# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs -run-pass=si-fix-sgpr-copies -o - %s | FileCheck %s

# Regression for the gfx90a waterfall lane reader: when an inreg call
# argument's source virtual register lives in an AV_* class (which
# allows AGPR), emitLoadScalarOpsFromVGPRLoop must legalize the
# ScalarOp into a class compatible with V_READFIRSTLANE_B32 by
# inserting a COPY to the common subclass (e.g. VGPR_32). Without the
# COPY, the loop emits `V_READFIRSTLANE_B32 %vreg:av_32`, and
# `-verify-machineinstrs` fails with:
#   *** Bad machine code: Illegal virtual register for instruction ***
#   Expected a VGPR_32 register, but got a AV_32 register

# AV_32 ScalarOp: a COPY to VGPR_32 must be inserted in the loop header
# before V_READFIRSTLANE_B32 reads the lane, so the readfirstlane
# source is a pure VGPR class.

---
name:            waterfall_si_call_inreg_av32
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
  hasCalls: true
machineFunctionInfo:
  scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
  stackPtrOffsetReg: '$sgpr32'
  occupancy: 8
body:             |
  ; CHECK-LABEL: name: waterfall_si_call_inreg_av32
  ; CHECK: bb.0:
  ; CHECK:   [[SRC:%[0-9]+]]:av_32 = COPY $vgpr0
  ; CHECK:   [[DEST:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr5
  ; CHECK: bb.1:
  ; CHECK:   [[VCOPY:%[0-9]+]]:vgpr_32 = COPY [[SRC]]
  ; CHECK:   [[LANE:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[VCOPY]], implicit $exec
  ; CHECK:   {{.*}}:sreg_64_xexec = V_CMP_EQ_U32_e64 [[LANE]], [[VCOPY]], implicit $exec
  ; CHECK:   {{.*}}:sreg_64_xexec = S_AND_SAVEEXEC_B64
  ; CHECK: bb.2:
  ; CHECK:   ADJCALLSTACKUP
  ; CHECK:   $sgpr16 = COPY [[LANE]]
  ; CHECK:   SI_CALL_ISEL [[DEST]], 0, csr_amdgpu_gfx90ainsts, implicit killed $sgpr16
  ; CHECK:   ADJCALLSTACKDOWN
  ; CHECK:   SI_WATERFALL_LOOP %bb.1
  bb.0:
    liveins: $vgpr0, $sgpr4_sgpr5

    %0:av_32 = COPY $vgpr0
    %1:sreg_64_xexec = COPY $sgpr4_sgpr5

    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    $sgpr16 = COPY %0
    SI_CALL_ISEL killed %1, 0, csr_amdgpu_gfx90ainsts, implicit $sgpr16
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    S_ENDPGM 0
...

# Baseline: a VGPR_32 ScalarOp is already a legal V_READFIRSTLANE_B32
# source, so no extra COPY should appear in the loop header and
# V_READFIRSTLANE_B32 should read the original VGPR directly.

---
name:            waterfall_si_call_inreg_vgpr32
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
  hasCalls: true
machineFunctionInfo:
  scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
  stackPtrOffsetReg: '$sgpr32'
  occupancy: 8
body:             |
  ; CHECK-LABEL: name: waterfall_si_call_inreg_vgpr32
  ; CHECK: bb.0:
  ; CHECK:   [[SRC:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  ; CHECK:   [[DEST:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr5
  ; CHECK: bb.1:
  ; CHECK-NOT: vgpr_32 = COPY [[SRC]]
  ; CHECK:   [[LANE:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[SRC]], implicit $exec
  ; CHECK:   {{.*}}:sreg_64_xexec = V_CMP_EQ_U32_e64 [[LANE]], [[SRC]], implicit $exec
  bb.0:
    liveins: $vgpr0, $sgpr4_sgpr5

    %0:vgpr_32 = COPY $vgpr0
    %1:sreg_64_xexec = COPY $sgpr4_sgpr5

    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    $sgpr16 = COPY %0
    SI_CALL_ISEL killed %1, 0, csr_amdgpu_gfx90ainsts, implicit $sgpr16
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    S_ENDPGM 0
...