When preparing arguments for a function call, the JIT must detect and
resolve dependencies (cycles) between registers. This occurs when,
for example, the value for the first argument is currently in the
register that should hold the second argument, and vice-versa.
The logic incorrectly used the register's 'kind' (the register class)
instead of its 'abi' (the specific data type/width) to identify the
source operand. This caused the dependency solver to fail or emit
incorrect moves on 64-bit platforms where floating-point arguments
are passed in registers.
This issue is "masked" on 32-bit x86 (IA32) because it primarily
uses the stack for passing double-precision arguments, avoiding
register-to-register swaps for the ABI registers.
The fix has been verified by adjusting 'tests/call_10_2.c' to force
a register swap (F0/F1). Formerly the modified test failed on:
- Linux x86_64 (System V ABI)
- Linux aarch64
- Windows 64-bit (Microsoft x64 ABI) via Cygwin, MINGW64, UCRT64,
and CLANG64.
* lightening/lightening.c (prepare_args): Use 'src[j].abi' instead of
'src[j].kind' when resolving register-to-register dependencies.
* tests/call_10_2.c (run_test_double): Swap the order in which dargs[0]
and dargs[1] are loaded into registers F0 and F1. This forces a
register cycle during the call, exposing the dependency resolution bug.
```
When preparing arguments for a function call, the JIT must detect and
resolve dependencies (cycles) between registers. This occurs when,
for example, the value for the first argument is currently in the
register that should hold the second argument, and vice-versa.
The logic incorrectly used the register's 'kind' (the register class)
instead of its 'abi' (the specific data type/width) to identify the
source operand. This caused the dependency solver to fail or emit
incorrect moves on 64-bit platforms where floating-point arguments
are passed in registers.
This issue is "masked" on 32-bit x86 (IA32) because it primarily
uses the stack for passing double-precision arguments, avoiding
register-to-register swaps for the ABI registers.
The fix has been verified by adjusting 'tests/call_10_2.c' to force
a register swap (F0/F1). Formerly the modified test failed on:
- Linux x86_64 (System V ABI)
- Linux aarch64
- Windows 64-bit (Microsoft x64 ABI) via Cygwin, MINGW64, UCRT64,
and CLANG64.
* lightening/lightening.c (prepare_args): Use 'src[j].abi' instead of
'src[j].kind' when resolving register-to-register dependencies.
* tests/call_10_2.c (run_test_double): Swap the order in which dargs[0]
and dargs[1] are loaded into registers F0 and F1. This forces a
register cycle during the call, exposing the dependency resolution bug.
```