r345395 ([X86][SSE] Move 2-input limit up from getFauxShuffleMask to resolveTargetShuffleInputs) introduced a new compiler failure.  This was later reverted in r345451 due to a report of new failures, but without a testcase just yet.

I'm not sure if this is the same failure that was observed, but we found this in our testing so I'm reporting it, as presumably we'll want to re-land r345395 in some form at some point.


$ e:\work\upstream-llvm\build-vs2015-native-ninja\bin\clang.exe --version
clang version 8.0.0 (trunk 345380) (llvm/trunk 345395)
Target: x86_64-pc-windows-msvc
Thread model: posix
InstalledDir: e:\work\upstream-llvm\build-vs2015-native-ninja\bin

$ cat 1.cpp
struct a {
  float e;
  float f;
  float h;
};
float b(a &c, a &d) { return c.e * d.e + c.f * d.f + c.h * d.h; }
float g[6];
int j;
a *k;
void fn2(a &c) {
  for (int i = 0; i < j; ++i)
    g[i] = b(k[i], c);
}

$ e:\work\upstream-llvm\build-vs2015-native-ninja\bin\clang.exe -c 1.cpp -O2 -march=btver2
Assertion failed: 0 <= InputIdx1 && "Unknown target shuffle input", file E:\work\upstream-llvm\llvm\lib\Target\X86\X86ISelLowering.cpp, line 30877

I would love to provide a reduced .ll input, but bugpoint has decided so far not to be cooperative today.  Regardless, the above should be easily usable for reproducing.