[mips][msa] Prevent output operand from commuting for dpadd_[su].df instructions

[smaksimovic]

Summary

Implementation of TargetInstrInfo::findCommutedOpIndices for MIPS target, restricting commutativity to second and third operand only for dpaadd_[su].df instructions therein.

Prior to this change, there were cases where the vector that is to be added to the dot product of the other two could take a position other than the first one in the instruction, generating false output in the destination vector.

Such behavior has been noticed in the two functions generating v2i64 output values so far. Other ones may exhibit such behavior as well, just not for the vector operands which are present in the test at the moment.

Tests altered so that the function's first operand is a constant splat so that it can be loaded with a ldi instruction, since that is the case in which the erroneous instruction operand placement has occurred. We check that the register which is present in the ldi instruction is placed as the first operand in the corresponding dpadd instruction.

[sdardis]

Hi, this patch is technically correct but it misses the case that the second and third operands to dpadd_(s|u).[hwd] are commutable.

Unfortunately, the instruction definition has the typical tied format of (set MSA128DOpnd:$wd, (int_mips_dpadd_s_d MSA128DOpnd:$wd_in, MSA128WOpnd:$ws, MSA128WOpnd:$wt). The two address pass commutes the $wd_in and $ws.

Rather than removing the isCommutable class from those instructions, can you have a look at implementing TargetInstrInfo::commuteInstructionImpl for MIPS and handling these cases there?

Thanks,
Simon

[smaksimovic]

Hi Simon,

As per your suggestion, we've taken a look at TargetInstrInfo::commuteInstructionImpl and ended up reimplementing TargetInstrInfo::findCommutedOpIndices
for the MIPS target and handled our instructions there, keeping the isCommutable flag in the .td file.

The latter function is being called prior to the former, setting its parameters accordingly.

[sdardis]

LGTM with inline comment addressed. Don't forget to update the patch message before committing.

lib/Target/Mips/MipsInstrInfo.cpp

522–525 ↗

(On Diff #92146)

I don't think we need to assign to CommutableOpIdx here. For these 6 cases, we know that it is always operands 2 and 3 that can be commuted.

llvm/trunk/lib/Target/Mips/MipsInstrInfo.h

virtual void adjustStackPtr(unsigned SP, int64_t Amount,
                               MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator I) const = 0;
 
   /// Create an instruction which has the same operands and memory operands
   /// as MI but has a new opcode.
   MachineInstrBuilder genInstrWithNewOpc(unsigned NewOpc,
                                          MachineBasicBlock::iterator I) const;
 
+  bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
+                             unsigned &SrcOpIdx2) const override;
+
 protected:
   bool isZeroImm(const MachineOperand &op) const;
 
   MachineMemOperand *GetMemOperand(MachineBasicBlock &MBB, int FI,
                                    MachineMemOperand::Flags Flags) const;
 
 private:
   virtual unsigned getAnalyzableBrOpc(unsigned Opc) const = 0;

llvm/trunk/lib/Target/Mips/MipsInstrInfo.cpp

if (NewOpc == Mips::JIC || NewOpc == Mips::JIALC || NewOpc == Mips::JIC64 ||
     }
   }
 
   MIB.copyImplicitOps(*I);
 
   MIB.setMemRefs(I->memoperands_begin(), I->memoperands_end());
   return MIB;
 }
+
+bool MipsInstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
+                                          unsigned &SrcOpIdx2) const {
+  assert(!MI.isBundle() &&
+         "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
+
+  const MCInstrDesc &MCID = MI.getDesc();
+  if (!MCID.isCommutable())
+    return false;
+
+  switch (MI.getOpcode()) {
+  case Mips::DPADD_U_H:
+  case Mips::DPADD_U_W:
+  case Mips::DPADD_U_D:
+  case Mips::DPADD_S_H:
+  case Mips::DPADD_S_W:
+  case Mips::DPADD_S_D: {
+    // The first operand is both input and output, so it should not commute
+    if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2, 3))
+      return false;
+
+    if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg())
+      return false;
+    return true;
+  }
+  }
+  return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
+}

llvm/trunk/test/CodeGen/Mips/msa/3r_4r_widen.ll

 ; Test the MSA intrinsics that are encoded with the 3R instruction format and
 ; use the result as a third operand and results in wider elements than the
 ; operands had.
 
