[AMDGPU] Vectorize misaligned global loads & stores

[jrbyrnes]

Summary

Based on experimentation on gfx906,908,90a and 1030, wider global loads / stores are more performant than multiple narrower ones independent of alignment -- this is especially true when combining 8 bit loads / stores, in which case speedup was usually 2x across all alignments.

Change-Id: I1713c6edfc189052b8a71dc1135f9a436c1042e0

[jrbyrnes]

Edit comment.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

1551

Should also cover "other" address space. We have isFlatGlobalAddrSpace but probably need one that excludes flat

[rampitec]

LGTM, modulo Matt's comment.

[jrbyrnes]

Also include "other" address space in updated model.

[rampitec]

llvm/lib/Target/AMDGPU/AMDGPU.h

431 ↗

(On Diff #502013)

I do not like the name here. It suggests it is global address space only.

[jrbyrnes]

Naming

[rampitec]

LGTM, please give Matt a chance to review.

[jrbyrnes]

Thanks for review -- passed psdb.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (AddrSpace == AMDGPUAS::FLAT_ADDRESS &&
       !Subtarget->hasUnalignedScratchAccess()) {
     bool AlignedBy4 = Alignment >= Align(4);
     if (IsFast)
       *IsFast = AlignedBy4;
 
     return AlignedBy4;
   }
 
-  if (Subtarget->hasUnalignedBufferAccessEnabled()) {
-    // If we have a uniform constant load, it still requires using a slow
-    // buffer instruction if unaligned.
-    if (IsFast) {
-      // Accesses can really be issued as 1-byte aligned or 4-byte aligned, so
-      // 2-byte alignment is worse than 1 unless doing a 2-byte access.
-      *IsFast = (AddrSpace == AMDGPUAS::CONSTANT_ADDRESS ||
-                 AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT) ?
-        Alignment >= Align(4) : Alignment != Align(2);
-    }
+  // So long as they are legal, wide global memory operations perform better
+  // than multiple smaller memory ops -- even when misaligned
+  if (AddrSpace == AMDGPUAS::GLOBAL_ADDRESS ||
+      AddrSpace == AMDGPUAS::CONSTANT_ADDRESS ||
+      AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
+    if (IsFast)
+      *IsFast = Size;
 
-    return true;
+    return Alignment >= Align(4) ||
+           Subtarget->hasUnalignedBufferAccessEnabled();
   }
 
   // Smaller than dword value must be aligned.
   if (Size < 32)
     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.

llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll

 ; GFX7-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
 ; GFX7-ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; GFX7-ALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GFX7-ALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-UNALIGNED-LABEL: global_load_2xi16_align2:
 ; GFX7-UNALIGNED:       ; %bb.0:
 ; GFX7-UNALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-UNALIGNED-NEXT:    v_add_i32_e32 v2, vcc, 2, v0
-; GFX7-UNALIGNED-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
-; GFX7-UNALIGNED-NEXT:    flat_load_ushort v2, v[2:3]
-; GFX7-UNALIGNED-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX7-UNALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; GFX7-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
+; GFX7-UNALIGNED-NEXT:    flat_load_dword v0, v[0:1]
 ; GFX7-UNALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; GFX7-UNALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GFX7-UNALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: global_load_2xi16_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_ushort v2, v[0:1], off
-; GFX9-NEXT:    global_load_ushort v3, v[0:1], off offset:2
+; GFX9-NEXT:    global_load_dword v0, v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_load_2xi16_align2:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    s_clause 0x1
-; GFX10-NEXT:    global_load_ushort v2, v[0:1], off
-; GFX10-NEXT:    global_load_ushort v3, v[0:1], off offset:2
+; GFX10-NEXT:    global_load_dword v0, v[0:1], off
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: global_load_2xi16_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    global_load_u16 v2, v[0:1], off
-; GFX11-NEXT:    global_load_u16 v0, v[0:1], off offset:2
+; GFX11-NEXT:    global_load_b32 v0, v[0:1], off
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v2
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, ptr addrspace(1) %p, i64 1
   %p.0 = load i16, ptr addrspace(1) %p, align 2
   %p.1 = load i16, ptr addrspace(1) %gep.p, align 2
   %zext.0 = zext i16 %p.0 to i32
   %zext.1 = zext i16 %p.1 to i32
   %shl.1 = shl i32 %zext.1, 16
   %or = or i32 %zext.0, %shl.1

