Now that tailduplication does not bork SSA form, run mem2reg earlier in gccas.

This tremendously improves the code generated by the LLVM optimizer, primarily
by making the inliner more aggressive.  For example, it improves the stepanov
benchmark from 55.56 mega-additions/sec to 98.04 Ma/s.  It also improves the
oopack/iterator benchmark from 338.3MFLOPS/s to 1103.4MFLOPS/s.  Less noteworthy,
it improves oopack/matrix from 573 -> 641 MFLOPS/s.

llvm-svn: 11053

1 files changed, 4 insertions, 2 deletions

diff --git a/llvm/tools/gccas/gccas.cpp b/llvm/tools/gccas/gccas.cpp
index 40053303db6..1abf251fce3 100644
--- a/llvm/tools/gccas/gccas.cpp
+++ b/llvm/tools/gccas/gccas.cpp
@@ -65,18 +65,20 @@ void AddConfiguredTransformationPasses(PassManager &PM) {
 
   if (DisableOptimizations) return;
 
-  addPass(PM, createCFGSimplificationPass());    // Clean up disgusting code
   addPass(PM, createRaiseAllocationsPass());     // call %malloc -> malloc inst
+  addPass(PM, createCFGSimplificationPass());    // Clean up disgusting code
+  addPass(PM, createPromoteMemoryToRegister());  // Kill useless allocas
   addPass(PM, createGlobalDCEPass());            // Remove unused globals
   addPass(PM, createIPConstantPropagationPass());// IP Constant Propagation
   addPass(PM, createDeadArgEliminationPass());   // Dead argument elimination
+  addPass(PM, createInstructionCombiningPass()); // Clean up after IPCP & DAE
+  addPass(PM, createCFGSimplificationPass());    // Clean up after IPCP & DAE
 
   addPass(PM, createPruneEHPass());              // Remove dead EH info
 
   if (!DisableInline)
     addPass(PM, createFunctionInliningPass());   // Inline small functions
 
-  addPass(PM, createInstructionCombiningPass()); // Cleanup code for raise
   addPass(PM, createRaisePointerReferencesPass());// Recover type information
   addPass(PM, createTailDuplicationPass());      // Simplify cfg by copying code
   addPass(PM, createCFGSimplificationPass());    // Merge & remove BBs