[PowerPC] Fix assertion due to assuming a type is simple.

[sfertile]

Summary

The CTRLoop pass performs checks on the argument of certain libcalls/intrinsics, and assumes the arguments must be of a simple type. This isn't always the case though. For example if we unroll and vectorize a loop we may end up with vectors larger then the largest legal type, along with intrinsics that operate on those wider types. This happened in the ffmpeg build, where we unrolled a loop and ended up with a sqrt intrinsic that operated on V16f64, triggering an assertion.

[nemanjai]

OK, I don't think there's an issue with this - if the vector just happens to be wider, we'll end up needing more of the respective instructions, but it won't turn anything into a library call that wouldn't be a library call otherwise.
LGTM.

lib/Target/PowerPC/PPCCTRLoops.cpp

407 ↗

(On Diff #128709)

Nit: I imagine it was probably clang-format that formatted this so, but I would say that putting the entire initialization expression on a separate line is probably more readable.

[hfinkel]

lib/Target/PowerPC/PPCCTRLoops.cpp

409 ↗

(On Diff #128709)

You should just be able to write:

if (EVTy == MVT::Other)

(otherwise, this LGTM too)

llvm/trunk/lib/Target/PowerPC/PPCCTRLoops.cpp

#endif
           case LibFunc_fmax:
           case LibFunc_fmaxf:
           case LibFunc_fmaxl:
             Opcode = ISD::FMAXNUM; break;
           }
         }
 
         if (Opcode) {
-          MVT VTy = TLI->getSimpleValueType(
-              *DL, CI->getArgOperand(0)->getType(), true);
-          if (VTy == MVT::Other)
+          EVT EVTy =
+              TLI->getValueType(*DL, CI->getArgOperand(0)->getType(), true);
+
+          if (EVTy == MVT::Other)
             return true;
 
-          if (TLI->isOperationLegalOrCustom(Opcode, VTy))
+          if (TLI->isOperationLegalOrCustom(Opcode, EVTy))
             continue;
-          else if (VTy.isVector() &&
-                   TLI->isOperationLegalOrCustom(Opcode, VTy.getScalarType()))
+          else if (EVTy.isVector() &&
+                   TLI->isOperationLegalOrCustom(Opcode, EVTy.getScalarType()))
             continue;
 
           return true;
         }
       }
 
       return true;
     } else if (isa<BinaryOperator>(J) &&

llvm/trunk/test/CodeGen/PowerPC/non-simple-args-intrin.ll

+; RUN: llc -verify-machineinstrs <%s | FileCheck %s
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le-unknown-linux-gnu"
+
+; Ensure that that the  CTRLoop pass can compile intrinsics with
+; non-simple arguments. eg: @llvm.sqrt.v16f64.
+
+; Function Attrs: nounwind
+define void @filter_prewitt() {
+; CHECK-LABEL: filter_prewitt:
+entry:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %entry
+  %wide.load = load <16 x i8>, <16 x i8>* undef, align 1, !tbaa !1, !alias.scope !4
+  %0 = zext <16 x i8> %wide.load to <16 x i32>
+  %wide.load279 = load <16 x i8>, <16 x i8>* undef, align 1, !tbaa !1, !alias.scope !4
+  %1 = zext <16 x i8> %wide.load279 to <16 x i32>
+  %2 = add nuw nsw <16 x i32> %1, %0
+  %3 = add nuw nsw <16 x i32> %2, zeroinitializer
+  %4 = sub nsw <16 x i32> zeroinitializer, %3
+  %5 = add nsw <16 x i32> %4, zeroinitializer
+  %6 = add nsw <16 x i32> %5, zeroinitializer
+  %7 = sub nsw <16 x i32> zeroinitializer, %0
+  %8 = sub nsw <16 x i32> %7, zeroinitializer
+  %9 = add nsw <16 x i32> %8, zeroinitializer
+  %10 = sub nsw <16 x i32> %9, zeroinitializer
+  %11 = add nsw <16 x i32> %10, zeroinitializer
+  %12 = mul nsw <16 x i32> %6, %6
+  %13 = mul nsw <16 x i32> %11, %11
+  %14 = add nuw nsw <16 x i32> %13, %12
+  %15 = sitofp <16 x i32> %14 to <16 x double>
+  %16 = call nsz <16 x double> @llvm.sqrt.v16f64(<16 x double> %15)
+  %17 = fmul nsz <16 x double> %16, undef
+  %18 = fadd nsz <16 x double> %17, undef
+  %19 = fptosi <16 x double> %18 to <16 x i32>
+  %20 = sub nsw <16 x i32> zeroinitializer, %19
+  %21 = ashr <16 x i32> %20, <i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31>
+  %22 = select <16 x i1> undef, <16 x i32> %21, <16 x i32> %19
+  %23 = trunc <16 x i32> %22 to <16 x i8>
+  store <16 x i8> %23, <16 x i8>* undef, align 1, !tbaa !1, !alias.scope !7, !noalias !9
+  br label %vector.body
+}
+
+; Function Attrs: nounwind readnone speculatable
+declare <16 x double> @llvm.sqrt.v16f64(<16 x double>) #1
+
+attributes #1 = { nounwind readnone speculatable }
+
+!1 = !{!2, !2, i64 0}
+!2 = !{!"omnipotent char", !3, i64 0}
+!3 = !{!"Simple C/C++ TBAA"}
+!4 = !{!5}
+!5 = distinct !{!5, !6}
+!6 = distinct !{!6, !"LVerDomain"}
+!7 = !{!8}
+!8 = distinct !{!8, !6}
+!9 = !{!10, !11, !5}
+!10 = distinct !{!10, !6}
+!11 = distinct !{!11, !6}