RFR(S): 8243670: Unexpected test result caused by C2 MergeMemNode::Ideal

[Yangfei (Felix)]

Hi,

  Please help review this patch fixing a C2 issue.  
      Bug: https://bugs.openjdk.java.net/browse/JDK-8243670 
      Webrev: http://cr.openjdk.java.net/~fyang/8243670/webrev.00/

As described on the issue, C2 generates incorrect code for the following OSR compile:
    420    4 %  b  4       TestReplaceEquivPhis::test @ 25 (107 bytes)

  v = iFld;      // load from field "iFld"             
  iFld = TestReplaceEquivPhis.instanceCount; // store to field "iFld"

From the C2 JIT code, load and store of field "iFld" are misplaced.
Looks like this is initially caused by the replace equivalent phis transformation in MergeMemNode::Ideal.

Call trace:
#0  MergeMemNode::Ideal (this=0x7fff580c70c0, phase=0x7fff7d9407b0, can_reshape=true) at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/memnode.cpp:4621
#1  0x00007ffff6020bcd in PhaseGVN::apply_ideal (this=0x7fff7d9407b0, k=0x7fff580c70c0, can_reshape=true)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/phaseX.cpp:806
#2  0x00007ffff60223ef in PhaseIterGVN::transform_old (this=0x7fff7d9407b0, n=0x7fff580c70c0)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/phaseX.cpp:1229
#3  0x00007ffff602217a in PhaseIterGVN::optimize (this=0x7fff7d9407b0) at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/phaseX.cpp:1175
#4  0x00007ffff5e32618 in PhaseIdealLoop::build_and_optimize (this=0x7fff7d93fa90, mode=LoopOptsDefault)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/loopnode.cpp:3192
#5  0x00007ffff5800831 in PhaseIdealLoop::PhaseIdealLoop (this=0x7fff7d93fa90, igvn=..., mode=LoopOptsDefault)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/loopnode.hpp:951
#6  0x00007ffff580092c in PhaseIdealLoop::optimize (igvn=..., mode=LoopOptsDefault) at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/loopnode.hpp:1026
#7  0x00007ffff57f4553 in Compile::optimize_loops (this=0x7fff7d942d00, igvn=..., mode=LoopOptsDefault)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/compile.cpp:1970
#8  0x00007ffff57f5308 in Compile::Optimize (this=0x7fff7d942d00) at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/compile.cpp:2182
#9  0x00007ffff57ee5a3 in Compile::Compile (this=0x7fff7d942d00, ci_env=0x7fff7d943810, target=0x7fff580eeb70, osr_bci=25, subsume_loads=true,
    do_escape_analysis=true, eliminate_boxing=true, directive=0x7ffff031b430) at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/compile.cpp:736
#10 0x00007ffff56ebc23 in C2Compiler::compile_method (this=0x7ffff035e940, env=0x7fff7d943810, target=0x7fff580eeb70, entry_bci=25, directive=0x7ffff031b430)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/opto/c2compiler.cpp:111
#11 0x00007ffff5808cd0 in CompileBroker::invoke_compiler_on_method (task=0x7ffff03b6a10)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/compiler/compileBroker.cpp:2210
#12 0x00007ffff58079eb in CompileBroker::compiler_thread_loop () at /home/yangfei/openjdk-jdk/src/hotspot/share/compiler/compileBroker.cpp:1894
#13 0x00007ffff62557f1 in compiler_thread_entry (thread=0x7ffff035f800, __the_thread__=0x7ffff035f800)
    at /home/yangfei/openjdk-jdk/src/hotspot/share/runtime/thread.cpp:3454
#14 0x00007ffff6250a11 in JavaThread::thread_main_inner (this=0x7ffff035f800) at /home/yangfei/openjdk-jdk/src/hotspot/share/runtime/thread.cpp:1969
#15 0x00007ffff62508bf in JavaThread::run (this=0x7ffff035f800) at /home/yangfei/openjdk-jdk/src/hotspot/share/runtime/thread.cpp:1952
#16 0x00007ffff624cb2c in Thread::call_run (this=0x7ffff035f800) at /home/yangfei/openjdk-jdk/src/hotspot/share/runtime/thread.cpp:399
#17 0x00007ffff5fbf288 in thread_native_entry (thread=0x7ffff035f800) at /home/yangfei/openjdk-jdk/src/hotspot/os/linux/os_linux.cpp:789
#18 0x00007ffff71976db in start_thread (arg=0x7fff7d944700) at pthread_create.c:463
#19 0x00007ffff78f188f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

graph before the transformation looks like:

  1: Phi1( Phi1 and Phi2 have the same input edges)  #memory  Memory: @TestReplaceEquivPhis+12 *, name=iFld, idx=5;
  2: Phi2( Phi1 and Phi2 have the same input edges)  #memory  Memory: @BotPTR *+bot, idx=Bot;
  3: LoadI( input: 1)   =>  name=iFld, idx=5
  4: MergeMem( input:1, 2)
  5: MemBarAcqure( input: 4)
  6: Proj( input: 5)
  7: StoreI( input: 6)   => name=iFld, idx=5

Here Phi1 and Phi2 have same input edges.  Input from Phi1 to MergeMem is simplified by MergeMemoryNode::Ideal.

graph after the transformation looks like:

  1: Phi1( Phi1 and Phi2 have the same input edges)  #memory  Memory: @TestReplaceEquivPhis+12 *, name=iFld, idx=5;
  2: Phi2( Phi1 and Phi2 have the same input edges)  #memory  Memory: @BotPTR *+bot, idx=Bot;
  3: LoadI( input: 1)    =>  name=iFld, idx=5
  4: MergeMem( input: 2)
  5: MemBarAcqure( input: 4)
  6: Proj( input: 5)
  7: StoreI( input: 6)    => name=iFld, idx=5

As a result, PhaseCFG::insert_anti_dependences won't insert an anti-dependence edge between 3 and 4 to place the load correctly.
This transformation is there from day one.  With -XX:-SplitIfBlocks option, it triggers more errors.  Proposed webrev simply disables it.
Tier1-3 tested on x86-64-linux-gnu.  Specjbb2015 shows no performance regression with this change.
Suggestions?

Thanks,
Felix