[llvm-bugs] [Bug 39982] New: __attribute__((pcs("aapcs-vfp"))) is not consistent

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=39982

            Bug ID: 39982
           Summary: __attribute__((pcs("aapcs-vfp"))) is not consistent
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: ARM
          Assignee: unassignedbugs at nondot.org
          Reporter: srhines at google.com
                CC: kristof.beyls at arm.com, llvm-bugs at lists.llvm.org,
                    peter.smith at linaro.org, Ties.Stuij at arm.com

This is from an internal Google bug, but really has to do with ARM backend
handling of aapcs-vfp. From that bug:

float __attribute__((pcs("aapcs-vfp"))) baz(float x, float y) {
    return y;
}

struct S {
     float f;
     float d;
};

float __attribute__((pcs("aapcs-vfp"))) foo(S s) {
    return s.d;
}

float __attribute__((pcs("aapcs"))) bar(S s) {
    return s.d;
}


Compilation in hard-fp mode produces expected output (
https://godbolt.org/z/SdXIsN ):

baz(float, float):
  vmov.f32 s0, s1
  bx lr
foo(S):
  vmov.f32 s0, s1
  bx lr
bar(S):
  mov r0, r1
  bx lr

Compilation in soft-fp mode produces very strange output (
https://godbolt.org/z/HVVDtj ):

baz(float, float):
  vmov.f32 s0, s1
  bx lr
foo(S):
  vmov s0, r1
  bx lr
bar(S):
  mov r0, r1
  bx lr

It looks as if "foo" function uses hard-fp mode for result but soft-fp for
input.

Unfortunatelly vulkan uses pcs("aapcs-vfp") thus we couldn't just ignore that
issue (even if it looks as if currently vulkan does not suffer from that issue
- there very few functions with float arguments and they all are compiled
identically in hardfp and pcs("aapcs-vfp") modes).

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20181212/c798b94a/attachment-0001.html>