 ; GFX7-ALIGNED-NEXT:    v_mov_b32_e32 v2, 2
 ; GFX7-ALIGNED-NEXT:    v_mov_b32_e32 v1, s3
 ; GFX7-ALIGNED-NEXT:    flat_store_short v[0:1], v2
 ; GFX7-ALIGNED-NEXT:    s_endpgm
 ;
 ; GFX7-UNALIGNED-LABEL: global_store_2xi16_align2:
 ; GFX7-UNALIGNED:       ; %bb.0:
 ; GFX7-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
-; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v2, 1
+; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v2, 0x20001
 ; GFX7-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v0, s0
-; GFX7-UNALIGNED-NEXT:    s_add_u32 s2, s0, 2
 ; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v1, s1
-; GFX7-UNALIGNED-NEXT:    s_addc_u32 s3, s1, 0
-; GFX7-UNALIGNED-NEXT:    flat_store_short v[0:1], v2
-; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v0, s2
-; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v2, 2
-; GFX7-UNALIGNED-NEXT:    v_mov_b32_e32 v1, s3
-; GFX7-UNALIGNED-NEXT:    flat_store_short v[0:1], v2
+; GFX7-UNALIGNED-NEXT:    flat_store_dword v[0:1], v2
 ; GFX7-UNALIGNED-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: global_store_2xi16_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-NEXT:    v_mov_b32_e32 v1, 1
-; GFX9-NEXT:    v_mov_b32_e32 v2, 2
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x20001
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX9-NEXT:    global_store_short v0, v2, s[0:1] offset:2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_store_2xi16_align2:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10-NEXT:    v_mov_b32_e32 v1, 1
-; GFX10-NEXT:    v_mov_b32_e32 v2, 2
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0x20001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    global_store_short v0, v1, s[0:1]
-; GFX10-NEXT:    global_store_short v0, v2, s[0:1] offset:2
+; GFX10-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_store_2xi16_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 1
-; GFX11-NEXT:    v_mov_b32_e32 v2, 2
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x20001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    global_store_b16 v0, v1, s[0:1]
-; GFX11-NEXT:    global_store_b16 v0, v2, s[0:1] offset:2
+; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %gep.r = getelementptr i16, ptr addrspace(1) %r, i64 1
   store i16 1, ptr addrspace(1) %r, align 2
   store i16 2, ptr addrspace(1) %gep.r, align 2
   ret void
 }
 

