Index: lib/Target/AMDGPU/SIInstrInfo.cpp =================================================================== --- lib/Target/AMDGPU/SIInstrInfo.cpp +++ lib/Target/AMDGPU/SIInstrInfo.cpp @@ -1634,6 +1634,29 @@ return AMDGPU::NoRegister; } +static bool shouldReadExec(const MachineInstr &MI) { + if (SIInstrInfo::isVALU(MI)) { + switch (MI.getOpcode()) { + case AMDGPU::V_READLANE_B32: + case AMDGPU::V_READLANE_B32_si: + case AMDGPU::V_READLANE_B32_vi: + case AMDGPU::V_WRITELANE_B32: + case AMDGPU::V_WRITELANE_B32_si: + case AMDGPU::V_WRITELANE_B32_vi: + return false; + } + + return true; + } + + if (SIInstrInfo::isGenericOpcode(MI.getOpcode()) || + SIInstrInfo::isSALU(MI) || + SIInstrInfo::isSMRD(MI)) + return false; + + return true; +} + bool SIInstrInfo::verifyInstruction(const MachineInstr *MI, StringRef &ErrInfo) const { uint16_t Opcode = MI->getOpcode(); @@ -1755,7 +1778,7 @@ // Make sure we aren't losing exec uses in the td files. This mostly requires // being careful when using let Uses to try to add other use registers. - if (!isGenericOpcode(Opcode) && !isSALU(Opcode) && !isSMRD(Opcode)) { + if (shouldReadExec(*MI)) { if (!MI->hasRegisterImplicitUseOperand(AMDGPU::EXEC)) { ErrInfo = "VALU instruction does not implicitly read exec mask"; return false; Index: lib/Target/AMDGPU/SIInstructions.td =================================================================== --- lib/Target/AMDGPU/SIInstructions.td +++ lib/Target/AMDGPU/SIInstructions.td @@ -1582,7 +1582,8 @@ } // End isCommutable = 1 -let isConvergent = 1 in { +// These are special and do not read the exec mask. +let isConvergent = 1, Uses = [] in { defm V_READLANE_B32 : VOP2SI_3VI_m < vop3 <0x001, 0x289>,