diff options
author | Richard Sandiford <rsandifo@linux.vnet.ibm.com> | 2013-08-05 10:58:53 +0000 |
---|---|---|
committer | Richard Sandiford <rsandifo@linux.vnet.ibm.com> | 2013-08-05 10:58:53 +0000 |
commit | 66fbb4781841a8411a772b6909a7e0de182b896f (patch) | |
tree | 282862d1fa64afb72878b1e7dec031fb42ad3d3f /lib/Target/SystemZ/SystemZTargetMachine.cpp | |
parent | 13e6e9171f79a481d7f814aad958460dfd867c71 (diff) | |
download | external_llvm-66fbb4781841a8411a772b6909a7e0de182b896f.zip external_llvm-66fbb4781841a8411a772b6909a7e0de182b896f.tar.gz external_llvm-66fbb4781841a8411a772b6909a7e0de182b896f.tar.bz2 |
[SystemZ] Split out comparison elimination into a separate pass
Perhaps predictably, doing comparison elimination on the fly during
SystemZLongBranch turned out to be a bad idea. The next patches make
use of LOAD AND TEST and BRANCH ON COUNT, both of which require
changes to earlier instructions.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187718 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/SystemZ/SystemZTargetMachine.cpp')
-rw-r--r-- | lib/Target/SystemZ/SystemZTargetMachine.cpp | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/lib/Target/SystemZ/SystemZTargetMachine.cpp b/lib/Target/SystemZ/SystemZTargetMachine.cpp index 437ea61..2bacc2b 100644 --- a/lib/Target/SystemZ/SystemZTargetMachine.cpp +++ b/lib/Target/SystemZ/SystemZTargetMachine.cpp @@ -65,6 +65,28 @@ bool SystemZPassConfig::addPreSched2() { } bool SystemZPassConfig::addPreEmitPass() { + // We eliminate comparisons here rather than earlier because some + // transformations can change the set of available CC values and we + // generally want those transformations to have priority. This is + // especially true in the commonest case where the result of the comparison + // is used by a single in-range branch instruction, since we will then + // be able to fuse the compare and the branch instead. + // + // For example, two-address NILF can sometimes be converted into + // three-address RISBLG. NILF produces a CC value that indicates whether + // the low word is zero, but RISBLG does not modify CC at all. On the + // other hand, 64-bit ANDs like NILL can sometimes be converted to RISBG. + // The CC value produced by NILL isn't useful for our purposes, but the + // value produced by RISBG can be used for any comparison with zero + // (not just equality). So there are some transformations that lose + // CC values (while still being worthwhile) and others that happen to make + // the CC result more useful than it was originally. + // + // Doing it so late makes it more likely that a register will be reused + // between the comparison and the branch, but it isn't clear whether + // preventing that would be a win or not. + if (getOptLevel() != CodeGenOpt::None) + addPass(createSystemZElimComparePass(getSystemZTargetMachine())); addPass(createSystemZLongBranchPass(getSystemZTargetMachine())); return true; } |