caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
72d4a41ba622bd3a653db2cd9bf314ad83caf032
clang-p2996/llvm/test/CodeGen/AMDGPU/unallocatable-bundle-regression.mir
Matt Arsenault 81b2c23b77 AMDGPU: Use kill instruction to hint soft clause live ranges 
Previously we would use a bundle to hint the register allocator to not
overwrite the pointers in a sequence of loads to avoid breaking soft
clauses. This bundling was based on a fuzzy register pressure
heuristic, so we could not guarantee using more registers than are
really available. This would result in register allocator failing on
unsatisfiable bundles. Use a kill to artificially extend the live
ranges, so we can always succeed at register allocation even if it
means extra spills in the worst case.

This seems to capture most of the benefit of the bundle while avoiding
most of the risk presented by the bundle. However the lit tests do
show a handful of regressions. In some cases with sequences of
volatile loads, unused load components end up getting reallocated to
the next load which forces a wait between. There are also a few small
scheduling regressions where a hazard used to be avoided, and one
spill torture test which for some reason nearly doubles the stack
usage. There is also a bit of noise from leftover kills (it may make
sense for post-RA pseudos to strip all of these out).
2021-02-26 18:26:40 -05:00

170 lines
14 KiB
YAML
Raw Blame History

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn-- -mcpu=gfx1010 -verify-machineinstrs -run-pass=si-form-memory-clauses,greedy,virtregrewriter -o - %s | FileCheck %s
# FIXME: -start-before=si-form-memory-clauses doesn't work correctly with -stop-after
# This used to produce a bundle that was impossible for the register
# allocator to handle.
---
name: unallocatable_clause_bundle
tracksRegLiveness: true
machineFunctionInfo:
isEntryFunction: true
scratchRSrcReg: '$sgpr100_sgpr101_sgpr102_sgpr103'
stackPtrOffsetReg: '$sgpr32'
argumentInfo:
privateSegmentBuffer: { reg: '$sgpr0_sgpr1_sgpr2_sgpr3' }
mode:
ieee: false
occupancy: 20
body: |
bb.0:
liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $vgpr0
; CHECK-LABEL: name: unallocatable_clause_bundle
; CHECK: liveins: $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3
; CHECK: renamable $sgpr4 = COPY $sgpr0
; CHECK: SI_SPILL_S128_SAVE $sgpr0_sgpr1_sgpr2_sgpr3, %stack.0, implicit $exec, implicit $sgpr32 :: (store 16 into %stack.0, align 4, addrspace 5)
; CHECK: renamable $sgpr5 = S_MOV_B32 0
; CHECK: renamable $sgpr76 = COPY renamable $sgpr5
; CHECK: renamable $sgpr77 = COPY renamable $sgpr5
; CHECK: renamable $sgpr78 = COPY renamable $sgpr5
; CHECK: renamable $sgpr0 = S_MOV_B32 1056964608
; CHECK: renamable $sgpr79 = COPY renamable $sgpr5
; CHECK: renamable $sgpr1 = COPY renamable $sgpr0
; CHECK: renamable $sgpr8 = COPY renamable $sgpr5
; CHECK: renamable $sgpr9 = COPY renamable $sgpr5
; CHECK: renamable $sgpr10 = COPY renamable $sgpr5
; CHECK: renamable $sgpr11 = COPY renamable $sgpr5
; CHECK: renamable $sgpr12 = COPY renamable $sgpr5
; CHECK: renamable $sgpr13 = COPY renamable $sgpr5
; CHECK: renamable $sgpr14 = COPY renamable $sgpr5
; CHECK: renamable $sgpr15 = COPY renamable $sgpr5
; CHECK: renamable $vgpr5_vgpr6 = COPY renamable $sgpr0_sgpr1
; CHECK: renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1088, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr80_sgpr81_sgpr82_sgpr83 = S_LOAD_DWORDX4_IMM renamable $sgpr4_sgpr5, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: renamable $sgpr0 = S_MOV_B32 1200
