Index: lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
+++ lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
@@ -122,14 +122,17 @@
 
     VectorType *VT = dyn_cast<VectorType>(ArgTy);
     bool IsV3 = VT && VT->getNumElements() == 3;
+    bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType();
+
     VectorType *V4Ty = nullptr;
 
     int64_t AlignDownOffset = alignDown(EltOffset, 4);
     int64_t OffsetDiff = EltOffset - AlignDownOffset;
-    unsigned AdjustedAlign = MinAlign(KernArgBaseAlign, AlignDownOffset);
+    unsigned AdjustedAlign = MinAlign(DoShiftOpt ? AlignDownOffset : EltOffset,
+                                      KernArgBaseAlign);
 
     Value *ArgPtr;
-    if (Size < 32 && !ArgTy->isAggregateType()) { // FIXME: Handle aggregate types
+    if (DoShiftOpt) { // FIXME: Handle aggregate types
       // Since we don't have sub-dword scalar loads, avoid doing an extload by
       // loading earlier than the argument address, and extracting the relevant
       // bits.
@@ -147,7 +150,7 @@
     } else {
       ArgPtr = Builder.CreateConstInBoundsGEP1_64(
         KernArgSegment,
-        AlignDownOffset,
+        EltOffset,
         Arg.getName() + ".kernarg.offset");
       ArgPtr = Builder.CreateBitCast(ArgPtr, ArgTy->getPointerTo(AS),
                                      ArgPtr->getName() + ".cast");
@@ -198,7 +201,7 @@
 
     // TODO: Convert noalias arg to !noalias
 
-    if (Size < 32 && !ArgTy->isAggregateType()) {
+    if (DoShiftOpt) {
       Value *ExtractBits = OffsetDiff == 0 ?
         Load : Builder.CreateLShr(Load, OffsetDiff * 8);
 
Index: test/CodeGen/AMDGPU/kernel-args.ll
===================================================================
--- test/CodeGen/AMDGPU/kernel-args.ll
+++ test/CodeGen/AMDGPU/kernel-args.ll
@@ -739,10 +739,10 @@
 ; multiple.
 ; FUNC-LABEL: {{^}}packed_struct_argument_alignment:
 ; HSA-GFX9: kernarg_segment_byte_size = 28
+; HSA-GFX9: global_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:13
+; HSA-GFX9: global_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:17
 ; HSA-GFX9: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
 ; HSA-GFX9: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x4
-; HSA-GFX9: s_load_dword s{{[0-9]+}}, s[4:5], 0xc
-; HSA-GFX9: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x10
 define amdgpu_kernel void @packed_struct_argument_alignment(<{i32, i64}> %arg0, i8, <{i32, i64}> %arg1) {
   %val0 = extractvalue <{i32, i64}> %arg0, 0
   %val1 = extractvalue <{i32, i64}> %arg0, 1
@@ -789,10 +789,18 @@
 ; FIXME: Why not all scalar loads?
 ; GCN-LABEL: {{^}}array_3xi16:
 ; HSA-GFX9: global_load_ushort v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:2
-; HSA-GFX9: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
-; HSA-GFX9: s_load_dword s{{[0-9]+}}, s[4:5], 0x4
+; HSA-GFX9: global_load_ushort v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:4
+; HSA-GFX9: global_load_ushort v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:6
 define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
   store volatile i8 %arg0, i8 addrspace(1)* undef
   store volatile [3 x i16] %arg1, [3 x i16] addrspace(1)* undef
   ret void
 }
+
+; GCN-LABEL: {{^}}small_array_round_down_offset:
+; HSA-GFX9: global_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:1
+define amdgpu_kernel void @small_array_round_down_offset(i8, [1 x i8] %arg) {
+  %val = extractvalue [1 x i8] %arg, 0
+  store volatile i8 %val, i8 addrspace(1)* undef
+  ret void
+}