[PATCH] D105827: [SLP]Workaround for InsertSubVector cost.

Thu Jul 15 14:19:46 PDT 2021

lebedev.ri added a comment.

In D105827#2881468 <https://reviews.llvm.org/D105827#2881468>, @ABataev wrote:

> In D105827#2880704 <https://reviews.llvm.org/D105827#2880704>, @aeubanks wrote:
>
>> This is causing hangs
>> `$ clang++ -O2 -c reduced2.ll -o /dev/null`
>> F17952182: reduced2.ll <https://reviews.llvm.org/F17952182>
>
> Investigated it, actually hangs in `X86 DAG->DAG Instruction Selection on function (_ZN12SpinningCube6UpdateEv)` even without SLP vectorizer with the attached reproducer (hangs with 12.0.0 and trunk, 11.0.1 does not recognize `poison` but passes if I replace `poison` by `undef`. 12.0.0 and trunk still hangs even in this case). Command to reproduce is `llc reduced.ll -o /dev/null`. Looks like the bug was introduced somewhat between llvm 11 and 12. Hangs in the loop in lines 1523-1598, `lib/CodeGen/SelectionDAG/DAGCombiner.cpp` F17954903: reduced.ll <https://reviews.llvm.org/F17954903>

Looks like endless X86 shuffle combine loop

  Legalizing: t89: v4f32 = X86ISD::SHUFP t87, t80, TargetConstant:i8<36>
  Legal node: nothing to do

  Combining: t89: v4f32 = X86ISD::SHUFP t87, t80, TargetConstant:i8<36>
  Creating fp constant: t87: v4f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>
  Creating fp constant: t87: v4f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>
  Creating new node: t1580: v4f32 = undef
  Creating fp constant: t87: v4f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>
  Creating new node: t1581: v4f32 = X86ISD::UNPCKL t68, undef:v4f32

  Replacing.2 t1578: v4f32 = X86ISD::SHUFP t87, t68, TargetConstant:i8<-44>

  With: t1581: v4f32 = X86ISD::UNPCKL t68, undef:v4f32

  Legalizing: t1581: v4f32 = X86ISD::UNPCKL t68, undef:v4f32
  Legal node: nothing to do

  Combining: t1581: v4f32 = X86ISD::UNPCKL t68, undef:v4f32
  Creating fp constant: t87: v4f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>
  Creating constant: t1582: i8 = TargetConstant<-44>
  Creating new node: t1583: v4f32 = X86ISD::SHUFP t87, t68, TargetConstant:i8<-44>
   ... into: t1583: v4f32 = X86ISD::SHUFP t87, t68, TargetConstant:i8<-44>

  Legalizing: t1583: v4f32 = X86ISD::SHUFP t87, t68, TargetConstant:i8<-44>
  Legal node: nothing to do

  Combining: t1583: v4f32 = X86ISD::SHUFP t87, t68, TargetConstant:i8<-44>
  Creating new node: t1584: v4f32 = undef
  Creating fp constant: t87: v4f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>

  Legalizing: t1582: i8 = TargetConstant<-44>

  Combining: t1582: i8 = TargetConstant<-44>

  Legalizing: t89: v4f32 = X86ISD::SHUFP t87, t80, TargetConstant:i8<36>
  Legal node: nothing to do

  Combining: t89: v4f32 = X86ISD::SHUFP t87, t80, TargetConstant:i8<36>
  <here we go again>

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105827/new/

https://reviews.llvm.org/D105827