search

llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
27.21k stars 11.14k forks source link

SI Whole Quad Mode causes Bad machine code: Live segment doesn't end at a valid instruction #54202

Closed jayfoad closed 10 months ago

jayfoad commented 2 years ago

With the attached test case I get:

$ ~/llvm-debug/bin/llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs reduced.txt

# After SI Whole Quad Mode
********** INTERVALS **********
EXEC_LO_HI16 [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
EXEC_HI_LO16 [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
EXEC_HI_HI16 [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
VGPR0_LO16 [0B,16r:0) 0@0B-phi
VGPR0_HI16 [0B,16r:0) 0@0B-phi
%0 [192r,272r:0)[544B,608B:0) 0@192r  weight:0.000000e+00
%2 [304r,432B:0)[608B,624r:0) 0@304r  weight:0.000000e+00
%4 [464r,544B:0)[672B,704r:0) 0@464r  weight:0.000000e+00
%5 [16r,112r:0) 0@16r  weight:0.000000e+00
%34 [112r,176r:0) 0@112r  weight:0.000000e+00
%44 [32r,96B:3)[128r,256B:0)[256B,432B:2)[432B,480r:4)[576r,608B:1)[608B,672B:2) 0@128r 1@576r 2@256B-phi 3@32r 4@432B-phi  weight:0.000000e+00
%46 [160r,192r:0) 0@160r  weight:0.000000e+00
%47 [176r,208r:0) 0@176r  weight:0.000000e+00
%48 [272r,304r:0) 0@272r  weight:0.000000e+00
%49 [480r,496r:0) 0@480r  weight:0.000000e+00
%50 [8r,376r:1)[376r,376d:0)[544B,608B:1) 0@376r 1@8r  weight:0.000000e+00
RegMasks:
********** MACHINEINSTRS **********
# Machine code for function _amdgpu_ps_main: NoPHIs, TracksLiveness, TiedOpsRewritten
Function Live Ins: $vgpr0 in %5

0B  bb.0..entry:
      successors: %bb.1(0x40000000), %bb.5(0x40000000); %bb.1(50.00%), %bb.5(50.00%)
      liveins: $vgpr0
8B    %50:sreg_64 = COPY $exec
16B   %5:vgpr_32 = COPY $vgpr0
32B   %44:sreg_64 = S_MOV_B64 -1
64B   S_CBRANCH_SCC1 %bb.5, implicit undef $scc
80B   S_BRANCH %bb.1

96B bb.1 (%ir-block.7):
    ; predecessors: %bb.0
      successors: %bb.6(0x40000000), %bb.2(0x40000000); %bb.6(50.00%), %bb.2(50.00%)

112B      %34:sreg_64 = nofpexcept V_CMP_NGT_F32_e64 0, 0, 0, %5:vgpr_32, 0, implicit $mode, implicit $exec
128B      %44:sreg_64 = S_MOV_B64 0
160B      %46:sreg_64 = COPY $exec, implicit-def $exec
176B      %47:sreg_64 = S_AND_B64 %46:sreg_64, %34:sreg_64, implicit-def dead $scc
192B      %0:sreg_64 = S_XOR_B64 %47:sreg_64, %46:sreg_64, implicit-def dead $scc
208B      $exec = S_MOV_B64_term %47:sreg_64
224B      S_CBRANCH_EXECZ %bb.2, implicit $exec
240B      S_BRANCH %bb.6

256B    bb.2.Flow1:
    ; predecessors: %bb.1, %bb.6
      successors: %bb.3(0x40000000), %bb.7(0x40000000); %bb.3(50.00%), %bb.7(50.00%)

272B      %48:sreg_64 = S_OR_SAVEEXEC_B64 %0:sreg_64, implicit-def $exec, implicit-def $scc, implicit $exec
304B      %2:sreg_64 = S_AND_B64 $exec, %48:sreg_64, implicit-def $scc
320B      $exec = S_XOR_B64_term $exec, %2:sreg_64, implicit-def $scc
336B      S_CBRANCH_EXECZ %bb.7, implicit $exec
352B      S_BRANCH %bb.3

