Host-dependent linearscan regalloc spiller

[llvmbot]

Bugzilla Link	9590
Resolution	FIXED
Resolved on	Mar 31, 2011 07:13
Version	2.9
OS	Windows XP
Blocks	llvm/llvm-bugzilla-archive#9591
Reporter	LLVM Bugzilla Contributor
CC	@stoklund

Extended Description

$ g++ --version
g++.exe (TDM-1 mingw32) 4.4.0

$ bin/clang++.exe --version
clang version 2.9 (git://github.com/chapuni/clang.git http://llvm.org/git/clang.git ssh://chapuni@192.168.1.193/home/chapuni/clang.git b769bbbda7024bb990e8bad698e414b8e386ba5a)
Target: i686-pc-mingw32
Thread model: posix

--

With g++, std::pow(float,float) aka __builtin_powf() is expanded to pow().
With clang, std::pow(float,float) is still powf().

Due to precision issue, result of -regalloc=linearscan might not be identical.
(I have not tried other regallocs with clang)

Workaround is as below;

--- a/lib/CodeGen/LiveIntervalAnalysis.cpp
+++ b/lib/CodeGen/LiveIntervalAnalysis.cpp
@@ -1703,7 +1703,7 @@ LiveIntervals::getSpillWeight(bool isDef, bool isUse, unsigned loopDepth) {
// overflow a float. This expression behaves like 10^d for small d, but is
// more tempered for large d. At d=200 we get 6.7e33 which leaves a bit of
// headroom before overflow.

• float lc = std::pow(1 + (100.0f / (loopDepth+10)), (float)loopDepth);

• float lc = ::pow(1 + (100.0f / (loopDepth+10)), (float)loopDepth);

  return (isDef + isUse) * lc;
  }

[stoklund]

Takumi, your patch looks good to me. Please commit, but also change the arguments to double precision. That is, use 100.0 and (double)loopDepth.

[llvmbot]

Committed in r128629, thank you, Jakob!

On Cygwin, also StackSlotColoring might not be identical between gcc and clang.

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#9591