Cannot split vector result of AVX intrinsic _mm256_rsqrt_ps

Project Infrastructure LLVM Dev List Archives
#1

I'm getting this on LLVM trunk:

SplitVectorResult #0: 0x27e6250: v8f32 = llvm.x86.avx.rsqrt.ps.256
0x2739310, 0x2739420 [ORD=16] [ID=0]
LLVM ERROR: Do not know how to split the result of this operator!
clang: error: linker command failed with exit code 1 (use -v to see invocation)

Oddly, when I build the same code without -flto I don't see this issue.

I see a similar bug was reported over a year ago and has had no
activity (though it doesn't use LTO):
http://llvm.org/bugs/show_bug.cgi?id=17909

#2

This sounds like the somehow the frontend is allowing the intrinsic to be used but the backend thinks it not legal and is trying to split the vector type to a legal type.

Can you provide your source and command line.

#3

Steven’s problem sounds similar to:
http://llvm.org/bugs/show_bug.cgi?id=19416

The cpu/feature flags aren’t being passed through correctly with -flto?

In bug 17909, the run script shows “-target-cpu core2” for code that includes AVX.

Is this expected behavior?

$ cat avx.ll
define <4 x double> @wrap_a_round(<4 x double> %a) {
%round = tail call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %a, i32 0)
ret <4 x double> %round
}

declare <4 x double> @llvm.x86.avx.round.pd.256(<4 x double>, i32)

$ ./llc -mcpu=core2 avx.ll
SplitVectorResult #0: 0x7fdc29849970: v4f64 = llvm.x86.avx.round.pd.256 0x7fdc29849750, 0x7fdc29849640, 0x7fdc29849860 [ORD=2] [ID=0]

LLVM ERROR: Do not know how to split the result of this operator!