; CHECK: renamable $sgpr1 = COPY renamable $sgpr5
; CHECK: renamable $sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1152, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr84_sgpr85_sgpr86_sgpr87 = S_LOAD_DWORDX4_IMM renamable $sgpr0_sgpr1, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: KILL renamable $sgpr0, renamable $sgpr1
; CHECK: renamable $sgpr0 = S_MOV_B32 1264
; CHECK: renamable $sgpr1 = COPY renamable $sgpr5
; CHECK: renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1216, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr88_sgpr89_sgpr90_sgpr91 = S_LOAD_DWORDX4_IMM renamable $sgpr0_sgpr1, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: KILL renamable $sgpr0, renamable $sgpr1
; CHECK: renamable $sgpr0 = S_MOV_B32 1328
; CHECK: renamable $sgpr1 = COPY renamable $sgpr5
; CHECK: renamable $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1280, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr92_sgpr93_sgpr94_sgpr95 = S_LOAD_DWORDX4_IMM renamable $sgpr0_sgpr1, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: KILL renamable $sgpr0, renamable $sgpr1
; CHECK: renamable $sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1344, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr0 = S_MOV_B32 1392
; CHECK: renamable $sgpr1 = COPY renamable $sgpr5
; CHECK: renamable $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 0, 0, 0 :: (load 32, addrspace 6)
; CHECK: renamable $sgpr2 = S_MOV_B32 1456
; CHECK: renamable $sgpr3 = COPY renamable $sgpr5
; CHECK: renamable $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75 = S_LOAD_DWORDX8_IMM renamable $sgpr4_sgpr5, 1472, 0, 0 :: (dereferenceable load 32, addrspace 6)
; CHECK: renamable $sgpr4 = S_MOV_B32 1520
; CHECK: renamable $sgpr96_sgpr97_sgpr98_sgpr99 = S_LOAD_DWORDX4_IMM renamable $sgpr2_sgpr3, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: renamable $sgpr4_sgpr5_sgpr6_sgpr7 = S_LOAD_DWORDX4_IMM renamable $sgpr4_sgpr5, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: renamable $sgpr0_sgpr1_sgpr2_sgpr3 = S_LOAD_DWORDX4_IMM renamable $sgpr0_sgpr1, 0, 0, 0 :: (load 16, addrspace 6)
; CHECK: renamable $vgpr7 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, renamable $sgpr76_sgpr77_sgpr78_sgpr79, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr8 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, killed renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23, killed renamable $sgpr80_sgpr81_sgpr82_sgpr83, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr9 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, killed renamable $sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31, killed renamable $sgpr84_sgpr85_sgpr86_sgpr87, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr10 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43, renamable $sgpr88_sgpr89_sgpr90_sgpr91, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr11 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51, renamable $sgpr92_sgpr93_sgpr94_sgpr95, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr12 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67, renamable $sgpr96_sgpr97_sgpr98_sgpr99, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr13 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75, renamable $sgpr4_sgpr5_sgpr6_sgpr7, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $vgpr14 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 renamable $vgpr5_vgpr6, renamable $sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59, renamable $sgpr0_sgpr1_sgpr2_sgpr3, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
