binary16 float-point arithmetic crashes LLVM on s390x

[Zixing Liu]

binary16 float-point arithmetic crashes LLVM on s390x due to the LLVM backend being unable to lower the LLVM IR to the machine code.

You can find a short reproducer below:

define noundef zeroext i1 @test_func(half noundef %f) unnamed_addr {
start:
  %self = fdiv half %f, 0xHC700
  %_4 = bitcast half %self to i16
  %_0 = icmp slt i16 %_4, 0
  ret i1 %_0
}

You can save this LLVM IR text to a file called `test-ir.ll` and build it using `clang test-ir.ll -c -o test-ir.o` on a s390x device and watch LLVM explodes with:

warning: overriding the module target triple with s390x-unknown-linux-gnu [-Woverride-module]
fatal error: error in backend: Cannot select: 0x6005f020ae20: i32 = fp_to_fp16 0x6005f020adb0
  0x6005f020adb0: f32 = fdiv 0x6005f020af70, 0x6005f020b4b0
    0x6005f020af70: f32 = fp16_to_fp 0x6005f020b520
      0x6005f020b520: i32 = fp_to_fp16 0x6005f020a9c0
        0x6005f020a9c0: f32,ch = CopyFromReg 0x6005f01e5220, Register:f32 %0
          0x6005f020a950: f32 = Register %0
    0x6005f020b4b0: f32,ch = load<(load (s32) from constant-pool)> 0x6005f01e5220, 0x6005f020b280, undef:i64
      0x6005f020b280: i64 = SystemZISD::PCREL_WRAPPER TargetConstantPool:i64<float -7.000000e+00> 0
        0x6005f020b1a0: i64 = TargetConstantPool<float -7.000000e+00> 0
      0x6005f020b130: i64 = undef
In function: test_func

[Zixing Liu (liushuyu-011)]

After some digging, I think the s390x is missing a fallback ISel path for the f16 type. Here's my preliminary fix based on PPC64 logic (may not be entirely correct):

diff --git a/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp b/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
index 383393914a16..a2351b205bdb 100644
--- a/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
+++ b/llvm/lib/Target/SystemZ/SystemZISelLowering.cpp
@@ -529,6 +529,14 @@ SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &TM,
       setOperationAction(ISD::FSINCOS, VT, Expand);
       setOperationAction(ISD::FREM, VT, Expand);
       setOperationAction(ISD::FPOW, VT, Expand);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
+ setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
+ setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
+ setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
+ setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
+ setTruncStoreAction(MVT::f64, MVT::f16, Expand);
+ setTruncStoreAction(MVT::f32, MVT::f16, Expand);

       // Special treatment.
       setOperationAction(ISD::IS_FPCLASS, VT, Custom);

[Ubuntu Foundations Team Bug Bot (crichton)]

The attachment "0001-fix-f16-load-store-s390x.patch" seems to be a patch. If it isn't, please remove the "patch" flag from the attachment, remove the "patch" tag, and if you are a member of the ~ubuntu-reviewers, unsubscribe the team.

[This is an automated message performed by a Launchpad user owned by ~brian-murray, for any issues please contact him.]

[Zixing Liu (liushuyu-011)]

Rust upstream pointed me to https://github.com/llvm/llvm-project/issues/97981. It seems to affect PowerPC as well.

[Zixing Liu (liushuyu-011)]

Another report from LLVM: https://github.com/llvm/llvm-project/issues/50374