I found a concerning issue when running with the latest Fortran compiler version (IFX 2025.2.0 20250605) involving the use of both the -fp-model=strict and -fimf-arch-consistency=true options. Compiling with both of these options on source files with certain floating point math operations results in IFX crashing with: "error #5633: **Internal compiler error: segmentation violation signal raised**". Below are a few sample cases of source files I was able to make that reproduce the issue:
SUBROUTINE case_1
INTEGER (KIND = 4) :: x
REAL (KIND = 8) :: res
x = 5
res = 2.0D0**REAL(x)
END SUBROUTINE case_1
SUBROUTINE case_2
REAL (KIND = 4) :: res
res = SQRT(1.40)
END SUBROUTINE case_2
SUBROUTINE case_3
REAL (KIND = 4) :: res
res = SIN(1.40)
END SUBROUTINE case_3
I'm assuming this issue extends to other common mathematical operations/functions in the math library, though I didn't see the issue for basic arithmetic operators (+, -, *, /). All I did to get the ICE for these sample cases was to run the command: "ifx -c source_file_name.F90 -fp-model=strict -fimf-arch-consistency=true". Running with only one option or the other doesn't produce the error. Value-safe optimization and consistency in floating point math is important for certain use cases, so I believe this issue should be looked into. In the meantime, I will look at using IFX 2025.1 instead.
