[PATCH] D82916: LIS: fix handleMove to properly extend main range

Tue Jul 7 07:34:37 PDT 2020

rampitec added a comment.

In D82916#2135752 <https://reviews.llvm.org/D82916#2135752>, @vpykhtin wrote:

> Can you please add a LiveInterval dump (possibly truncated) for the test before and after each move to this review? It's a bit hard to follow what happens there.

Sure! This is the LIS at each step w/o this patch:

Before 1st move:

  %1 [16r,32B:0)[32B,80r:4)[80r,80d:3)[96r,112r:1)[112r,160B:2)  0 at 16r 1 at 96r 2 at 112r 3 at 80r 4 at 32B-phi
  L000000000000000C [16r,32B:0)[32B,64r:1)  0 at 16r 1 at 32B-phi
  L00000000000000C0 [16r,16d:0)  0 at 16r
  L0000000000000003 [16r,16d:1)[96r,96d:0)  0 at 96r 1 at 16r
  L0000000000000030 [16r,32B:2)[32B,48r:3)[80r,80d:1)[112r,160B:0)  0 at 112r 1 at 80r 2 at 16r 3 at 32B-phi
  weight:0.000000e+00
  %2 [48r,80r:0)  0 at 48r weight:0.000000e+00
  %3 [64r,64d:0)  0 at 64r weight:0.000000e+00
  RegMasks:
  ********** MACHINEINSTRS **********
  # Machine code for function func: NoPHIs

  0B      bb.0:
            successors: %bb.1(0x80000000); %bb.1(100.00%)

  16B       %1:sgpr_128 = IMPLICIT_DEF

  32B     bb.1:
          ; predecessors: %bb.0, %bb.1
            successors: %bb.1(0x40000000), %bb.2(0x40000000); %bb.1(50.00%), %bb.2(50.00%)

  48B       %2:sgpr_32 = COPY %1.sub2:sgpr_128
  64B       dead %3:sgpr_32 = COPY %1.sub1:sgpr_128
  80B       dead %1.sub2:sgpr_128 = COPY %2:sgpr_32
  96B       %1.sub0:sgpr_128 = IMPLICIT_DEF
  112B      %1.sub2:sgpr_128 = IMPLICIT_DEF
  128B      S_CBRANCH_SCC1 %bb.1, implicit undef $scc
  144B      S_BRANCH %bb.2

  160B    bb.2:
          ; predecessors: %bb.1

Before 2nd move (still correct):

  %1 [16r,32B:0)[32B,80r:4)[80r,80d:3)[88r,96r:1)[96r,160B:2)  0 at 16r 1 at 88r 2 at 96r 3 at 80r 4 at 32B-phi
  L000000000000000C [16r,32B:0)[32B,64r:1)  0 at 16r 1 at 32B-phi
  L00000000000000C0 [16r,16d:0)  0 at 16r
  L0000000000000003 [16r,16d:1)[96r,96d:0)  0 at 96r 1 at 16r
  L0000000000000030 [16r,32B:2)[32B,48r:3)[80r,80d:1)[88r,160B:0)  0 at 88r 1 at 80r 2 at 16r 3 at 32B-phi
  weight:0.000000e+00
  %2 [48r,80r:0)  0 at 48r weight:0.000000e+00
  %3 [64r,64d:0)  0 at 64r weight:0.000000e+00
  RegMasks:
  ********** MACHINEINSTRS **********
  # Machine code for function func: NoPHIs

  0B      bb.0:
            successors: %bb.1(0x80000000); %bb.1(100.00%)

  16B       %1:sgpr_128 = IMPLICIT_DEF

  32B     bb.1:
          ; predecessors: %bb.0, %bb.1
            successors: %bb.1(0x40000000), %bb.2(0x40000000); %bb.1(50.00%), %bb.2(50.00%)

  48B       %2:sgpr_32 = COPY %1.sub2:sgpr_128
  64B       dead %3:sgpr_32 = COPY %1.sub1:sgpr_128
  80B       dead %1.sub2:sgpr_128 = COPY %2:sgpr_32
  88B       %1.sub2:sgpr_128 = IMPLICIT_DEF
  96B       %1.sub0:sgpr_128 = IMPLICIT_DEF
  128B      S_CBRANCH_SCC1 %bb.1, implicit undef $scc
  144B      S_BRANCH %bb.2

  160B    bb.2:
          ; predecessors: %bb.1

After 2nd move from 64B to 92B, LIS is broken:

  %1 [16r,32B:0)[32B,80r:4)[80r,80d:3)[88r,96r:1)[96r,160B:2)  0 at 16r 1 at 88r 2 at 96r 3 at 80r 4 at 32B-phi
  L000000000000000C [16r,32B:0)[32B,92r:1)  0 at 16r 1 at 32B-phi
  L00000000000000C0 [16r,16d:0)  0 at 16r
  L0000000000000003 [16r,16d:1)[96r,96d:0)  0 at 96r 1 at 16r
  L0000000000000030 [16r,32B:2)[32B,48r:3)[80r,80d:1)[88r,160B:0)  0 at 88r 1 at 80r 2 at 16r 3 at 32B-phi
  weight:0.000000e+00
  %2 [48r,80r:0)  0 at 48r weight:0.000000e+00
  %3 [92r,92d:0)  0 at 92r weight:0.000000e+00
  RegMasks:
  ********** MACHINEINSTRS **********
  # Machine code for function func: NoPHIs

  0B      bb.0:
            successors: %bb.1(0x80000000); %bb.1(100.00%)

  16B       %1:sgpr_128 = IMPLICIT_DEF

  32B     bb.1:
          ; predecessors: %bb.0, %bb.1
            successors: %bb.1(0x40000000), %bb.2(0x40000000); %bb.1(50.00%), %bb.2(50.00%)

  48B       %2:sgpr_32 = COPY %1.sub2:sgpr_128
  80B       dead %1.sub2:sgpr_128 = COPY %2:sgpr_32
  88B       %1.sub2:sgpr_128 = IMPLICIT_DEF
  92B       dead %3:sgpr_32 = COPY %1.sub1:sgpr_128
  96B       %1.sub0:sgpr_128 = IMPLICIT_DEF
  128B      S_CBRANCH_SCC1 %bb.1, implicit undef $scc
  144B      S_BRANCH %bb.2

  160B    bb.2:
          ; predecessors: %bb.1

  *** Bad machine code: A Subrange is not covered by the main range ***
  - function:    func
  - interval:    %1 [16r,32B:0)[32B,80r:4)[80r,80d:3)[88r,96r:1)[96r,160B:2)  0 at 16r 1 at 88r 2 at 96r 3 at 80r 4 at 32B-phi L000000000000000C [16r,32B:0)[32B,92r:1)  0 at 16r 1 at 32B-phi L00000000000000C0 [16r,16d:0)  0 at 16r L0000000000000003 [16r,16d:1)[96r,96d:0)  0 at 96r 1 at 16r L0000000000000030 [16r,32B:2)[32B,48r:3)[80r,80d:1)[88r,160B:0)  0 at 88r 1 at 80r 2 at 16r 3 at 32B-phi weight:0.000000e+00
  LLVM ERROR: Found 1 machine code errors.

The difference after the patch which occurs after the 2nd move:

  %1 [16r,32B:0)[32B,80r:4)[80r,88r:3)[88r,96r:2)[96r,160B:1)  0 at 16r 1 at 96r 2 at 88r 3 at 80r 4 at 32B-phi

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82916/new/

https://reviews.llvm.org/D82916