Internal error in Constant.hs

[tmcdonell]

I'm getting the following error when attempting moduleAST:

*** Exception: src/LLVM/General/Internal/Constant.hs:(197,13)-(239,14): Non-exhaustive patterns in case

The module in question that is hitting this problem is the math library that you need to link against when targeting PTX. I've uploaded a copy here for convenience, but you can also get it by installing the CUDA toolkit available from NVIDIA here (version 7.5).

Here is a small example to trigger the error:

import LLVM.General
import LLVM.General.Context
import Control.Monad.Except

exceptError :: Show e => ExceptT e IO a -> IO a
exceptError = either (error . show) return <=< runExceptT

main :: IO ()
main = do
  ll <- readFile "libdevice.compute_30.10.ll"

  withContext                                     $ \ctx -> do
  exceptError $ withModuleFromLLVMAssembly ctx ll $ \mdl -> do
    print =<< moduleAST mdl

Tested with llvm-general-3.5.1.1 and llvm-general-3.4.6.0.

[bscarlet]

That error should correlate with some specific construct in the input file. Could you try to narrow it down please, and attach it directly to this bug?

[tmcdonell]

This function will trigger it:

; ModuleID = 'libdevice.compute_30.10.bc'
target triple = "nvptx-unknown-unknown"

@"$str" = private addrspace(1) constant [11 x i8] c"__CUDA_FTZ\00"

; Function Attrs: alwaysinline inlinehint
define float @__nv_floorf(float %f) #0 {
  %call = call i32 @__nvvm_reflect(i8* addrspacecast (i8 addrspace(1)* getelementptr inbounds ([11 x i8] addrspace(1)* @"$str", i32 0, i32 0) to i8*))
  %1 = icmp ne i32 %call, 0
  br i1 %1, label %2, label %4

; <label>:2                                       ; preds = %0
  %3 = call float @llvm.nvvm.floor.ftz.f(float %f)
  br label %6

; <label>:4                                       ; preds = %0
  %5 = call float @llvm.nvvm.floor.f(float %f)
  br label %6

; <label>:6                                       ; preds = %4, %2
  %retval.0 = phi float [ %3, %2 ], [ %5, %4 ]
  ret float %retval.0
}

declare i32 @__nvvm_reflect(i8*)

; Function Attrs: nounwind readnone
declare float @llvm.nvvm.floor.ftz.f(float) #1

; Function Attrs: nounwind readnone
declare float @llvm.nvvm.floor.f(float) #1

attributes #1 = { nounwind readnone }

[tmcdonell]

Example pruned down slightly more. This is indeed the first call to __nvvm_reflect in the file.

; ModuleID = 'libdevice.compute_30.10.bc'
target triple = "nvptx-unknown-unknown"

@"$str" = private addrspace(1) constant [11 x i8] c"__CUDA_FTZ\00"

; Function Attrs: alwaysinline inlinehint
define float @__nv_floorf(float %f) #0 {
  %call = call i32 @__nvvm_reflect(i8* addrspacecast (i8 addrspace(1)* getelementptr inbounds ([11 x i8] addrspace(1)* @"$str", i32 0, i32 0) to i8*))

  ret float 0x00000000
}

declare i32 @__nvvm_reflect(i8*)

Apologies for not submitting the small test case initially.

[tmcdonell]

Not sure if this will help you, but in the last working version that I could find (CUDA 6.5) that call line instead looked like this:

  %1 = call i8* @llvm.nvvm.ptr.global.to.gen.p0i8.p1i8(i8 addrspace(1)* getelementptr inbounds ([11 x i8] addrspace(1)* @"$str", i32 0, i32 0))
  %call = call i32 @__nvvm_reflect(i8* %1)

plus:

declare i8* @llvm.nvvm.ptr.global.to.gen.p0i8.p1i8(i8 addrspace(1)*) #1

[bscarlet]

The problem is likely the addrspacecast instruction, or rather the corresponding constant expression. The instruction's new in this llvm 3.5.

[bscarlet]

Should be fixed in llvm-general-pure-3.5.1.0 / llvm-general-3.5.1.2 Please re-open if not.

[tmcdonell]

Everything seems to be working in the new version. Thanks!