368B    bb.3 (%ir-block.16):
    ; predecessors: %bb.2
      successors: %bb.4(0x80000000); %bb.4(100.00%)

376B      dead %50:sreg_64 = S_ANDN2_B64 %50:sreg_64, $exec, implicit-def $scc
380B      SI_EARLY_TERMINATE_SCC0 implicit $exec, implicit $scc
388B      $exec = S_MOV_B64_term 0

400B    bb.4 (%ir-block.16):
    ; predecessors: %bb.3
      successors: %bb.7(0x80000000); %bb.7(100.00%)

416B      S_BRANCH %bb.7

432B    bb.5.Flow:
    ; predecessors: %bb.0, %bb.7
      successors: %bb.8(0x40000000), %bb.9(0x40000000); %bb.8(50.00%), %bb.9(50.00%)

464B      %4:sreg_64 = COPY $exec, implicit-def $exec
480B      %49:sreg_64 = S_AND_B64 %4:sreg_64, %44:sreg_64, implicit-def dead $scc
496B      $exec = S_MOV_B64_term %49:sreg_64
512B      S_CBRANCH_EXECZ %bb.9, implicit $exec
528B      S_BRANCH %bb.8

544B    bb.6 (%ir-block.21):
    ; predecessors: %bb.1
      successors: %bb.2(0x80000000); %bb.2(100.00%)

576B      %44:sreg_64 = COPY $exec
592B      S_BRANCH %bb.2

608B    bb.7.Flow2:
    ; predecessors: %bb.2, %bb.4
      successors: %bb.5(0x80000000); %bb.5(100.00%)

624B      $exec = S_OR_B64 $exec, %2:sreg_64, implicit-def $scc
656B      S_BRANCH %bb.5

672B    bb.8 (%ir-block.22):
    ; predecessors: %bb.5
      successors: %bb.9(0x80000000); %bb.9(100.00%)

688B    bb.9.UnifiedReturnBlock:
    ; predecessors: %bb.5, %bb.8

704B      $exec = S_OR_B64 $exec, %4:sreg_64, implicit-def $scc
720B      S_ENDPGM 0

# End machine code for function _amdgpu_ps_main.

*** Bad machine code: No instruction at VNInfo def index ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_LO_HI16
- ValNo:       8 (def 396r)

*** Bad machine code: Live segment doesn't end at a valid instruction ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_LO_HI16
- segment:     [396r,396d:8)

*** Bad machine code: No instruction at VNInfo def index ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_HI_LO16
- ValNo:       8 (def 396r)

*** Bad machine code: Live segment doesn't end at a valid instruction ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_HI_LO16
- segment:     [396r,396d:8)

*** Bad machine code: No instruction at VNInfo def index ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_HI_HI16
- ValNo:       8 (def 396r)

*** Bad machine code: Live segment doesn't end at a valid instruction ***
- function:    _amdgpu_ps_main
- basic block: %bb.3  (0xb021a08) [368B;400B)
- liverange:   [160r,160d:7)[208r,208d:6)[272r,272d:5)[320r,320d:4)[396r,396d:8)[464r,464d:3)[496r,496d:2)[624r,624d:1)[704r,704d:0) 0@704r 1@624r 2@496r 3@464r 4@320r 5@272r 6@208r 7@160r 8@396r
- regunit:     EXEC_HI_HI16
- segment:     [396r,396d:8)
LLVM ERROR: Found 6 machine code errors.
PLEASE submit a bug report to https://bugs.llvm.org/ and include the crash backtrace.
llvmbot commented 2 years ago

@llvm/issue-subscribers-backend-amdgpu

jayfoad commented 10 months ago

I can't reproduce this.