llvm/test/CodeGen/AMDGPU/global-i16-load-store.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX1030 %s
+
+
+; Function Attrs: mustprogress nounwind willreturn
+define amdgpu_kernel void @half8(ptr addrspace(1) nocapture readonly %0, ptr addrspace(1) nocapture writeonly %1) local_unnamed_addr #0 {
+; GFX908-LABEL: half8:
+; GFX908:       ; %bb.0:
+; GFX908-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX908-NEXT:    v_mov_b32_e32 v4, 0
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    global_load_dwordx4 v[0:3], v4, s[0:1]
+; GFX908-NEXT:    s_waitcnt vmcnt(0)
+; GFX908-NEXT:    global_store_dwordx4 v4, v[0:3], s[2:3]
+; GFX908-NEXT:    s_endpgm
+;
+; GFX90A-LABEL: half8:
+; GFX90A:       ; %bb.0:
+; GFX90A-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX90A-NEXT:    v_mov_b32_e32 v4, 0
+; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT:    global_load_dwordx4 v[0:3], v4, s[0:1]
+; GFX90A-NEXT:    s_waitcnt vmcnt(0)
+; GFX90A-NEXT:    global_store_dwordx4 v4, v[0:3], s[2:3]
+; GFX90A-NEXT:    s_endpgm
+;
+; GFX1030-LABEL: half8:
+; GFX1030:       ; %bb.0:
+; GFX1030-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX1030-NEXT:    v_mov_b32_e32 v4, 0
+; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX1030-NEXT:    global_load_dwordx4 v[0:3], v4, s[0:1]
+; GFX1030-NEXT:    s_waitcnt vmcnt(0)
+; GFX1030-NEXT:    global_store_dwordx4 v4, v[0:3], s[2:3]
+; GFX1030-NEXT:    s_endpgm
+  %gep0 = getelementptr half, ptr addrspace(1) %0, i64 0
+  %gep1 = getelementptr half, ptr addrspace(1) %0, i64 1
+  %gep2 = getelementptr half, ptr addrspace(1) %0, i64 2
+  %gep3 = getelementptr half, ptr addrspace(1) %0, i64 3
+  %gep4 = getelementptr half, ptr addrspace(1) %0, i64 4
+  %gep5 = getelementptr half, ptr addrspace(1) %0, i64 5
+  %gep6 = getelementptr half, ptr addrspace(1) %0, i64 6
+  %gep7 = getelementptr half, ptr addrspace(1) %0, i64 7
+  %l0 = load half, ptr addrspace(1) %gep0, align 2
+  %l1 = load half, ptr addrspace(1) %gep1, align 2
+  %l2 = load half, ptr addrspace(1) %gep2, align 2
+  %l3 = load half, ptr addrspace(1) %gep3, align 2
+  %l4 = load half, ptr addrspace(1) %gep4, align 2
+  %l5 = load half, ptr addrspace(1) %gep5, align 2
+  %l6 = load half, ptr addrspace(1) %gep6, align 2
+  %l7 = load half, ptr addrspace(1) %gep7, align 2
+  %sgep0 = getelementptr half, ptr addrspace(1) %1, i64 0
+  %sgep1 = getelementptr half, ptr addrspace(1) %1, i64 1
+  %sgep2 = getelementptr half, ptr addrspace(1) %1, i64 2
+  %sgep3 = getelementptr half, ptr addrspace(1) %1, i64 3
+  %sgep4 = getelementptr half, ptr addrspace(1) %1, i64 4
+  %sgep5 = getelementptr half, ptr addrspace(1) %1, i64 5
+  %sgep6 = getelementptr half, ptr addrspace(1) %1, i64 6
+  %sgep7 = getelementptr half, ptr addrspace(1) %1, i64 7
+  store half %l0, ptr addrspace(1) %sgep0, align 2
+  store half %l1, ptr addrspace(1) %sgep1, align 2
+  store half %l2, ptr addrspace(1) %sgep2, align 2
+  store half %l3, ptr addrspace(1) %sgep3, align 2
+  store half %l4, ptr addrspace(1) %sgep4, align 2
+  store half %l5, ptr addrspace(1) %sgep5, align 2
+  store half %l6, ptr addrspace(1) %sgep6, align 2
+  store half %l7, ptr addrspace(1) %sgep7, align 2
+  ret void
+}
+
+; Function Attrs: mustprogress nounwind willreturn
+define amdgpu_kernel void @half6(ptr addrspace(1) nocapture readonly %0, ptr addrspace(1) nocapture writeonly %1) local_unnamed_addr #0 {
+; GFX908-LABEL: half6:
+; GFX908:       ; %bb.0:
+; GFX908-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX908-NEXT:    v_mov_b32_e32 v3, 0
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    global_load_dwordx3 v[0:2], v3, s[0:1]
+; GFX908-NEXT:    s_waitcnt vmcnt(0)
+; GFX908-NEXT:    global_store_dwordx3 v3, v[0:2], s[2:3]
+; GFX908-NEXT:    s_endpgm
+;
+; GFX90A-LABEL: half6:
+; GFX90A:       ; %bb.0:
+; GFX90A-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX90A-NEXT:    v_mov_b32_e32 v3, 0
+; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT:    global_load_dwordx3 v[0:2], v3, s[0:1]
+; GFX90A-NEXT:    s_waitcnt vmcnt(0)
+; GFX90A-NEXT:    global_store_dwordx3 v3, v[0:2], s[2:3]
+; GFX90A-NEXT:    s_endpgm
+;
