Reliably reproduces on x86_32 with vector tests, but the issue looks pre-existing.

$ CONF=linux-x86-server-fastdebug make run-test TEST=jdk/incubator/vector/Float128VectorTests.java TEST_VM_OPTS="-XX:UseSSE=0 -XX:UseAVX=0"

test Float128VectorTests.SINHFloat128VectorTests(float[cornerCaseValue(i)]): success
o39 SqrtF === _ o12 [[o40 2 ]]

--N: o39 SqrtF === _ o12 [[o40 2 ]]

   --N: o12 Parm === o3 [[o39 ]] Parm2: float
   REGFPR 0 REGFPR
   REGFPR1 0 REGFPR1
   REGF 100 MoveVL2F
   LEGREGF 200 MoveF2LEG
   VLREGF 0 VLREGF

# To suppress the following error report, specify this argument
# after -XX: or in .hotspotrc: SuppressErrorAt=/matcher.cpp:1716

The current match rules are:

// sqrtss instruction needs destination register to be pre initialized for best performance
// Therefore only the instruct rule where the input is pre-loaded into dst register is defined below
instruct sqrtF_reg(regF dst) %{
  predicate(UseSSE>=1);
  match(Set dst (SqrtF dst));
  format %{ "sqrtss $dst, $dst" %}
  ins_encode %{
    __ sqrtss($dst$$XMMRegister, $dst$$XMMRegister);
  %}
  ins_pipe(pipe_slow);
%}

The node is generated here:

//------------------------------Ideal------------------------------------------
// If we see pattern ConvF2D SomeDoubleOp ConvD2F, do operation as float.
Node *ConvD2FNode::Ideal(PhaseGVN *phase, bool can_reshape) {
  if ( in(1)->Opcode() == Op_SqrtD ) {
    Node* sqrtd = in(1);
    if ( sqrtd->in(1)->Opcode() == Op_ConvF2D ) {
      if ( Matcher::match_rule_supported(Op_SqrtF) ) {
        Node* convf2d = sqrtd->in(1);
        return new SqrtFNode(phase->C, sqrtd->in(0), convf2d->in(1));
      }
    }
  }
  return NULL;
}

I actually thought Matcher::match_rule_supported looks at predicate(...) in the match rule above, but apparently it just hard-codes the exceptions! Op_SqrtF (and Op_SqrtD) should be handled there.

This only affects x86_32, as x86_64 always runs with UseSSE >= 2.