 ; RUN: llc -march=mips -mattr=+msa,+fp64 < %s | FileCheck %s
 ; RUN: llc -march=mipsel -mattr=+msa,+fp64 < %s | FileCheck %s
 
-@llvm_mips_dpadd_s_h_ARG1 = global <8 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7>, align 16
 @llvm_mips_dpadd_s_h_ARG2 = global <16 x i8> <i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>, align 16
 @llvm_mips_dpadd_s_h_ARG3 = global <16 x i8> <i8 24, i8 25, i8 26, i8 27, i8 28, i8 29, i8 30, i8 31, i8 32, i8 33, i8 34, i8 35, i8 36, i8 37, i8 38, i8 39>, align 16
 @llvm_mips_dpadd_s_h_RES  = global <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, align 16
 
 define void @llvm_mips_dpadd_s_h_test() nounwind {
 entry:
-  %0 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_s_h_ARG1
-  %1 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_s_h_ARG2
-  %2 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_s_h_ARG3
-  %3 = tail call <8 x i16> @llvm.mips.dpadd.s.h(<8 x i16> %0, <16 x i8> %1, <16 x i8> %2)
-  store <8 x i16> %3, <8 x i16>* @llvm_mips_dpadd_s_h_RES
+  %0 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_s_h_ARG2
+  %1 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_s_h_ARG3
+  %2 = tail call <8 x i16> @llvm.mips.dpadd.s.h(<8 x i16> <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>, <16 x i8> %0, <16 x i8> %1)
+  store <8 x i16> %2, <8 x i16>* @llvm_mips_dpadd_s_h_RES
   ret void
 }
 
 declare <8 x i16> @llvm.mips.dpadd.s.h(<8 x i16>, <16 x i8>, <16 x i8>) nounwind
 
 ; CHECK: llvm_mips_dpadd_s_h_test:
 ; CHECK: ld.b
 ; CHECK: ld.b
-; CHECK: ld.h
-; CHECK: dpadd_s.h
+; CHECK: ldi.h [[R1:\$w[0-9]+]],
+; CHECK: dpadd_s.h [[R1]],
 ; CHECK: st.h
 ; CHECK: .size llvm_mips_dpadd_s_h_test
 ;
-@llvm_mips_dpadd_s_w_ARG1 = global <4 x i32> <i32 0, i32 1, i32 2, i32 3>, align 16
 @llvm_mips_dpadd_s_w_ARG2 = global <8 x i16> <i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11>, align 16
 @llvm_mips_dpadd_s_w_ARG3 = global <8 x i16> <i16 12, i16 13, i16 14, i16 15, i16 16, i16 17, i16 18, i16 19>, align 16
 @llvm_mips_dpadd_s_w_RES  = global <4 x i32> <i32 0, i32 0, i32 0, i32 0>, align 16
 
 define void @llvm_mips_dpadd_s_w_test() nounwind {
 entry:
-  %0 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_s_w_ARG1
-  %1 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_s_w_ARG2
-  %2 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_s_w_ARG3
-  %3 = tail call <4 x i32> @llvm.mips.dpadd.s.w(<4 x i32> %0, <8 x i16> %1, <8 x i16> %2)
-  store <4 x i32> %3, <4 x i32>* @llvm_mips_dpadd_s_w_RES
+  %0 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_s_w_ARG2
+  %1 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_s_w_ARG3
+  %2 = tail call <4 x i32> @llvm.mips.dpadd.s.w(<4 x i32> <i32 4, i32 4, i32 4, i32 4>, <8 x i16> %0, <8 x i16> %1)
+  store <4 x i32> %2, <4 x i32>* @llvm_mips_dpadd_s_w_RES
   ret void
 }
 
 declare <4 x i32> @llvm.mips.dpadd.s.w(<4 x i32>, <8 x i16>, <8 x i16>) nounwind
 
 ; CHECK: llvm_mips_dpadd_s_w_test:
 ; CHECK: ld.h
 ; CHECK: ld.h
-; CHECK: ld.w
-; CHECK: dpadd_s.w
+; CHECK: ldi.w [[R1:\$w[0-9]+]],
+; CHECK: dpadd_s.w [[R1]],
 ; CHECK: st.w
 ; CHECK: .size llvm_mips_dpadd_s_w_test
 ;
-@llvm_mips_dpadd_s_d_ARG1 = global <2 x i64> <i64 0, i64 1>, align 16
 @llvm_mips_dpadd_s_d_ARG2 = global <4 x i32> <i32 2, i32 3, i32 4, i32 5>, align 16
 @llvm_mips_dpadd_s_d_ARG3 = global <4 x i32> <i32 6, i32 7, i32 8, i32 9>, align 16
 @llvm_mips_dpadd_s_d_RES  = global <2 x i64> <i64 0, i64 0>, align 16
 
