
# RUN: at line 3
/builddir/build/BUILD/binaryen-126-build/binaryen-version_126/redhat-linux-build/bin/wasm-opt /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/test/lit/exec/table64.wast -all --fuzz-exec -q -o /dev/null 2>&1 | filecheck /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/test/lit/exec/table64.wast
# executed command: /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/redhat-linux-build/bin/wasm-opt /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/test/lit/exec/table64.wast -all --fuzz-exec -q -o /dev/null
# executed command: filecheck /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/test/lit/exec/table64.wast
# .---command stdout------------
# | /builddir/build/BUILD/binaryen-126-build/binaryen-version_126/test/lit/exec/table64.wast:47:17: error: CHECK-NEXT: is not on the line after the previous match
# |  ;; CHECK-NEXT: [trap out of bounds table access]
# |                 ^
# | <stdin>:16:1: note: 'next' match was here
# | [trap out of bounds table access]
# | ^
# | <stdin>:7:29: note: previous match ended here
# | [fuzz-exec] calling oob-huge
# |                             ^
# | <stdin>:8:1: note: non-matching line after previous match is here
# | [fuzz-exec] note result: oob-huge => 1337
# | ^
# `-----------------------------
# error: command failed with exit status: 1
--
********************

The same test passes in binaryen 125.

[kripken]

Thanks for the report @rathann !

The failing test is

 (func $oob-huge (export "oob-huge") (result i32)
  ;; This call traps on oob with a value over 32 bits, 2**32 + 1, which if we
  ;; truncated to 32 bits, would seem in bounds, and end up calling a valid
  ;; function.
  (call_indirect (type $i32)
   (i64.add
    (i64.const 0x100000000)
    (i64.const 1)
   )
  )
 )

It ends up calling that valid function - so it isn't doing the oob check right, likely doing it on 32-bit and not 64.

@stevenfontanella @tlively I am pretty sure the problem is here:

binaryen/src/ir/runtime-table.cpp

Lines 43 to 44 in 9656b8e

	Literal RealRuntimeTable::get(std::size_t i) const {
	if (i >= table.size()) {

I am guessing all those size_ts should instead be Address? Address is a general address in wasm, while size_t depends on the host bit size.

[stevenfontanella]

Yeah that sounds likely, especially if this only showed up in the latest release. Will take a look and send a fix. Thanks for reporting!

[stevenfontanella]

I was able to repro with

cmake -G Ninja -DCMAKE_CXX_FLAGS="-m32" -DCMAKE_C_FLAGS="-m32" -DCMAKE_BUILD_TYPE=Debug .
ninja
python3 check.py lit

In addition to the reported test, passes/O_fast-math.wast and passes/optimize-instructions-mvp.wast also fail for me. It looks like these two give different NaNs than expected:

$ "filecheck" "/usr/local/google/home/stevenfont/code/binaryen/test/lit/passes/O_fast-math.wast"
# command output:
/usr/local/google/home/stevenfont/code/binaryen/test/lit/passes/O_fast-math.wast:41:17: error: CHECK-NEXT: expected string not found in input
 ;; CHECK-NEXT: (f32.const -nan:0x23017a)
                ^
<stdin>:19:1: note: scanning from here
  (f32.const -nan:0x63017a)
^

error: command failed with exit status: 1

(and similar for the other test).

Will work on fixing table64.wast for now.

[stevenfontanella]

Filed another issue for the remaining failures: #8366