SelectionDAG: accommodate atomic floating stores

[t.p.northover]

Summary

We were applying a pointer truncation to floating types after r359676, which crashed LLVM. That is Not A Good Thing(TM).

[sanjoy]

LGTM

Thanks for the quick fix!

[t.p.northover]

Thanks. Committed as r360421.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
   Flags |= TLI.getMMOFlags(I);
 
   MachineFunction &MF = DAG.getMachineFunction();
   MachineMemOperand *MMO =
     MF.getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()), Flags,
                             MemVT.getStoreSize(), I.getAlignment(), AAMDNodes(),
                             nullptr, SSID, Ordering);
 
-  SDValue Val = DAG.getPtrExtOrTrunc(getValue(I.getValueOperand()), dl, MemVT);
+  SDValue Val = getValue(I.getValueOperand());
+  if (Val.getValueType() != MemVT)
+    Val = DAG.getPtrExtOrTrunc(Val, dl, MemVT);
+
   SDValue OutChain = DAG.getAtomic(ISD::ATOMIC_STORE, dl, MemVT, InChain,
                                    getValue(I.getPointerOperand()), Val, MMO);
 
 
   DAG.setRoot(OutChain);
 }
 
 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC

llvm/test/CodeGen/NVPTX/load-store.ll

define void @volatile(i8* %a, i16* %b, i32* %c, i64* %d) local_unnamed_addr {
   %d.add = add i64 %d.load, 1
   ; CHECK: st.volatile.u64 [%rd{{[0-9]+}}], %rd{{[0-9]+}}
   store volatile i64 %d.add, i64* %d
 
   ret void
 }
 
 ; CHECK-LABEL: monotonic
-define void @monotonic(i8* %a, i16* %b, i32* %c, i64* %d) local_unnamed_addr {
+define void @monotonic(i8* %a, i16* %b, i32* %c, i64* %d, float* %e) local_unnamed_addr {
   ; CHECK: ld.volatile.u8 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
   %a.load = load atomic i8, i8* %a monotonic, align 1
   %a.add = add i8 %a.load, 1
   ; CHECK: st.volatile.u8 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
   store atomic i8 %a.add, i8* %a monotonic, align 1
 
   ; CHECK: ld.volatile.u16 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
   %b.load = load atomic i16, i16* %b monotonic, align 2
   %b.add = add i16 %b.load, 1
   ; CHECK: st.volatile.u16 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
   store atomic i16 %b.add, i16* %b monotonic, align 2
 
   ; CHECK: ld.volatile.u32 %r{{[0-9]+}}, [%rd{{[0-9]+}}]
   %c.load = load atomic i32, i32* %c monotonic, align 4
   %c.add = add i32 %c.load, 1
   ; CHECK: st.volatile.u32 [%rd{{[0-9]+}}], %r{{[0-9]+}}
   store atomic i32 %c.add, i32* %c monotonic, align 4
 
   ; CHECK: ld.volatile.u64 %rd{{[0-9]+}}, [%rd{{[0-9]+}}]
   %d.load = load atomic i64, i64* %d monotonic, align 8
   %d.add = add i64 %d.load, 1
   ; CHECK: st.volatile.u64 [%rd{{[0-9]+}}], %rd{{[0-9]+}}
   store atomic i64 %d.add, i64* %d monotonic, align 8
 
+  ; CHECK: ld.volatile.f32 %f{{[0-9]+}}, [%rd{{[0-9]+}}]
+  %e.load = load atomic float, float* %e monotonic, align 4
+  %e.add = fadd float %e.load, 1.0
+  ; CHECK: st.volatile.f32 [%rd{{[0-9]+}}], %f{{[0-9]+}}
+  store atomic float %e.add, float* %e monotonic, align 4
+
   ret void
 }