 define void @llvm_mips_dpadd_s_d_test() nounwind {
 entry:
-  %0 = load <2 x i64>, <2 x i64>* @llvm_mips_dpadd_s_d_ARG1
-  %1 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_s_d_ARG2
-  %2 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_s_d_ARG3
-  %3 = tail call <2 x i64> @llvm.mips.dpadd.s.d(<2 x i64> %0, <4 x i32> %1, <4 x i32> %2)
-  store <2 x i64> %3, <2 x i64>* @llvm_mips_dpadd_s_d_RES
+  %0 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_s_d_ARG2
+  %1 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_s_d_ARG3
+  %2 = tail call <2 x i64> @llvm.mips.dpadd.s.d(<2 x i64> <i64 4, i64 4>, <4 x i32> %0, <4 x i32> %1)
+  store <2 x i64> %2, <2 x i64>* @llvm_mips_dpadd_s_d_RES
   ret void
 }
 
 declare <2 x i64> @llvm.mips.dpadd.s.d(<2 x i64>, <4 x i32>, <4 x i32>) nounwind
 
 ; CHECK: llvm_mips_dpadd_s_d_test:
+; CHECK: ldi.d [[R1:\$w[0-9]+]],
 ; CHECK: ld.w
 ; CHECK: ld.w
-; CHECK: ld.d
-; CHECK: dpadd_s.d
+; CHECK: dpadd_s.d [[R1]],
 ; CHECK: st.d
 ; CHECK: .size llvm_mips_dpadd_s_d_test
 ;
-@llvm_mips_dpadd_u_h_ARG1 = global <8 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7>, align 16
 @llvm_mips_dpadd_u_h_ARG2 = global <16 x i8> <i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>, align 16
 @llvm_mips_dpadd_u_h_ARG3 = global <16 x i8> <i8 24, i8 25, i8 26, i8 27, i8 28, i8 29, i8 30, i8 31, i8 32, i8 33, i8 34, i8 35, i8 36, i8 37, i8 38, i8 39>, align 16
 @llvm_mips_dpadd_u_h_RES  = global <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, align 16
 
 define void @llvm_mips_dpadd_u_h_test() nounwind {
 entry:
-  %0 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_u_h_ARG1
-  %1 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_u_h_ARG2
-  %2 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_u_h_ARG3
-  %3 = tail call <8 x i16> @llvm.mips.dpadd.u.h(<8 x i16> %0, <16 x i8> %1, <16 x i8> %2)
-  store <8 x i16> %3, <8 x i16>* @llvm_mips_dpadd_u_h_RES
+  %0 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_u_h_ARG2
+  %1 = load <16 x i8>, <16 x i8>* @llvm_mips_dpadd_u_h_ARG3
+  %2 = tail call <8 x i16> @llvm.mips.dpadd.u.h(<8 x i16> <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4>, <16 x i8> %0, <16 x i8> %1)
+  store <8 x i16> %2, <8 x i16>* @llvm_mips_dpadd_u_h_RES
   ret void
 }
 
 declare <8 x i16> @llvm.mips.dpadd.u.h(<8 x i16>, <16 x i8>, <16 x i8>) nounwind
 
 ; CHECK: llvm_mips_dpadd_u_h_test:
 ; CHECK: ld.b
 ; CHECK: ld.b
-; CHECK: ld.h
-; CHECK: dpadd_u.h
+; CHECK: ldi.h [[R1:\$w[0-9]+]],
+; CHECK: dpadd_u.h [[R1]],
 ; CHECK: st.h
 ; CHECK: .size llvm_mips_dpadd_u_h_test
 ;
-@llvm_mips_dpadd_u_w_ARG1 = global <4 x i32> <i32 0, i32 1, i32 2, i32 3>, align 16
 @llvm_mips_dpadd_u_w_ARG2 = global <8 x i16> <i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11>, align 16
 @llvm_mips_dpadd_u_w_ARG3 = global <8 x i16> <i16 12, i16 13, i16 14, i16 15, i16 16, i16 17, i16 18, i16 19>, align 16
 @llvm_mips_dpadd_u_w_RES  = global <4 x i32> <i32 0, i32 0, i32 0, i32 0>, align 16
 
