Index: lib/Target/AMDGPU/SIInstrInfo.cpp =================================================================== --- lib/Target/AMDGPU/SIInstrInfo.cpp +++ lib/Target/AMDGPU/SIInstrInfo.cpp @@ -2514,6 +2514,10 @@ return MI.readsRegister(AMDGPU::EXEC, &RI); } + // Make a conservative assumption about the callee. + if (MI.isCall()) + return true; + // Be conservative with any unhandled generic opcodes. if (!isTargetSpecificOpcode(MI.getOpcode())) return true; Index: test/CodeGen/AMDGPU/collapse-endcf2.mir =================================================================== --- /dev/null +++ test/CodeGen/AMDGPU/collapse-endcf2.mir @@ -0,0 +1,121 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass=si-optimize-exec-masking-pre-ra %s -o - | FileCheck -check-prefix=GCN %s + +--- | + define amdgpu_kernel void @call_no_explicit_exec_dependency () { + unreachable + } + + declare void @func() + +... + +# Call should be assumed to read exec +--- +name: call_no_explicit_exec_dependency +tracksRegLiveness: true +liveins: + - { reg: '$vgpr0', virtual-reg: '%0' } + - { reg: '$sgpr0_sgpr1', virtual-reg: '%1' } +machineFunctionInfo: + isEntryFunction: true +body: | + ; GCN-LABEL: name: call_no_explicit_exec_dependency + ; GCN: bb.0: + ; GCN: successors: %bb.1(0x40000000), %bb.4(0x40000000) + ; GCN: liveins: $vgpr0, $sgpr0_sgpr1 + ; GCN: [[COPY:%[0-9]+]]:sgpr_64 = COPY $sgpr0_sgpr1 + ; GCN: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0 + ; GCN: [[V_CMP_LT_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_LT_U32_e64 1, [[COPY1]], implicit $exec + ; GCN: [[COPY2:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec + ; GCN: [[S_AND_B64_:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY2]], [[V_CMP_LT_U32_e64_]], implicit-def dead $scc + ; GCN: $exec = S_MOV_B64_term [[S_AND_B64_]] + ; GCN: SI_MASK_BRANCH %bb.4, implicit $exec + ; GCN: S_BRANCH %bb.1 + ; GCN: bb.1: + ; GCN: successors: %bb.2(0x40000000), %bb.3(0x40000000) + ; GCN: undef %5.sub0_sub1:sgpr_128 = S_LOAD_DWORDX2_IMM [[COPY]], 9, 0, 0 :: (dereferenceable invariant load 8, align 4, addrspace 4) + ; GCN: undef %6.sub0:vreg_64 = V_LSHLREV_B32_e32 2, [[COPY1]], implicit $exec + ; GCN: %6.sub1:vreg_64 = V_MOV_B32_e32 0, implicit $exec + ; GCN: [[COPY3:%[0-9]+]]:vgpr_32 = COPY %5.sub1 + ; GCN: undef %8.sub0:vreg_64, %9:sreg_64_xexec = V_ADD_I32_e64 %5.sub0, %6.sub0, 0, implicit $exec + ; GCN: %8.sub1:vreg_64, dead %10:sreg_64_xexec = V_ADDC_U32_e64 0, [[COPY3]], %9, 0, implicit $exec + ; GCN: %5.sub3:sgpr_128 = S_MOV_B32 61440 + ; GCN: %5.sub2:sgpr_128 = S_MOV_B32 0 + ; GCN: BUFFER_STORE_DWORD_ADDR64 %6.sub1, %6, %5, 0, 0, 0, 0, 0, 0, implicit $exec :: (store 4, addrspace 1) + ; GCN: [[V_CMP_NE_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_NE_U32_e64 2, [[COPY1]], implicit $exec + ; GCN: [[COPY4:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec + ; GCN: [[S_AND_B64_1:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY4]], [[V_CMP_NE_U32_e64_]], implicit-def dead $scc + ; GCN: $exec = S_MOV_B64_term [[S_AND_B64_1]] + ; GCN: SI_MASK_BRANCH %bb.3, implicit $exec + ; GCN: S_BRANCH %bb.2 + ; GCN: bb.2: + ; GCN: successors: %bb.3(0x80000000) + ; GCN: %5.sub0:sgpr_128 = COPY %5.sub2 + ; GCN: %5.sub1:sgpr_128 = COPY %5.sub2 + ; GCN: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1, implicit $exec + ; GCN: BUFFER_STORE_DWORD_ADDR64 [[V_MOV_B32_e32_]], %8, %5, 0, 4, 0, 0, 0, 0, implicit $exec :: (store 4, addrspace 1) + ; GCN: bb.3: + ; GCN: successors: %bb.4(0x80000000) + ; GCN: $exec = S_OR_B64 $exec, [[COPY4]], implicit-def $scc + ; GCN: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF + ; GCN: dead %16:sreg_64 = SI_CALL [[DEF]], @func, csr_amdgpu_highregs + ; GCN: bb.4: + ; GCN: $exec = S_OR_B64 $exec, [[COPY2]], implicit-def $scc + ; GCN: [[V_MOV_B32_e32_1:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 3, implicit $exec + ; GCN: [[V_MOV_B32_e32_2:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec + ; GCN: $m0 = S_MOV_B32 -1 + ; GCN: DS_WRITE_B32 [[V_MOV_B32_e32_2]], [[V_MOV_B32_e32_1]], 0, 0, implicit $m0, implicit $exec :: (store 4, addrspace 3) + ; GCN: S_ENDPGM 0 + bb.0: + successors: %bb.1, %bb.4 + liveins: $vgpr0, $sgpr0_sgpr1 + + %1:sgpr_64 = COPY $sgpr0_sgpr1 + %0:vgpr_32 = COPY $vgpr0 + %2:sreg_64 = V_CMP_LT_U32_e64 1, %0, implicit $exec + %3:sreg_64 = COPY $exec, implicit-def $exec + %4:sreg_64 = S_AND_B64 %3, %2, implicit-def dead $scc + $exec = S_MOV_B64_term %4 + SI_MASK_BRANCH %bb.4, implicit $exec + S_BRANCH %bb.1 + + bb.1: + successors: %bb.2, %bb.3 + + undef %5.sub0_sub1:sgpr_128 = S_LOAD_DWORDX2_IMM %1, 9, 0, 0 :: (dereferenceable invariant load 8, align 4, addrspace 4) + undef %6.sub0:vreg_64 = V_LSHLREV_B32_e32 2, %0, implicit $exec + %6.sub1:vreg_64 = V_MOV_B32_e32 0, implicit $exec + %7:vgpr_32 = COPY %5.sub1 + undef %8.sub0:vreg_64, %9:sreg_64_xexec = V_ADD_I32_e64 %5.sub0, %6.sub0, 0, implicit $exec + %8.sub1:vreg_64, dead %10:sreg_64_xexec = V_ADDC_U32_e64 0, %7, %9, 0, implicit $exec + %5.sub3:sgpr_128 = S_MOV_B32 61440 + %5.sub2:sgpr_128 = S_MOV_B32 0 + BUFFER_STORE_DWORD_ADDR64 %6.sub1, %6, %5, 0, 0, 0, 0, 0, 0, implicit $exec :: (store 4, addrspace 1) + %11:sreg_64 = V_CMP_NE_U32_e64 2, %0, implicit $exec + %12:sreg_64 = COPY $exec, implicit-def $exec + %13:sreg_64 = S_AND_B64 %12, %11, implicit-def dead $scc + $exec = S_MOV_B64_term %13 + SI_MASK_BRANCH %bb.3, implicit $exec + S_BRANCH %bb.2 + + bb.2: + %5.sub0:sgpr_128 = COPY %5.sub2 + %5.sub1:sgpr_128 = COPY %5.sub2 + %14:vgpr_32 = V_MOV_B32_e32 1, implicit $exec + BUFFER_STORE_DWORD_ADDR64 %14, %8, %5, 0, 4, 0, 0, 0, 0, implicit $exec :: (store 4, addrspace 1) + + bb.3: + $exec = S_OR_B64 $exec, %12, implicit-def $scc + %20:sreg_64 = IMPLICIT_DEF + %21:sreg_64 = SI_CALL %20, @func, csr_amdgpu_highregs + + bb.4: + $exec = S_OR_B64 $exec, %3, implicit-def $scc + %17:vgpr_32 = V_MOV_B32_e32 3, implicit $exec + %18:vgpr_32 = V_MOV_B32_e32 0, implicit $exec + $m0 = S_MOV_B32 -1 + DS_WRITE_B32 %18, %17, 0, 0, implicit $m0, implicit $exec :: (store 4, addrspace 3) + S_ENDPGM 0 + +...