Open
Description
The following testcase fails verification:
# RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-coalescing -o - -run-pass=register-coalescer %s
---
name: test
tracksRegLiveness: true
liveins:
- { reg: '$vgpr0', virtual-reg: '%0' }
body: |
bb.0:
liveins: $vgpr0
%0:vgpr_32(s32) = COPY killed $vgpr0
%1:vgpr_32 = V_AND_B32_e32 1023, killed %0(s32), implicit $exec
undef %2.sub0:vreg_128_align2 = COPY %1
undef %3.sub1:vreg_256_align2 = COPY killed %1
%4:sgpr_32 = V_READFIRSTLANE_B32 killed %2.sub0, implicit $exec
%5:sgpr_32 = S_MOV_B32 0
undef %6.sub0:sgpr_128 = COPY killed %4
%6.sub1:sgpr_128 = COPY %5
%6.sub2:sgpr_128 = COPY %5
%6.sub3:sgpr_128 = COPY %5
%7:sgpr_32 = S_MOV_B32 65536
%8:sgpr_32 = V_READFIRSTLANE_B32 %3.sub1, implicit $exec
%9:sgpr_32 = V_READFIRSTLANE_B32 %3.sub3, implicit $exec
%10:sgpr_32 = V_READFIRSTLANE_B32 %3.sub5, implicit $exec
%11:sgpr_32 = V_READFIRSTLANE_B32 %3.sub6, implicit $exec
%12:sgpr_32 = V_READFIRSTLANE_B32 killed %3.sub7, implicit $exec
undef %13.sub0:sgpr_256 = COPY killed %7
%13.sub1:sgpr_256 = COPY killed %8
%13.sub2:sgpr_256 = COPY %5
%13.sub3:sgpr_256 = COPY killed %9
%13.sub4:sgpr_256 = COPY killed %5
%13.sub5:sgpr_256 = COPY killed %10
%13.sub6:sgpr_256 = COPY killed %11
%13.sub7:sgpr_256 = COPY killed %12
S_ENDPGM 0, implicit killed %6, implicit killed %13
...
It has multiple verification errors, essentially like this:
# Before register coalescing
...
*** Bad machine code: No live subrange at use ***
- function: test
- basic block: %bb.0 (0x57249c37bb48) [0B;416B)
- instruction: 208B %9:sgpr_32 = V_READFIRSTLANE_B32 %3.sub3:vreg_256_align2, implicit $exec
- operand 1: %3.sub3:vreg_256_align2
- interval: %3 [64r,256r:0) 0@64r L000000000000000C [64r,192r:0) 0@64r weight:0.000000e+00
- at: 208B
This is indeed true, only %3.sub1 is defined. These uses should be marked as undef, then no error occurs, but apparently there is nothing to set undef flag of these. There is detect-dead-lanes pass, but it only works with a small amount of opcodes. In similar cases register coalescer marks operands undef, but not here.
This is the same testcase, stopped before the detect-dead-lanes:
# RUN: llc -march=amdgcn -mcpu=gfx1100 -start-before=detect-dead-lanes -verify-coalescing -o - %s
---
name: test
tracksRegLiveness: true
liveins:
- { reg: '$vgpr0', virtual-reg: '%0' }
body: |
bb.0:
liveins: $vgpr0
%0:vgpr_32(s32) = COPY $vgpr0
%6:vgpr_32 = V_AND_B32_e32 1023, %0(s32), implicit $exec
%30:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
%29:vreg_128_align2 = REG_SEQUENCE %6, %subreg.sub0, %30, %subreg.sub1, %30, %subreg.sub2, %30, %subreg.sub3
%11:sreg_32 = IMPLICIT_DEF
%13:sreg_32 = IMPLICIT_DEF
%15:sreg_32 = IMPLICIT_DEF
%17:sreg_32 = IMPLICIT_DEF
%22:vgpr_32 = V_MOV_B32_e32 65536, implicit $exec
%24:vgpr_32 = COPY %11
%26:vgpr_32 = COPY %13
%27:vgpr_32 = COPY %15
%28:vgpr_32 = COPY %17
%21:vreg_256_align2 = REG_SEQUENCE killed %22, %subreg.sub0, %6, %subreg.sub1, %30, %subreg.sub2, killed %24, %subreg.sub3, %30, %subreg.sub4, killed %26, %subreg.sub5, killed %27, %subreg.sub6, killed %28, %subreg.sub7
%34:sgpr_32 = V_READFIRSTLANE_B32 %29.sub0, implicit $exec
%35:sgpr_32 = S_MOV_B32 0
%33:sgpr_128 = REG_SEQUENCE %34, %subreg.sub0, %35, %subreg.sub1, %35, %subreg.sub2, %35, %subreg.sub3
%39:sgpr_32 = S_MOV_B32 65536
%40:sgpr_32 = V_READFIRSTLANE_B32 %21.sub1, implicit $exec
%42:sgpr_32 = V_READFIRSTLANE_B32 %21.sub3, implicit $exec
%44:sgpr_32 = V_READFIRSTLANE_B32 %21.sub5, implicit $exec
%45:sgpr_32 = V_READFIRSTLANE_B32 %21.sub6, implicit $exec
%46:sgpr_32 = V_READFIRSTLANE_B32 %21.sub7, implicit $exec
%38:sgpr_256 = REG_SEQUENCE %39, %subreg.sub0, %40, %subreg.sub1, %35, %subreg.sub2, %42, %subreg.sub3, %35, %subreg.sub4, %44, %subreg.sub5, %45, %subreg.sub6, %46, %subreg.sub7
S_ENDPGM 0, implicit killed %33, implicit killed %38
...
Should it be marked undef with LIS creation itself?