[LegalizeDAG] Fix legalization of SETCC

[miyuki]

Summary

Currently when expanding a SETCC node into a SELECT_CC, LLVM uses
an incorrect type for determining BooleanContent of the result. This
patch fixes the issue.

Fixes PR36079.

[efriedma]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

3769 ↗

(On Diff #134188)

Shouldn't this be TLI.getBooleanContents(VT)? None of the operands of a SETCC are boolean.

[miyuki]

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

3769 ↗

(On Diff #134188)

As far as I understand, the parameter of getBooleanContents is the type of values being compared rather than the result type. At least, other uses of getBooleanContents in https://reviews.llvm.org/rL212697 suggest so.

[efriedma]

LGTM

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

3769 ↗

(On Diff #134188)

Oh, you're right; wasn't really thinking about it.

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

if (Legalized) {
       Results.push_back(Tmp1);
       break;
     }
 
     // Otherwise, SETCC for the given comparison type must be completely
     // illegal; expand it into a SELECT_CC.
     EVT VT = Node->getValueType(0);
     int TrueValue;
-    switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
+    switch (TLI.getBooleanContents(Tmp1.getValueType())) {
     case TargetLowering::ZeroOrOneBooleanContent:
     case TargetLowering::UndefinedBooleanContent:
       TrueValue = 1;
       break;
     case TargetLowering::ZeroOrNegativeOneBooleanContent:
       TrueValue = -1;
       break;
     }

llvm/trunk/test/CodeGen/ARM/2018-02-13-PR36079.ll

+; RUN: llc -mtriple=arm-eabi -mattr=+neon < %s -o - | FileCheck %s
+
+@c = global [4 x i32] [i32 3, i32 3, i32 3, i32 3], align 4
+@d = common global i32 0, align 4
+
+define void @foo() local_unnamed_addr nounwind norecurse {
+entry:
+  %0 = load <4 x i32>, <4 x i32>* bitcast ([4 x i32]* @c to <4 x i32>*), align 4
+  %1 = and <4 x i32> %0,
+           <i32 1,
+            i32 zext (i1 icmp ne (i32* getelementptr inbounds ([4 x i32], [4 x i32]* @c, i32 0, i32 1), i32* @d) to i32),
+            i32 zext (i1 icmp ne (i32* getelementptr inbounds ([4 x i32], [4 x i32]* @c, i32 0, i32 2), i32* @d) to i32),
+            i32 zext (i1 icmp ne (i32* getelementptr inbounds ([4 x i32], [4 x i32]* @c, i32 0, i32 3), i32* @d) to i32)>
+  store <4 x i32> %1, <4 x i32>* bitcast ([4 x i32]* @c to <4 x i32>*), align 4
+  ret void
+; CHECK-NOT: mvnne
+; CHECK: movne r{{[0-9]+}}, #1
+; CHECK-NOT: mvnne
+; CHECK: movne r{{[0-9]+}}, #1
+; CHECK-NOT: mvnne
+; CHECK: movne r{{[0-9]+}}, #1
+; CHECK-NOT: mvnne
+}