[ARM] Fix arm_get_fpscr intrinsic selection in ISel

[rs]

Summary

Hans found a bug in my patch https://bugs.llvm.org/show_bug.cgi?id=32134 , here was the review https://reviews.llvm.org/D30542 . I didn't check to see if fpscr intrinsic was being selected in ISel.

I've fixed this by changing the node type in the instruction selection from INTRINSIC_WO_CHAIN (no side-effects) to INTRINSIC_W_CHAIN (has side-effects) which makes it semantically correct with the definition of the intrinsic in the tablegen code. Also added a test based on Hans reproducer to make sure instruction selection on int_arm_get_fpscr works.

[hans]

Sounds good to me. Someone more familiar with ARM should probably sign off on this though.

[rengolin]

Makes sense. LGTM Thanks!

include/llvm/IR/IntrinsicsARM.td

def int_arm_dsb : GCCBuiltin<"__builtin_arm_dsb">, MSBuiltin<"__dsb">,
                   Intrinsic<[], [llvm_i32_ty]>;
 def int_arm_isb : GCCBuiltin<"__builtin_arm_isb">, MSBuiltin<"__isb">,
                   Intrinsic<[], [llvm_i32_ty]>;
 
 //===----------------------------------------------------------------------===//
 // VFP
 
 def int_arm_get_fpscr : GCCBuiltin<"__builtin_arm_get_fpscr">,
-                       Intrinsic<[llvm_i32_ty], [], [IntrNoMem]>;
+                       Intrinsic<[llvm_i32_ty], [], []>;
 def int_arm_set_fpscr : GCCBuiltin<"__builtin_arm_set_fpscr">,
                        Intrinsic<[], [llvm_i32_ty], []>;
 def int_arm_vcvtr     : Intrinsic<[llvm_float_ty], [llvm_anyfloat_ty],
                                   [IntrNoMem]>;
 def int_arm_vcvtru    : Intrinsic<[llvm_float_ty], [llvm_anyfloat_ty],
                                   [IntrNoMem]>;
 
 //===----------------------------------------------------------------------===//

lib/Target/ARM/ARMISelLowering.cpp

 
 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
                                             SelectionDAG &DAG) const {
   // The rounding mode is in bits 23:22 of the FPSCR.
   // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
   // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
   // so that the shift + and get folded into a bitfield extract.
   SDLoc dl(Op);
-  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
-                              DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
-                                              MVT::i32));
+  SDValue Ops[] = { DAG.getEntryNode(),
+                    DAG.getConstant(Intrinsic::arm_get_fpscr, dl, MVT::i32) };
+
+  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_W_CHAIN, dl, MVT::i32, Ops);
   SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
                                   DAG.getConstant(1U << 22, dl, MVT::i32));
   SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
                               DAG.getConstant(22, dl, MVT::i32));
   return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
                      DAG.getConstant(3, dl, MVT::i32));
 }
 

test/CodeGen/ARM/fpscr-intrinsics.ll

+; RUN: llc < %s -O0 -mtriple=armv7-eabi -mcpu=cortex-a8 -mattr=+neon,+fp-armv8 | FileCheck %s
+; RUN: llc < %s -O3 -mtriple=armv7-eabi -mcpu=cortex-a8 -mattr=+neon,+fp-armv8 | FileCheck %s
+
+@a = common global double 0.000000e+00, align 8
+
+; Function Attrs: noinline nounwind uwtable
+define void @strtod() {
+entry:
+  ; CHECK: vmrs r{{[0-9]+}}, fpscr
+  %0 = call i32 @llvm.flt.rounds()
+  %tobool = icmp ne i32 %0, 0
+  br i1 %tobool, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  store double 5.000000e-01, double* @a, align 8
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret void
+}
+
+; Function Attrs: nounwind
+define void @fn1(i32* nocapture %p) local_unnamed_addr {
+entry:
+  ; CHECK: vmrs r{{[0-9]+}}, fpscr
+  %0 = tail call i32 @llvm.arm.get.fpscr()
+  store i32 %0, i32* %p, align 4
+  ; CHECK: vmsr fpscr, r{{[0-9]+}}
+  tail call void @llvm.arm.set.fpscr(i32 1)
+  ; CHECK: vmrs r{{[0-9]+}}, fpscr
+  %1 = tail call i32 @llvm.arm.get.fpscr()
+  %arrayidx1 = getelementptr inbounds i32, i32* %p, i32 1
+  store i32 %1, i32* %arrayidx1, align 4
+  ret void
+}
+
+; Function Attrs: nounwind readonly
+declare i32 @llvm.arm.get.fpscr()
+
+; Function Attrs: nounwind writeonly
+declare void @llvm.arm.set.fpscr(i32)
+
+; Function Attrs: nounwind
+declare i32 @llvm.flt.rounds()