AMDGPU/SI: Don't allow unaligned scratch access

Summary: The hardware doesn't support this.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye

Differential Revision: https://reviews.llvm.org/D25523

llvm-svn: 284257
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index e0bde33..5f7033e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -67,6 +67,12 @@
   "Support unaligned global loads and stores"
 >;
 
+def FeatureUnalignedScratchAccess : SubtargetFeature<"unaligned-scratch-access",
+  "UnalignedScratchAccess",
+  "true",
+  "Support unaligned scratch loads and stores"
+>;
+
 def FeatureXNACK : SubtargetFeature<"xnack",
   "EnableXNACK",
   "true",
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index aba1263..bd69847 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -85,6 +85,7 @@
     FP64Denormals(false),
     FPExceptions(false),
     FlatForGlobal(false),
+    UnalignedScratchAccess(false),
     UnalignedBufferAccess(false),
 
     EnableXNACK(false),
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 1dd6506..5682d99 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -76,6 +76,7 @@
   bool FP64Denormals;
   bool FPExceptions;
   bool FlatForGlobal;
+  bool UnalignedScratchAccess;
   bool UnalignedBufferAccess;
   bool EnableXNACK;
   bool DebuggerInsertNops;
@@ -277,6 +278,10 @@
     return UnalignedBufferAccess;
   }
 
+  bool hasUnalignedScratchAccess() const {
+    return UnalignedScratchAccess;
+  }
+
   bool isXNACKEnabled() const {
     return EnableXNACK;
   }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index be669c1..e632951 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -459,6 +459,15 @@
     return AlignedBy4;
   }
 
+  // FIXME: We have to be conservative here and assume that flat operations
+  // will access scratch.  If we had access to the IR function, then we
+  // could determine if any private memory was used in the function.
+  if (!Subtarget->hasUnalignedScratchAccess() &&
+      (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
+       AddrSpace == AMDGPUAS::FLAT_ADDRESS)) {
+    return false;
+  }
+
   if (Subtarget->hasUnalignedBufferAccess()) {
     // If we have an uniform constant load, it still requires using a slow
     // buffer instruction if unaligned.