[PowerPC] Disable CTR loops containing floating point conversion on half-precision

[vchuravy]

Summary

Follow-up on https://reviews.llvm.org/rG0a19fc3088f5 and fixes https://bugs.llvm.org/show_bug.cgi?id=48519
for fpext and fptrunc.

[nemanjai]

I think we should just fix all this together. Can you please review (and test) https://reviews.llvm.org/D94980?

[vchuravy]

Abandoned in favour of D94980

llvm/lib/Target/PowerPC/PPCTargetTransformInfo.cpp

if (ST->useSoftFloat()) {
       case Instruction::FPToSI:
       case Instruction::UIToFP:
       case Instruction::SIToFP:
       case Instruction::FCmp:
         return true;
       }
     }
 
+    if (!ST->isISA3_0()) {
+      switch (J->getOpcode()) {
+      case Instruction::FPTrunc:
+      case Instruction::FPExt: {
+        CastInst *CI = cast<CastInst>(J);
+        if (CI->getSrcTy()->getScalarType()->isHalfTy() ||
+            CI->getDestTy()->getScalarType()->isHalfTy())
+          return true;
+      }
+      }
+    }
+
     for (Value *Operand : J->operands())
       if (memAddrUsesCTR(Operand, TM, Visited))
         return true;
   }
 
   return false;
 }
 

llvm/test/CodeGen/PowerPC/pr48519.ll

L139:                                             ; preds = %L139, %top
   %.not101.not = icmp eq i64 %value_phi21, 0
   %5 = add i64 %value_phi21, 1
   br i1 %.not101.not, label %pass.1, label %L139
 
 pass.1:                                           ; preds = %L139
   unreachable
 }
 
+define void @julia__hypot_17() {
+; CHECK-LABEL: julia__hypot_17:
+; CHECK:       # %bb.0: # %top
+; CHECK-NEXT:    mflr r0
+; CHECK-NEXT:    .cfi_def_cfa_offset 48
+; CHECK-NEXT:    .cfi_offset lr, 16
+; CHECK-NEXT:    .cfi_offset r30, -16
+; CHECK-NEXT:    std r30, -16(r1) # 8-byte Folded Spill
+; CHECK-NEXT:    std r0, 16(r1)
+; CHECK-NEXT:    stdu r1, -48(r1)
+; CHECK-NEXT:    li r30, 3
+; CHECK-NEXT:    .p2align 5
+; CHECK-NEXT:  .LBB1_1: # %L57
+; CHECK-NEXT:    #
+; CHECK-NEXT:    addi r30, r30, -1
+; CHECK-NEXT:    cmpldi r30, 0
+; CHECK-NEXT:    beq cr0, .LBB1_3
+; CHECK-NEXT:  # %bb.2: # %L68
+; CHECK-NEXT:    #
+; CHECK-NEXT:    lhz r3, 0(0)
+; CHECK-NEXT:    bl __gnu_h2f_ieee
+; CHECK-NEXT:    nop
+; CHECK-NEXT:    fcmpu cr0, f1, f1
+; CHECK-NEXT:    bun cr0, .LBB1_1
+; CHECK-NEXT:  .LBB1_3: # %L78
+; CHECK-NEXT:    addi r1, r1, 48
+; CHECK-NEXT:    ld r0, 16(r1)
+; CHECK-NEXT:    ld r30, -16(r1) # 8-byte Folded Reload
+; CHECK-NEXT:    mtlr r0
+; CHECK-NEXT:    blr
+top:
+  br label %L57
+
+L57:                                              ; preds = %L68, %top
+  %value_phi117 = phi i64 [ %0, %L68 ], [ 2, %top ]
+  %exitcond = icmp eq i64 %value_phi117, 4
+  br i1 %exitcond, label %L78, label %L68
+
+L68:                                              ; preds = %L57
+  %0 = add nuw nsw i64 %value_phi117, 1
+  %1 = load half, half* null, align 2
+  %2 = fpext half %1 to float
+  %3 = fcmp uno float %2, 0.000000e+00
+  %4 = or i1 %3, false
+  br i1 %4, label %L57, label %L78
+
+L78:                                              ; preds = %L68, %L57
+  ret void
+}
+
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare { i64, i1 } @llvm.ssub.with.overflow.i64(i64, i64) #0
 
 attributes #0 = { nounwind }