FTL: split overflow checks into non-overflow arithmetic and an additional call to the overflow intrinsic check.
https://bugs.webkit.org/show_bug.cgi?id=122170

Patch by Nadav Rotem <nrotem@apple.com> on 2013-10-01
Reviewed by Filip Pizlo.

Overflow intrinsics are preventing SCEV and other LLVM analysis passes from analyzing loops. This patch changes the FTL-IR gen by splitting arithmetic calculations into two parts:
1. Generate the arithmetic calculation (that may overflow)
2. Generate the overflow check (that is only used by the OSR-exit logic).

We trust LLVM (SelectionDAG) to merge these calculations into a single opcode.

This JS function:

function foo() {
    for (i=0; i < 10000000; i++) { }
}

Is now compiled into this LLVM-IR:

"OSR exit continuation for @24<Int32>":           ; preds = %"Block #0", %"OSR exit continuation for @24<Int32>2"
  %4 = phi i64 [ %10, %"OSR exit continuation for @24<Int32>2" ], [ -281474976710656, %"Block #0" ]
  %5 = trunc i64 %4 to i32
  %6 = add i32 %5, 1
  %7 = tail call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %5, i32 1)
  %8 = extractvalue { i32, i1 } %7, 1
  br i1 %8, label %"OSR exit failCase for @24<Int32>1", label %"OSR exit continuation for @24<Int32>2"

 And into this assembly:

LBB0_1:                                 ## %OSR exit continuation for @24<Int32>
                                ## =>This Inner Loop Header: Depth=1
    movl  %ecx, %esi
    incl  %esi
    jo  LBB0_4

* ftl/FTLLowerDFGToLLVM.cpp:
(JSC::FTL::LowerDFGToLLVM::compileAddSub):
(JSC::FTL::LowerDFGToLLVM::compileArithMul):
(JSC::FTL::LowerDFGToLLVM::compileArithNegate):

git-svn-id: http://svn.webkit.org/repository/webkit/trunk@156746 268f45cc-cd09-0410-ab3c-d52691b4dbfc