using experimental intrinsics failed

Project Infrastructure LLVM Dev List Archives
#1

Hi:

Sorry I need to send email directly. I am new to llvm and trying to write interval arithmetic, which requires changing rounding mode during computation.

The document I found https://llvm.org/docs/LangRef.html#constrained-floating-point-intrinsics, seems to be doing the trick. Here is the piece of code that I did:

%5 = call double @llvm.experimental.constrained.fadd(double %0, double %2, metadata !"round.downward", metadata !"fpexcept.ignore")

It is taking the first and third argument and adding them together with the downward rounding mode. However, when I try to convert it to bit code, I got the error:

error: use of undefined value '@llvm.experimental.constrained.fadd'

Is there any flags I need to add in order to use this intrinsic? Or this is just a wrong function call. I am using the llvm version 10.0.1.

Thank you
Xuan Tang

#2

I think you need a declaration like this at module scope in your IR file. Everything that’s called either needs to be defined or declared even if it is an intrinsic. It doesn’t matter if the declaration is before or after the use.

declare double @llvm.experimental.constrained.fadd(double, double, metadata", metadata)

#3

Hi Craig:

I tried that, now the function is like this:
; Function Attrs: norecurse nounwind readnone ssp uwtable
define { double, double } @add(double, double, double, double) local_unnamed_addr #0 {
%5 = call double @llvm.experimental.constrained.fadd(double %0, double %2, metadata !"round.downward", metadata !"fpexcept.ignore")
%6 = fadd double %1, %3
%7 = insertvalue { double, double } undef, double %5, 0
%8 = insertvalue { double, double } %7, double %6, 1
ret { double, double } %8
}

declare double @llvm.experimental.constrained.fadd(double, double, metadata, metadata)

But now I get the error at the line of declaration:
error: expected function name

I tried the example at https://llvm.org/docs/LangRef.html#variable-argument-handling-intrinsics and that compiles fine, which means the declaration code should work as well.

Thank you
Xuan Tang

#4

I just pasted that code and ran it through llc, opt, and llvm-as. All passed without issue. What version of llvm are you using?

#5

Hi Craig:

It’s llvm 10.1 I think, I did the same thing, removing everything except the code block but still failed with llvm-as.

Thank you
Xuan Tang

#6

Maybe also useful to add im using clang version 11.0.3 on macOS

Thank you
Xuan Tang

#7

Hi Craig:

I found the issue, the space before the metadata is somehow a tab in my original code, and that caused the problem.

I guess we should close the issue.

Thank you
Xuan Tang

#8

Note that the intrinsics don’t affect the rounding mode at run time.

What they do is the tell llvm what rounding mode llvm should expect the code to be using when it does eventually get run. You’ll need to change the rounding mode in your fp environment yourself by calling fpsetround() or whatever your system uses.

Also, note that the work to implement the constrained intrinsics is ongoing. Some hosts support it (X86, SystemZ, and recently PowerPC), but roughly no optimizations are enabled yet. So don’t expect performance yet.

I apologize for being behind in my email.