Author: [email protected]
Date: Thu Jan 15 01:29:05 2009
New Revision: 1075
Modified:
branches/experimental/toiger/src/codegen-ia32.cc
Log:
Experimental: processor flags are not preserved across CFG edges
targeting labeled basic blocks. Merge code may be generated at the
block exit which could potentially destroy a value in the flags
register.
This solution is to always materialize the flag result in a register.
The generated code is known to be bad but (more) correct and will have
to be fixed as a future change.
Review URL: http://codereview.chromium.org/18225
Modified: branches/experimental/toiger/src/codegen-ia32.cc
==============================================================================
--- branches/experimental/toiger/src/codegen-ia32.cc (original)
+++ branches/experimental/toiger/src/codegen-ia32.cc Thu Jan 15 01:29:05
2009
@@ -1269,25 +1269,46 @@
}
CompareStub stub(cc, strict);
Result answer = frame_->CallStub(&stub, &right_side, &left_side, 0);
+ ASSERT(answer.is_register() && answer.reg().is_byte_register());
+ frame_->Spill(answer.reg());
if (cc == equal) {
__ test(answer.reg(), Operand(answer.reg()));
} else {
__ cmp(answer.reg(), 0);
}
- answer.Unuse();
+ __ setcc(cc, answer.reg());
+ __ and_(Operand(answer.reg()), Immediate(1));
// The expected frame at JumpTarget "done" is bound to the current frame.
// This current frame is spilled, due to the call to CallStub.
// It would be better if the fast SMI case controlled the expected frame.
- done.Jump();
+ done.Jump(&answer);
is_smi.Bind(&left_side, &right_side);
+ left_side.ToRegister();
+ right_side.ToRegister();
+ // Find a byte register for the flag value.
+ if (!left_side.reg().is_byte_register() &&
+ !right_side.reg().is_byte_register()) {
+ // If we allocate a register here, it should be a byte register
+ // because there is only one non-reserved register on IA32 that
+ // does not have a byte part (namely edi) and it must be left_side
+ // (and right_side).
+ answer = allocator_->Allocate();
+ ASSERT(answer.is_register() && answer.reg().is_byte_register());
+ } else {
+ answer = left_side.reg().is_byte_register() ? left_side : right_side;
+ frame_->Spill(answer.reg());
+ }
__ cmp(left_side.reg(), Operand(right_side.reg()));
right_side.Unuse();
left_side.Unuse();
- // Fall through to |done|.
+ __ setcc(cc, answer.reg());
+ __ and_(Operand(answer.reg()), Immediate(1));
- done.Bind();
- cc_reg_ = cc;
+ done.Bind(&answer);
+ answer.ToRegister();
+ __ test(answer.reg(), Operand(answer.reg()));
+ cc_reg_ = not_zero;
}
@@ -1324,11 +1345,12 @@
ASSERT(value.is_valid());
__ Set(value.reg(), Immediate(Smi::FromInt(int_value_)));
Result result = cgen->frame()->CallStub(&stub, &argument, &value, 0);
- ASSERT(result.is_register());
+ ASSERT(result.is_register() && result.reg().is_byte_register());
+ cgen->frame()->Spill(result.reg());
__ cmp(result.reg(), 0);
- result.Unuse();
- // The actual result is returned in the flags.
- exit()->Jump();
+ __ setcc(cc_, result.reg());
+ __ and_(Operand(result.reg()), Immediate(1));
+ exit()->Jump(&result);
}
@@ -1348,11 +1370,28 @@
comparee.ToRegister();
__ test(comparee.reg(), Immediate(kSmiTagMask));
deferred->enter()->Branch(not_zero, &comparee, not_taken);
+ // Find a byte register to hold the flag value. If the comparee is
+ // a byte register we can use it, otherwise allocating will give us
+ // one since there is only one non-reserved IA32 register that does
+ // not have a byte part (namely edi).
+ Result flag(this);
+ if (comparee.reg().is_byte_register()) {
+ flag = comparee;
+ frame_->Spill(flag.reg());
+ } else {
+ flag = allocator_->Allocate();
+ ASSERT(flag.is_register() && flag.reg().is_byte_register());
+ }
// Test smi equality and comparison by signed int comparison.
__ cmp(Operand(comparee.reg()), Immediate(value));
comparee.Unuse();
- deferred->exit()->Bind();
- cc_reg_ = cc;
+ __ setcc(cc, flag.reg());
+ __ and_(Operand(flag.reg()), Immediate(1));
+
+ deferred->exit()->Bind(&flag);
+ flag.ToRegister();
+ __ test(flag.reg(), Operand(flag.reg()));
+ cc_reg_ = not_zero;
}
@@ -3590,7 +3629,7 @@
void CodeGenerator::GenerateIsArray(ZoneList<Expression*>* args) {
ASSERT(args->length() == 1);
LoadAndSpill(args->at(0));
- JumpTarget answer(this);
+ Label answer;
// We need the CC bits to come out as not_equal in the case where the
// object is a smi. This can't be done with the usual test opcode so
// we copy the object to ecx and do some destructive ops on it that
@@ -3599,13 +3638,13 @@
__ mov(ecx, Operand(eax));
__ and_(ecx, kSmiTagMask);
__ xor_(ecx, kSmiTagMask);
- answer.Branch(not_equal, not_taken);
+ __ j(not_equal, &answer, not_taken);
// It is a heap object - get map.
__ mov(eax, FieldOperand(eax, HeapObject::kMapOffset));
__ movzx_b(eax, FieldOperand(eax, Map::kInstanceTypeOffset));
// Check if the object is a JS array or not.
__ cmp(eax, JS_ARRAY_TYPE);
- answer.Bind();
+ __ bind(&answer);
cc_reg_ = equal;
}
--~--~---------~--~----~------------~-------~--~----~
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
-~----------~----~----~----~------~----~------~--~---
