P4 to Git Change 1288106 by jsjodin@jsjodin-git2p4-llvm on 2016/07/06 18:17:28

SWDEV-3 - AMDGPU: Improve load/store of illegal types.

	There was a combine before to handle the simple copy case.
	Split this into handling loads and stores separately.

	We might want to change how this handles some of the vector
	extloads, since this can result in large code size increases.

	git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@274394 91177308-0d34-0410-b5e6-96231b3b80d8

	GitHash: df587174ebb04a30c8b55aef601121c25af11986

Affected files ...

... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/lib/Target/AMDGPU/AMDGPUISelLowering.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/lib/Target/AMDGPU/AMDGPUISelLowering.h#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/lib/Target/AMDGPU/SIISelLowering.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/add.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/copy-illegal-type.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/ctpop64.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/half.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/insert_vector_elt.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-constant-i16.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-constant-i32.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-constant-i64.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-constant-i8.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-global-f64.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-global-i16.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-global-i32.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-global-i64.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-global-i8.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-local-f32.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-local-i16.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-local-i8.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/load-weird-sizes.ll#1 add
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/scalar_to_vector.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/sext-in-reg.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/store-barrier.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/store.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/sub.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/trunc-bitcast-vector.ll#2 edit
... //depot/stg/opencl/drivers/opencl/compiler/llvm.git/test/CodeGen/AMDGPU/trunc-store.ll#2 edit

rocclr/runtime/device/pal/palresource.cpp

@@ -1072,10 +1072,9 @@ Resource::free()
 void
 Resource::writeRawData(
     VirtualGPU& gpu,
-    size_t      offset,
-    size_t      size,
+    size_t size,
     const void* data,
-    bool        waitForEvent) const
+    bool waitForEvent) const
 {
     GpuEvent    event;
 
@@ -1083,8 +1082,11 @@ Resource::writeRawData(
     // size needs to be DWORD aligned
     assert((size & 3) == 0);
     gpu.eventBegin(MainEngine);
+    //! @todo Remove cache flush
+    //! It's a workaround for a PAL crash with embedded data, allocated before any command
+    gpu.flushCUCaches();
     gpu.queue(MainEngine).addCmdMemRef(iMem());
-    gpu.iCmd()->CmdUpdateMemory(*iMem(), offset, size, reinterpret_cast<const uint32_t*>(data));
+    gpu.iCmd()->CmdUpdateMemory(*iMem(), 0, size, reinterpret_cast<const uint32_t*>(data));
     gpu.eventEnd(MainEngine, event);
 
     setBusy(gpu, event);
@@ -1936,7 +1938,7 @@ Resource::warmUpRenames(VirtualGPU& gpu)
         uint    dummy = 0;
         const bool NoWait = false;
         // Write 0 for the buffer paging by VidMM
-        writeRawData(gpu, 0, sizeof(dummy), &dummy, NoWait);
+        writeRawData(gpu, sizeof(dummy), &dummy, NoWait);
         const bool Force = true;
         rename(gpu, Force);
     }