[llvm] r207126 - R600/SI: Use address space in allowsUnalignedMemoryAccesses

Matt Arsenault Matthew.Arsenault at amd.com
Thu Apr 24 10:08:26 PDT 2014


Author: arsenm
Date: Thu Apr 24 12:08:26 2014
New Revision: 207126

URL: http://llvm.org/viewvc/llvm-project?rev=207126&view=rev
Log:
R600/SI: Use address space in allowsUnalignedMemoryAccesses

Modified:
    llvm/trunk/lib/Target/R600/SIISelLowering.cpp

Modified: llvm/trunk/lib/Target/R600/SIISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIISelLowering.cpp?rev=207126&r1=207125&r2=207126&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/R600/SIISelLowering.cpp Thu Apr 24 12:08:26 2014
@@ -223,10 +223,40 @@ SITargetLowering::SITargetLowering(Targe
 bool SITargetLowering::allowsUnalignedMemoryAccesses(EVT  VT,
                                                      unsigned AddrSpace,
                                                      bool *IsFast) const {
+  if (IsFast)
+    *IsFast = false;
+
   // XXX: This depends on the address space and also we may want to revist
   // the alignment values we specify in the DataLayout.
+
+  // TODO: I think v3i32 should allow unaligned accesses on CI with DS_READ_B96,
+  // which isn't a simple VT.
   if (!VT.isSimple() || VT == MVT::Other)
     return false;
+
+  // XXX - CI changes say "Support for unaligned memory accesses" but I don't
+  // see what for specifically. The wording everywhere else seems to be the
+  // same.
+
+  // 3.6.4 - Operations using pairs of VGPRs (for example: double-floats) have
+  // no alignment restrictions.
+  if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) {
+    // Using any pair of GPRs should be the same as any other pair.
+    if (IsFast)
+      *IsFast = true;
+    return VT.bitsGE(MVT::i64);
+  }
+
+  // XXX - The only mention I see of this in the ISA manual is for LDS direct
+  // reads the "byte address and must be dword aligned". Is it also true for the
+  // normal loads and stores?
+  if (AddrSpace == AMDGPUAS::LOCAL_ADDRESS)
+    return false;
+
+  // 8.1.6 - For Dword or larger reads or writes, the two LSBs of the
+  // byte-address are ignored, thus forcing Dword alignment.
+  if (IsFast)
+    *IsFast = true;
   return VT.bitsGT(MVT::i32);
 }
 





More information about the llvm-commits mailing list