Remove the rest of the *_sfp Neon instruction patterns.

Use the same COPY_TO_REGCLASS approach as for the 2-register *_sfp instructions.
This change made a big difference in the code generated for the
CodeGen/Thumb2/cross-rc-coalescing-2.ll test: The coalescer is still doing
a fine job, but some instructions that were previously moved outside the loop
are not moved now.  It's using fewer VFP registers now, which is generally
a good thing, so I think the estimates for register pressure changed and that
affected the LICM behavior.  Since that isn't obviously wrong, I've just
changed the test file.  This completes the work for Radar 8711675.

llvm-svn: 121730

1 files changed, 0 insertions, 5 deletions

diff --git a/llvm/utils/TableGen/ARMDecoderEmitter.cpp b/llvm/utils/TableGen/ARMDecoderEmitter.cpp
index c4e78f358ec..10e507f603b 100644
--- a/llvm/utils/TableGen/ARMDecoderEmitter.cpp
+++ b/llvm/utils/TableGen/ARMDecoderEmitter.cpp
@@ -1613,11 +1613,6 @@ ARMDEBackend::populateInstruction(const CodeGenInstruction &CGI,
         Name == "VNEGScc")
       return false;
 
-    // Ignore the *_sfp instructions when decoding.  They are used by the
-    // compiler to implement scalar floating point operations using vector
-    // operations in order to work around some performance issues.
-    if (Name.find("_sfp") != std::string::npos) return false;
-
     // LDMIA_RET is a special case of LDM (Load Multiple) where the registers
     // loaded include the PC, causing a branch to a loaded address.  Ignore
     // the LDMIA_RET instruction when decoding.