Files
clang-p2996/llvm/test/CodeGen/AMDGPU/spill192.mir
Christudasan Devadasan 7a98f084c4 [AMDGPU][SILowerSGPRSpills] Spill SGPRs to virtual VGPRs
Currently, the custom SGPR spill lowering pass spills
SGPRs into physical VGPR lanes and the remaining VGPRs
are used by regalloc for vector regclass allocation.
This imposes many restrictions that we ended up with
unsuccessful SGPR spilling when there won't be enough
VGPRs and we are forced to spill the leftover into
memory during PEI. The custom spill handling during PEI
has many edge cases and often breaks the compiler time
to time.

This patch implements spilling SGPRs into virtual VGPR
lanes. Since we now split the register allocation for
SGPRs and VGPRs, the virtual registers introduced for
the spill lanes would get allocated automatically in
the subsequent regalloc invocation for VGPRs.

Spill to virtual registers will always be successful,
even in the high-pressure situations, and hence it avoids
most of the edge cases during PEI. We are now left with
only the custom SGPR spills during PEI for special registers
like the frame pointer which is an unproblematic case.

Differential Revision: https://reviews.llvm.org/D124196
2023-07-07 23:14:32 +05:30

119 lines
5.7 KiB
YAML

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -march=amdgcn -mcpu=tahiti -run-pass=regallocfast -o - %s | FileCheck -check-prefix=SPILLED %s
# RUN: llc -march=amdgcn -mcpu=tahiti -run-pass=regallocfast,si-lower-sgpr-spills -o - %s | FileCheck -check-prefix=EXPANDED %s
# Make sure spill/restore of 192 bit registers works. We have to
# settle for a MIR test for now since inlineasm fails without 192-bit
# MVT.
---
name: spill_restore_sgpr192
tracksRegLiveness: true
machineFunctionInfo:
scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
stackPtrOffsetReg: $sgpr32
body: |
; SPILLED-LABEL: name: spill_restore_sgpr192
; SPILLED: bb.0:
; SPILLED-NEXT: successors: %bb.1(0x80000000)
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; SPILLED-NEXT: SI_SPILL_S192_SAVE killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9, %stack.0, implicit $exec, implicit $sgpr32 :: (store (s192) into %stack.0, align 4, addrspace 5)
; SPILLED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: bb.1:
; SPILLED-NEXT: successors: %bb.2(0x80000000)
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: S_NOP 1
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: bb.2:
; SPILLED-NEXT: $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9 = SI_SPILL_S192_RESTORE %stack.0, implicit $exec, implicit $sgpr32 :: (load (s192) from %stack.0, align 4, addrspace 5)
; SPILLED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-LABEL: name: spill_restore_sgpr192
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr9, 5, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
bb.0:
S_NOP 0, implicit-def %0:sgpr_192
S_CBRANCH_SCC1 implicit undef $scc, %bb.1
bb.1:
S_NOP 1
bb.2:
S_NOP 0, implicit %0
...
---
name: spill_restore_vgpr192
tracksRegLiveness: true
machineFunctionInfo:
scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
stackPtrOffsetReg: $sgpr32
body: |
; SPILLED-LABEL: name: spill_restore_vgpr192
; SPILLED: bb.0:
; SPILLED-NEXT: successors: %bb.1(0x80000000)
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: S_NOP 0, implicit-def renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5
; SPILLED-NEXT: SI_SPILL_V192_SAVE killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, %stack.0, $sgpr32, 0, implicit $exec :: (store (s192) into %stack.0, align 4, addrspace 5)
; SPILLED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: bb.1:
; SPILLED-NEXT: successors: %bb.2(0x80000000)
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: S_NOP 1
; SPILLED-NEXT: {{ $}}
; SPILLED-NEXT: bb.2:
; SPILLED-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = SI_SPILL_V192_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s192) from %stack.0, align 4, addrspace 5)
; SPILLED-NEXT: S_NOP 0, implicit killed renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5
; EXPANDED-LABEL: name: spill_restore_vgpr192
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5
; EXPANDED-NEXT: SI_SPILL_V192_SAVE killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, %stack.0, $sgpr32, 0, implicit $exec :: (store (s192) into %stack.0, align 4, addrspace 5)
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
; EXPANDED-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = SI_SPILL_V192_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s192) from %stack.0, align 4, addrspace 5)
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5
bb.0:
S_NOP 0, implicit-def %0:vreg_192
S_CBRANCH_SCC1 implicit undef $scc, %bb.1
bb.1:
S_NOP 1
bb.2:
S_NOP 0, implicit %0
...