 define void @llvm_mips_dpadd_u_w_test() nounwind {
 entry:
-  %0 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_u_w_ARG1
-  %1 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_u_w_ARG2
-  %2 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_u_w_ARG3
-  %3 = tail call <4 x i32> @llvm.mips.dpadd.u.w(<4 x i32> %0, <8 x i16> %1, <8 x i16> %2)
-  store <4 x i32> %3, <4 x i32>* @llvm_mips_dpadd_u_w_RES
+  %0 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_u_w_ARG2
+  %1 = load <8 x i16>, <8 x i16>* @llvm_mips_dpadd_u_w_ARG3
+  %2 = tail call <4 x i32> @llvm.mips.dpadd.u.w(<4 x i32> <i32 4, i32 4, i32 4, i32 4>, <8 x i16> %0, <8 x i16> %1)
+  store <4 x i32> %2, <4 x i32>* @llvm_mips_dpadd_u_w_RES
   ret void
 }
 
 declare <4 x i32> @llvm.mips.dpadd.u.w(<4 x i32>, <8 x i16>, <8 x i16>) nounwind
 
 ; CHECK: llvm_mips_dpadd_u_w_test:
 ; CHECK: ld.h
 ; CHECK: ld.h
-; CHECK: ld.w
-; CHECK: dpadd_u.w
+; CHECK: ldi.w [[R1:\$w[0-9]+]],
+; CHECK: dpadd_u.w [[R1]],
 ; CHECK: st.w
 ; CHECK: .size llvm_mips_dpadd_u_w_test
 ;
-@llvm_mips_dpadd_u_d_ARG1 = global <2 x i64> <i64 0, i64 1>, align 16
 @llvm_mips_dpadd_u_d_ARG2 = global <4 x i32> <i32 2, i32 3, i32 4, i32 5>, align 16
 @llvm_mips_dpadd_u_d_ARG3 = global <4 x i32> <i32 6, i32 7, i32 8, i32 9>, align 16
 @llvm_mips_dpadd_u_d_RES  = global <2 x i64> <i64 0, i64 0>, align 16
 
 define void @llvm_mips_dpadd_u_d_test() nounwind {
 entry:
-  %0 = load <2 x i64>, <2 x i64>* @llvm_mips_dpadd_u_d_ARG1
-  %1 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_u_d_ARG2
-  %2 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_u_d_ARG3
-  %3 = tail call <2 x i64> @llvm.mips.dpadd.u.d(<2 x i64> %0, <4 x i32> %1, <4 x i32> %2)
-  store <2 x i64> %3, <2 x i64>* @llvm_mips_dpadd_u_d_RES
+  %0 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_u_d_ARG2
+  %1 = load <4 x i32>, <4 x i32>* @llvm_mips_dpadd_u_d_ARG3
+  %2 = tail call <2 x i64> @llvm.mips.dpadd.u.d(<2 x i64> <i64 4, i64 4>, <4 x i32> %0, <4 x i32> %1)
+  store <2 x i64> %2, <2 x i64>* @llvm_mips_dpadd_u_d_RES
   ret void
 }
 
 declare <2 x i64> @llvm.mips.dpadd.u.d(<2 x i64>, <4 x i32>, <4 x i32>) nounwind
 
 ; CHECK: llvm_mips_dpadd_u_d_test:
+; CHECK: ldi.d [[R1:\$w[0-9]+]],
 ; CHECK: ld.w
 ; CHECK: ld.w
-; CHECK: ld.d
-; CHECK: dpadd_u.d
+; CHECK: dpadd_u.d [[R1]],
 ; CHECK: st.d
 ; CHECK: .size llvm_mips_dpadd_u_d_test
 ;
 @llvm_mips_dpsub_s_h_ARG1 = global <8 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7>, align 16
 @llvm_mips_dpsub_s_h_ARG2 = global <16 x i8> <i8 8, i8 9, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>, align 16
 @llvm_mips_dpsub_s_h_ARG3 = global <16 x i8> <i8 24, i8 25, i8 26, i8 27, i8 28, i8 29, i8 30, i8 31, i8 32, i8 33, i8 34, i8 35, i8 36, i8 37, i8 38, i8 39>, align 16
 @llvm_mips_dpsub_s_h_RES  = global <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, align 16