; CHECK: renamable $sgpr8_sgpr9_sgpr10_sgpr11 = SI_SPILL_S128_RESTORE %stack.0, implicit $exec, implicit $sgpr32 :: (load 16 from %stack.0, align 4, addrspace 5)
; CHECK: renamable $vgpr1_vgpr2_vgpr3_vgpr4 = BUFFER_LOAD_FORMAT_XYZW_IDXEN renamable $vgpr0, renamable $sgpr8_sgpr9_sgpr10_sgpr11, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "BufferResource", align 1, addrspace 4)
; CHECK: KILL killed renamable $sgpr4_sgpr5_sgpr6_sgpr7
; CHECK: KILL killed renamable $sgpr92_sgpr93_sgpr94_sgpr95
; CHECK: KILL killed renamable $sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75
; CHECK: KILL killed renamable $sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; CHECK: KILL killed renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43
; CHECK: KILL killed renamable $vgpr5_vgpr6
; CHECK: KILL killed renamable $sgpr0_sgpr1_sgpr2_sgpr3
; CHECK: KILL killed renamable $sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59
; CHECK: KILL killed renamable $sgpr96_sgpr97_sgpr98_sgpr99
; CHECK: KILL killed renamable $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51
; CHECK: KILL killed renamable $sgpr8_sgpr9_sgpr10_sgpr11
; CHECK: KILL killed renamable $sgpr88_sgpr89_sgpr90_sgpr91
; CHECK: KILL killed renamable $vgpr0
; CHECK: renamable $vgpr0 = nofpexcept V_MAX_F32_e32 killed $vgpr7, killed $vgpr8, implicit $mode, implicit $exec
; CHECK: renamable $vgpr0 = V_MAX3_F32_e64 0, killed $vgpr0, 0, killed $vgpr9, 0, killed $vgpr10, 0, 0, implicit $mode, implicit $exec
; CHECK: renamable $vgpr1 = nofpexcept V_ADD_F32_e32 -1083321614, killed $vgpr12, implicit $mode, implicit $exec
; CHECK: renamable $vgpr3 = nofpexcept V_ADD_F32_e32 -1090988802, killed $vgpr13, implicit $mode, implicit $exec
; CHECK: renamable $vgpr0 = V_MAX3_F32_e64 0, killed $vgpr0, 0, killed $vgpr11, 0, killed $vgpr14, 0, 0, implicit $mode, implicit $exec
; CHECK: renamable $vgpr0 = V_MAX3_F32_e64 0, killed $vgpr0, 0, killed $vgpr1, 0, killed $vgpr3, 0, 0, implicit $mode, implicit $exec
; CHECK: renamable $sgpr0 = nofpexcept V_CMP_GT_F32_e64 0, 1028443341, 0, killed $vgpr0, 0, implicit $mode, implicit $exec
; CHECK: renamable $vgpr0 = V_CNDMASK_B32_e64 0, 0, 0, 1065353216, killed $sgpr0, implicit $exec
; CHECK: EXP_DONE 12, killed renamable $vgpr0, renamable $vgpr2, undef renamable $vgpr0, undef renamable $vgpr0, -1, 0, 15, implicit $exec
; CHECK: S_ENDPGM 0
%0:vgpr_32 = COPY $vgpr0
undef %1.sub0:sgpr_64 = COPY $sgpr0
%2:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
%1.sub1:sgpr_64 = S_MOV_B32 0
undef %3.sub0:sgpr_128 = COPY %1.sub1
%3.sub1:sgpr_128 = COPY %1.sub1
%3.sub2:sgpr_128 = COPY %1.sub1
undef %4.sub0:sgpr_64 = S_MOV_B32 1056964608
%3.sub3:sgpr_128 = COPY %1.sub1
%4.sub1:sgpr_64 = COPY %4.sub0
undef %5.sub0:sgpr_256 = COPY %1.sub1
%5.sub1:sgpr_256 = COPY %1.sub1
%5.sub2:sgpr_256 = COPY %1.sub1
%5.sub3:sgpr_256 = COPY %1.sub1
%5.sub4:sgpr_256 = COPY %1.sub1
%5.sub5:sgpr_256 = COPY %1.sub1
%5.sub6:sgpr_256 = COPY %1.sub1
%5.sub7:sgpr_256 = COPY %1.sub1
%6:vreg_64 = COPY %4
%7:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1088, 0, 0 :: (dereferenceable load 32, addrspace 6)
%8:sgpr_128 = S_LOAD_DWORDX4_IMM %1, 0, 0, 0 :: (load 16, addrspace 6)
undef %9.sub0:sreg_64_xexec = S_MOV_B32 1200
