[X86] When handling addcarry intrinsic, create the flag result with the correct type...

Summary:
Previously we were creating the flag result with MVT::Other which is interpretted as a Chain node. If we used a memory form of the instruction we would end up with a copyToReg that consumed the chain result of the adcx instruction instead of the flag result.

Pretty sure we should be using MVT::i32 here, that's what we do other places we create these node types.

We should probably consider this for 5.0 as well.

Reviewers: RKSimon, zvi, spatel

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D36645

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310784 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5109b74..23f2700 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -20618,8 +20618,8 @@ static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget &Subtarget,
   }
   // ADC/ADCX/SBB
   case ADX: {
-    SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
-    SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
+    SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
+    SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::i32);
     SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
                                 DAG.getConstant(-1, dl, MVT::i8));
     SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),

diff --git a/test/CodeGen/X86/adx-intrinsics.ll b/test/CodeGen/X86/adx-intrinsics.ll
index 0498177..819a5df 100644 (file)
--- a/test/CodeGen/X86/adx-intrinsics.ll
+++ b/test/CodeGen/X86/adx-intrinsics.ll
@@ -75,3 +75,30 @@ define i8 @test_subborrow_u64(i8 %c, i64 %a, i64 %b, i8* %ptr) {
   ret i8 %ret;
 }
 
+; Try a version with loads. Previously we crashed on this.
+define i32 @load_crash(i64* nocapture readonly %a, i64* nocapture readonly %b, i64* %res)  {
+; CHECK-LABEL: load_crash
+; CHECK: addb
+; ADX: adcxq
+; CHECK: setb
+; CHECK: retq
+  %1 = load i64, i64* %a, align 8
+  %2 = load i64, i64* %b, align 8
+  %3 = bitcast i64* %res to i8*
+  %4 = tail call i8 @llvm.x86.addcarryx.u64(i8 0, i64 %1, i64 %2, i8* %3)
+  %conv = zext i8 %4 to i32
+  ret i32 %conv
+}
+
+; Try a really simple all zero input case, which also used to crash
+define void @allzeros() {
+; CHECK-LABEL: allzeros
+; CHECK: xorl
+; CHECK: addb
+; CHECK: sbbq
+; CHECK: andl
+; CHECK: retq
+entry:
+  %0 = tail call i8 @llvm.x86.addcarryx.u64(i8 0, i64 0, i64 0, i8* null)
+  ret void
+}