Crash in ARM backend for fmaxnum/fminnum

[Quuxplusone]

Bugzilla Link	PR35372
Status	NEW
Importance	P enhancement
Reported by	Stephen Hines (srhines@google.com)
Reported on	2017-11-20 17:27:53 -0800
Last modified on	2017-11-21 12:18:21 -0800
Version	trunk
Hardware	PC Linux
CC	diana.picus@linaro.org, kristof.beyls@arm.com, llvm-bugs@lists.llvm.org, pirama@google.com, srhines@google.com, yvan.roux@linaro.org
Fixed by commit(s)
Attachments	bugpoint-reduced-simplified.ll (2123 bytes, text/plain)
Blocks
Blocked by
See also

Created attachment 19451
reduced ll file

The attached test case breaks for fmaxnum. We also tested a few variations in
this source file. Changing the call to fminnum also will crash similarly. If we
remove the fp-armv8 flag, then the bug goes away (but the new instructions
aren't used either). There are also slightly different crashes if you change a
constant value from -3.276800e+04 to -3.376800e+04, but it isn't obvious if
this is just a different manifestation of the same bug.

$ llc bugpoint-reduced-simplified.ll
LLVM ERROR: Cannot select: t12: v4f32 = fmaxnum t22, undef:v4f32
  t22: v4f32 = bitcast t21
    t21: v4i32 = ARMISD::VMOVIMM TargetConstant:i32<1735>
      t20: i32 = TargetConstant<1735>
  t11: v4f32 = undef
In function: _ZN7android11volumeMultiILi4ELi8EsffisEEvPT1_jPKT2_PT4_PKT3_T5_

[Quuxplusone]

Attached bugpoint-reduced-simplified.ll (2123 bytes, text/plain): reduced ll file

[Quuxplusone]

I should note that part of the problem might be related to armv8 + "krait" (as opposed to "kryo"). Pirama noticed this particular aspect of the issue, so it is possible that the build misconfiguration is causing the problem. Still, I would expect that we get a proper error message for that misconfiguration, and not a crash.

[Quuxplusone]

Right, after changing "target-cpu"="krait" to "target-cpu"="kryo" in the function attributes, this no longer crashes.

krait seems to be a v7a core, kryo a v8a core.

I'm not sure what the root issue is here. Maybe it's the front-end generating a call to the llvm.maxnum intrinsic (see https://llvm.org/docs/LangRef.html#llvm-maxnum-intrinsic ) which may not be supported when targeting pre-v8a?

[Quuxplusone]

My guess is that clang was called with something like:

clang -target armv7--linux-android -mcpu=krait -mfpu=neon-fp-armv8 ...

Otherwise I don't see how target-features can contain "...,+fp-armv8,...+neon,+vfp4,..." in section attributes #0, given that krait cpu is indeed defined as an ARMv7a core with VPFv4 features.

I'm still too new in llvm to say where and how incoherent options has to be handled, but it seems to be the issue for me here (fmaxnum/fminnum are used to generate vmaxnm/vminnm instructions which are only available on ARMv8 cores, see http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0802b/CIHFCJCF.html).