%9.sub1:sreg_64_xexec = COPY %1.sub1
%10:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1152, 0, 0 :: (dereferenceable load 32, addrspace 6)
%11:sgpr_128 = S_LOAD_DWORDX4_IMM %9, 0, 0, 0 :: (load 16, addrspace 6)
undef %12.sub0:sreg_64_xexec = S_MOV_B32 1264
%12.sub1:sreg_64_xexec = COPY %1.sub1
%13:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1216, 0, 0 :: (dereferenceable load 32, addrspace 6)
%14:sgpr_128 = S_LOAD_DWORDX4_IMM %12, 0, 0, 0 :: (load 16, addrspace 6)
undef %15.sub0:sreg_64_xexec = S_MOV_B32 1328
%15.sub1:sreg_64_xexec = COPY %1.sub1
%16:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1280, 0, 0 :: (dereferenceable load 32, addrspace 6)
%17:sgpr_128 = S_LOAD_DWORDX4_IMM %15, 0, 0, 0 :: (load 16, addrspace 6)
%18:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1344, 0, 0 :: (dereferenceable load 32, addrspace 6)
undef %19.sub0:sreg_64_xexec = S_MOV_B32 1392
%19.sub1:sreg_64_xexec = COPY %1.sub1
%20:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 0, 0, 0 :: (load 32, addrspace 6)
undef %21.sub0:sreg_64_xexec = S_MOV_B32 1456
%21.sub1:sreg_64_xexec = COPY %1.sub1
%22:sgpr_256 = S_LOAD_DWORDX8_IMM %1, 1472, 0, 0 :: (dereferenceable load 32, addrspace 6)
%1.sub0:sgpr_64 = S_MOV_B32 1520
%23:sgpr_128 = S_LOAD_DWORDX4_IMM %21, 0, 0, 0 :: (load 16, addrspace 6)
%24:sgpr_128 = S_LOAD_DWORDX4_IMM %1, 0, 0, 0 :: (load 16, addrspace 6)
%25:sgpr_128 = S_LOAD_DWORDX4_IMM %19, 0, 0, 0 :: (load 16, addrspace 6)
%26:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %5, %3, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%27:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %7, %8, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%28:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %10, %11, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%29:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %13, %14, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%30:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %16, %17, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%31:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %20, %23, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%32:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %22, %24, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%33:vgpr_32 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 %6, %18, %25, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
%34:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN %0, %2, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "BufferResource", align 1, addrspace 4)
%35:vgpr_32 = nofpexcept V_MAX_F32_e32 %26, %27, implicit $mode, implicit $exec
%36:vgpr_32 = V_MAX3_F32_e64 0, %35, 0, %28, 0, %29, 0, 0, implicit $mode, implicit $exec
%37:vgpr_32 = nofpexcept V_ADD_F32_e32 -1083321614, %31, implicit $mode, implicit $exec
%38:vgpr_32 = nofpexcept V_ADD_F32_e32 -1090988802, %32, implicit $mode, implicit $exec
%39:vgpr_32 = V_MAX3_F32_e64 0, %36, 0, %30, 0, %33, 0, 0, implicit $mode, implicit $exec
%40:vgpr_32 = V_MAX3_F32_e64 0, %39, 0, %37, 0, %38, 0, 0, implicit $mode, implicit $exec
%41:sreg_32_xm0_xexec = nofpexcept V_CMP_GT_F32_e64 0, 1028443341, 0, %40, 0, implicit $mode, implicit $exec
%42:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1065353216, %41, implicit $exec
EXP_DONE 12, %42, %34.sub1, undef %43:vgpr_32, undef %44:vgpr_32, -1, 0, 15, implicit $exec
S_ENDPGM 0
...
Reference in New Issue View Git Blame Copy Permalink