diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -912,6 +912,22 @@
   }
 
   if (!FoldingImmLike) {
+    if (OpToFold.isReg() && ST->needsAlignedVGPRs()) {
+      // Don't fold if OpToFold doesn't hold an aligned register.
+      const TargetRegisterClass *RC =
+          TRI->getRegClassForReg(*MRI, OpToFold.getReg());
+      if (TRI->hasVectorRegisters(RC) && OpToFold.getSubReg()) {
+        unsigned SubReg = OpToFold.getSubReg();
+        const TargetRegisterClass *SubRC = TRI->getSubRegClass(RC, SubReg);
+        RC = TRI->getCompatibleSubRegClass(RC, SubRC, SubReg);
+        if (RC)
+          RC = SubRC;
+      }
+
+      if (!RC || !TRI->isProperlyAlignedRC(*RC))
+        return;
+    }
+
     tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII);
 
     // FIXME: We could try to change the instruction from 64-bit to 32-bit
diff --git a/llvm/test/CodeGen/AMDGPU/si-fold-aligned-agprs.mir b/llvm/test/CodeGen/AMDGPU/si-fold-aligned-agprs.mir
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-aligned-agprs.mir
@@ -0,0 +1,244 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX90A %s
+
+# Should fold only the aligned agpr copies with their uses.
+
+---
+name:            aligned_agpr_64
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_64
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_64_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_64
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_64
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_64 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_64 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_96_sub0_subg1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_96_sub0_subg1
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_96_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_96_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_96_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_96_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_96_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_96_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_96
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_96
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_96_align2 = IMPLICIT_DEF
+    %2:areg_96_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_96
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_96
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_96 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_96 = IMPLICIT_DEF
+    %2:areg_96_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_128_sub0_sub1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1.sub0_sub1:areg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_128_sub2_sub3
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_128_sub2_sub3
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub2_sub3, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1.sub2_sub3:areg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_128_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_64_align2 = COPY killed [[DEF]].sub1_sub2
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_64_align2 = COPY killed %1.sub1_sub2:areg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_128_sub0_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_128_sub0_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub0_sub1_sub2, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_96_align2 = COPY killed %1.sub0_sub1_sub2:areg_128_align2
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_128_sub1_sub2_sub3
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_128_sub1_sub2_sub3
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_96_align2 = COPY killed [[DEF]].sub1_sub2_sub3
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_96_align2 = COPY killed %1.sub1_sub2_sub3:areg_128_align2
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_agpr_128
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: aligned_agpr_128
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128_align2 = IMPLICIT_DEF
+    %2:areg_128_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_agpr_128
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX90A-LABEL: name: unaligned_agpr_128
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:areg_128 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:areg_128_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:areg_128 = IMPLICIT_DEF
+    %2:areg_128_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
+...
diff --git a/llvm/test/CodeGen/AMDGPU/si-fold-aligned-vgprs.mir b/llvm/test/CodeGen/AMDGPU/si-fold-aligned-vgprs.mir
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/si-fold-aligned-vgprs.mir
@@ -0,0 +1,323 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX908 %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX90A %s
+
+# Should fold all copies for gfx908. In case of gfx90a, the folding must happen only for the aligned cases.
+
+---
+name:            aligned_vgpr_64
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_64
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_64_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_64
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_64_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_64_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_64
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_64
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_64
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_64_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_64 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_96_sub0_subg1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_96_sub0_subg1
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_96_sub0_subg1
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_96_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1.sub0_sub1:vreg_96_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_96_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_96_sub1_sub2
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub1_sub2, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_96_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_64_align2 = COPY killed [[DEF]].sub1_sub2
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_96_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1.sub1_sub2:vreg_96_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_96
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_96
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_96
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_96_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_96_align2 = IMPLICIT_DEF
+    %2:vreg_96_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_96
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_96
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_96 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_96
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_96 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_96_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_96 = IMPLICIT_DEF
+    %2:vreg_96_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_128_sub0_sub1
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_128_sub0_sub1
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_128_sub0_sub1
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub0_sub1, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1.sub0_sub1:vreg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_128_sub2_sub3
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_128_sub2_sub3
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub2_sub3, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_128_sub2_sub3
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub2_sub3, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1.sub2_sub3:vreg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_128_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_128_sub1_sub2
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[DEF]].sub1_sub2, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_128_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_64_align2 = COPY killed [[DEF]].sub1_sub2
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX2 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_64_align2 = COPY killed %1.sub1_sub2:vreg_128_align2
+    GLOBAL_STORE_DWORDX2 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_128_sub0_sub1_sub2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_128_sub0_sub1_sub2
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub0_sub1_sub2, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_128_sub0_sub1_sub2
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub0_sub1_sub2, 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_96_align2 = COPY killed %1.sub0_sub1_sub2:vreg_128_align2
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_128_sub1_sub2_sub3
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_128_sub1_sub2_sub3
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[DEF]].sub1_sub2_sub3, 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_128_sub1_sub2_sub3
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_96_align2 = COPY killed [[DEF]].sub1_sub2_sub3
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX3 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_96_align2 = COPY killed %1.sub1_sub2_sub3:vreg_128_align2
+    GLOBAL_STORE_DWORDX3 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            aligned_vgpr_128
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: aligned_vgpr_128
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: aligned_vgpr_128
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128_align2 = IMPLICIT_DEF
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128_align2 = IMPLICIT_DEF
+    %2:vreg_128_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
+...
+
+---
+name:            unaligned_vgpr_128
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; GFX908-LABEL: name: unaligned_vgpr_128
+    ; GFX908: liveins: $vgpr0_vgpr1
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX908-NEXT: [[DEF:%[0-9]+]]:vreg_128 = IMPLICIT_DEF
+    ; GFX908-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[DEF]], 0, 0, implicit $exec
+    ; GFX90A-LABEL: name: unaligned_vgpr_128
+    ; GFX90A: liveins: $vgpr0_vgpr1
+    ; GFX90A-NEXT: {{  $}}
+    ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vreg_64_align2 = COPY $vgpr0_vgpr1
+    ; GFX90A-NEXT: [[DEF:%[0-9]+]]:vreg_128 = IMPLICIT_DEF
+    ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vreg_128_align2 = COPY killed [[DEF]]
+    ; GFX90A-NEXT: GLOBAL_STORE_DWORDX4 [[COPY]], [[COPY1]], 0, 0, implicit $exec
+    %0:vreg_64_align2 = COPY $vgpr0_vgpr1
+    %1:vreg_128 = IMPLICIT_DEF
+    %2:vreg_128_align2 = COPY killed %1
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, implicit $exec
+...