llvm/test/CodeGen/AMDGPU/sched-dynamic-vgpr-blocks.mir

# RUN: llc -mtriple=amdgpu12.00 -run-pass=machine-scheduler -amdgpu-stress-vgpr=18 -o - %s | FileCheck %s
# RUN: %if asserts %{ llc -mtriple=amdgpu12.00 -run-pass=machine-scheduler -amdgpu-stress-vgpr=18 -debug-only=machine-scheduler -filetype=null %s 2>&1 | FileCheck %s --check-prefix=DEBUG %}

# In dynamic VGPR mode we don't want to waste VGPR blocks, so we revert any
# schedule that increases the number of blocks we need to allocate. The
# scheduler will try to hoist the loads above the dependent adds, which will
# increase the pressure in the region from 5 to 17 VGPRs, i.e. from one block of
# 16 to 2. This should trigger the code that performs the revert based on the
# number of VGPR blocks.
# The -amdgpu-stress-vgpr flag is only there to keep the test small (the
# scheduler bails out early for regions that stay well below the VGPR limit).

--- |
  define amdgpu_cs void @dvgpr_blocks() #0 { ret void }
  attributes #0 = { "amdgpu-dynamic-vgpr-block-size"="16" }
...
---
# DEBUG: Region register pressure: VGPRs: 5
# DEBUG: Pressure after scheduling: VGPRs: 17
# DEBUG-NEXT: Region: 0
# DEBUG-NEXT: Occupancy before scheduling: 16, after 16.
# DEBUG-NEXT: Reverting scheduling for region 0

# CHECK-LABEL: name: dvgpr_blocks
# CHECK:      %acc0:vgpr_32 = V_MOV_B32_e32
# CHECK-NEXT: %l0:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 0,
# CHECK-NEXT: %acc1:vgpr_32 = V_ADD_U32_e32 %acc0, %l0,
# CHECK-NEXT: %l1:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 4,
# CHECK-NEXT: %acc2:vgpr_32 = V_ADD_U32_e32 %acc1, %l1,
# CHECK-NEXT: %l2:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 8,
# CHECK-NEXT: %acc3:vgpr_32 = V_ADD_U32_e32 %acc2, %l2,
name: dvgpr_blocks
tracksRegLiveness: true
machineFunctionInfo:
  dynamicVGPRBlockSize: 16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1
    %addr:vreg_64 = COPY $vgpr0_vgpr1
    %acc0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
    %l0:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 0, 0, implicit $exec
    %acc1:vgpr_32 = V_ADD_U32_e32 %acc0, %l0, implicit $exec
    %l1:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 4, 0, implicit $exec
    %acc2:vgpr_32 = V_ADD_U32_e32 %acc1, %l1, implicit $exec
    %l2:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 8, 0, implicit $exec
    %acc3:vgpr_32 = V_ADD_U32_e32 %acc2, %l2, implicit $exec
    %l3:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 12, 0, implicit $exec
    %acc4:vgpr_32 = V_ADD_U32_e32 %acc3, %l3, implicit $exec
    %l4:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 16, 0, implicit $exec
    %acc5:vgpr_32 = V_ADD_U32_e32 %acc4, %l4, implicit $exec
    %l5:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 20, 0, implicit $exec
    %acc6:vgpr_32 = V_ADD_U32_e32 %acc5, %l5, implicit $exec
    %l6:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 24, 0, implicit $exec
    %acc7:vgpr_32 = V_ADD_U32_e32 %acc6, %l6, implicit $exec
    %l7:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 28, 0, implicit $exec
    %acc8:vgpr_32 = V_ADD_U32_e32 %acc7, %l7, implicit $exec
    %l8:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 32, 0, implicit $exec
    %acc9:vgpr_32 = V_ADD_U32_e32 %acc8, %l8, implicit $exec
    %l9:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 36, 0, implicit $exec
    %acc10:vgpr_32 = V_ADD_U32_e32 %acc9, %l9, implicit $exec
    %l10:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 40, 0, implicit $exec
    %acc11:vgpr_32 = V_ADD_U32_e32 %acc10, %l10, implicit $exec
    %l11:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 44, 0, implicit $exec
    %acc12:vgpr_32 = V_ADD_U32_e32 %acc11, %l11, implicit $exec
    %l12:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 48, 0, implicit $exec
    %acc13:vgpr_32 = V_ADD_U32_e32 %acc12, %l12, implicit $exec
    %l13:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 52, 0, implicit $exec
    %acc14:vgpr_32 = V_ADD_U32_e32 %acc13, %l13, implicit $exec
    %l14:vgpr_32 = GLOBAL_LOAD_DWORD %addr, 56, 0, implicit $exec
    %acc15:vgpr_32 = V_ADD_U32_e32 %acc14, %l14, implicit $exec
    S_ENDPGM 0, implicit %acc15
...