[PowerPC] Prevent legalization loop from promoting SELECT_CC from v4i32 to v4i32

[nemanjai]

Summary

As reported in https://bugs.llvm.org/show_bug.cgi?id=45709 we can hit an infinite loop in legalization since we set the legalization action for ISD::SELECT_CC for all fixed length vector types to Promote. Without some different legalization action for the type being promoted to, the legalizer simply loops. Since we don't have patterns to match the node, the right legalization action should be Expand.

[adalava]

I confirmed it fixes the case Piotr Kubaj reported. Thank you!

[amyk]

llvm/test/CodeGen/PowerPC/pr45709.ll

6

This is kind of minor, but it may be good to document about the fix in the test case or give the function a more descriptive name.

[stefanp]

The name of the test is pr45709 which refers to the bug number so I guess anyone interested should look there for the details.
However, I still agree with Amy that adding a description inside the test would be useful since the code in the test is not self explanatory.
Otherwise, LGTM.

llvm/test/CodeGen/PowerPC/pr45709.ll

52

nit:
You can probably remove this comment if all we leave is nounwind.

[saghir]

LGTM other than the nits that Stefan mentioned.

[amyk]

I meant to approve this when I reviewed earlier. LGTM.

[nemanjai]

Add description to the test case as requested.

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
 
       for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
         setTruncStoreAction(VT, InnerVT, Expand);
         setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
         setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
         setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
       }
     }
+    setOperationAction(ISD::SELECT_CC, MVT::v4i32, Expand);
     if (!Subtarget.hasP8Vector()) {
       setOperationAction(ISD::SMAX, MVT::v2i64, Expand);
       setOperationAction(ISD::SMIN, MVT::v2i64, Expand);
       setOperationAction(ISD::UMAX, MVT::v2i64, Expand);
       setOperationAction(ISD::UMIN, MVT::v2i64, Expand);
     }
 
     for (auto VT : {MVT::v2i64, MVT::v4i32, MVT::v8i16, MVT::v16i8})

llvm/test/CodeGen/PowerPC/pr45709.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-unknown \
+; RUN:     -mcpu=pwr6 -ppc-asm-full-reg-names -mattr=-vsx \
+; RUN:     -ppc-vsr-nums-as-vr < %s | FileCheck %s
+
+; There is code in the SDAG to expand FMAX/FMIN with fast flags to SELECT_CC.
+; On PPC, we had SELECT_CC legalized using Promote for all vector types
+; (including the type that they are all promoted to - which caused an infinite
+; loop in legalization). This test just ensures that we terminate on such input.
+define dso_local void @_ZN1a1bEv(<4 x float> %in) local_unnamed_addr #0 align 2 {
+; CHECK-LABEL: _ZN1a1bEv:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    bclr 12, 4*cr5+lt, 0
+; CHECK-NEXT:  # %bb.1: # %.preheader
+; CHECK-NEXT:    addis r3, r2, .LCPI0_0@toc@ha
+; CHECK-NEXT:    vxor v3, v3, v3
+; CHECK-NEXT:    addi r3, r3, .LCPI0_0@toc@l
+; CHECK-NEXT:    lvx v4, 0, r3
+; CHECK-NEXT:    addi r3, r1, -48
+; CHECK-NEXT:    stvx v3, 0, r3
+; CHECK-NEXT:    addi r3, r1, -32
+; CHECK-NEXT:    vperm v2, v2, v2, v4
+; CHECK-NEXT:    stvx v2, 0, r3
+; CHECK-NEXT:    lwz r3, -48(r1)
+; CHECK-NEXT:    lwz r4, -32(r1)
+; CHECK-NEXT:    cmpw r4, r3
+; CHECK-NEXT:    bc 12, gt, .LBB0_2
+; CHECK-NEXT:    b .LBB0_3
+; CHECK-NEXT:  .LBB0_2: # %.preheader
+; CHECK-NEXT:    addi r3, r4, 0
+; CHECK-NEXT:  .LBB0_3: # %.preheader
+; CHECK-NEXT:    stw r3, -64(r1)
+; CHECK-NEXT:    addi r3, r1, -64
+; CHECK-NEXT:    lvx v2, 0, r3
+; CHECK-NEXT:    addi r3, r1, -16
+; CHECK-NEXT:    stvx v2, 0, r3
+; CHECK-NEXT:    blr
+  br i1 undef, label %7, label %1
+
+1:                                                ; preds = %1, %0
+  br i1 undef, label %2, label %1
+
+2:                                                ; preds = %1
+  %3 = shufflevector <4 x float> %in, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 1, i32 0>
+  %4 = call fast <4 x float> @llvm.maxnum.v4f32(<4 x float> %3, <4 x float> zeroinitializer)
+  %5 = call fast <4 x float> @llvm.maxnum.v4f32(<4 x float> %4, <4 x float> undef)
+  %6 = extractelement <4 x float> %5, i32 0
+  br label %7
+
+7:                                                ; preds = %2, %0
+  %8 = phi float [ %6, %2 ], [ undef, %0 ]
+  %9 = fcmp fast une float %8, 0.000000e+00
+  ret void
+}
+
+declare <4 x float> @llvm.maxnum.v4f32(<4 x float>, <4 x float>) #0
+
+attributes #0 = { nounwind }