+; GFX1030-LABEL: half6:
+; GFX1030:       ; %bb.0:
+; GFX1030-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX1030-NEXT:    v_mov_b32_e32 v3, 0
+; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX1030-NEXT:    global_load_dwordx3 v[0:2], v3, s[0:1]
+; GFX1030-NEXT:    s_waitcnt vmcnt(0)
+; GFX1030-NEXT:    global_store_dwordx3 v3, v[0:2], s[2:3]
+; GFX1030-NEXT:    s_endpgm
+  %gep0 = getelementptr half, ptr addrspace(1) %0, i64 0
+  %gep1 = getelementptr half, ptr addrspace(1) %0, i64 1
+  %gep2 = getelementptr half, ptr addrspace(1) %0, i64 2
+  %gep3 = getelementptr half, ptr addrspace(1) %0, i64 3
+  %gep4 = getelementptr half, ptr addrspace(1) %0, i64 4
+  %gep5 = getelementptr half, ptr addrspace(1) %0, i64 5
+  %l0 = load half, ptr addrspace(1) %gep0, align 1
+  %l1 = load half, ptr addrspace(1) %gep1, align 1
+  %l2 = load half, ptr addrspace(1) %gep2, align 1
+  %l3 = load half, ptr addrspace(1) %gep3, align 1
+  %l4 = load half, ptr addrspace(1) %gep4, align 1
+  %l5 = load half, ptr addrspace(1) %gep5, align 1
+  %sgep0 = getelementptr half, ptr addrspace(1) %1, i64 0
+  %sgep1 = getelementptr half, ptr addrspace(1) %1, i64 1
+  %sgep2 = getelementptr half, ptr addrspace(1) %1, i64 2
+  %sgep3 = getelementptr half, ptr addrspace(1) %1, i64 3
+  %sgep4 = getelementptr half, ptr addrspace(1) %1, i64 4
+  %sgep5 = getelementptr half, ptr addrspace(1) %1, i64 5
+  store half %l0, ptr addrspace(1) %sgep0, align 1
+  store half %l1, ptr addrspace(1) %sgep1, align 1
+  store half %l2, ptr addrspace(1) %sgep2, align 1
+  store half %l3, ptr addrspace(1) %sgep3, align 1
+  store half %l4, ptr addrspace(1) %sgep4, align 1
+  store half %l5, ptr addrspace(1) %sgep5, align 1
+  ret void
+}
+
+; Function Attrs: mustprogress nounwind willreturn
+define amdgpu_kernel void @half4(ptr addrspace(1) nocapture readonly %0, ptr addrspace(1) nocapture writeonly %1) local_unnamed_addr #0 {
+; GFX908-LABEL: half4:
+; GFX908:       ; %bb.0:
+; GFX908-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX908-NEXT:    v_mov_b32_e32 v2, 0
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    v_mov_b32_e32 v0, s0
+; GFX908-NEXT:    v_mov_b32_e32 v1, s1
+; GFX908-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX908-NEXT:    s_endpgm
+;
+; GFX90A-LABEL: half4:
+; GFX90A:       ; %bb.0:
+; GFX90A-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX90A-NEXT:    v_mov_b32_e32 v2, 0
+; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT:    v_pk_mov_b32 v[0:1], s[0:1], s[0:1] op_sel:[0,1]
+; GFX90A-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX90A-NEXT:    s_endpgm
+;
+; GFX1030-LABEL: half4:
+; GFX1030:       ; %bb.0:
+; GFX1030-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX1030-NEXT:    v_mov_b32_e32 v2, 0
+; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX1030-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX1030-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1030-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1030-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
+; GFX1030-NEXT:    s_endpgm
+  %gep0 = getelementptr half, ptr addrspace(1) %0, i64 0
+  %gep1 = getelementptr half, ptr addrspace(1) %0, i64 1
+  %gep2 = getelementptr half, ptr addrspace(1) %0, i64 2
+  %gep3 = getelementptr half, ptr addrspace(1) %0, i64 3
+  %l0 = load half, ptr addrspace(1) %gep0, align 4
+  %l1 = load half, ptr addrspace(1) %gep1, align 4
+  %l2 = load half, ptr addrspace(1) %gep2, align 4
+  %l3 = load half, ptr addrspace(1) %gep3, align 4
+  %sgep0 = getelementptr half, ptr addrspace(1) %1, i64 0
+  %sgep1 = getelementptr half, ptr addrspace(1) %1, i64 1
+  %sgep2 = getelementptr half, ptr addrspace(1) %1, i64 2
+  %sgep3 = getelementptr half, ptr addrspace(1) %1, i64 3
+  store half %l0, ptr addrspace(1) %sgep0, align 4
+  store half %l1, ptr addrspace(1) %sgep1, align 4
+  store half %l2, ptr addrspace(1) %sgep2, align 4
+  store half %l3, ptr addrspace(1) %sgep3, align 4
+  ret void
+}
+
+
+; Function Attrs: mustprogress nounwind willreturn
+define amdgpu_kernel void @half2(ptr addrspace(1) nocapture readonly %0, ptr addrspace(1) nocapture writeonly %1) local_unnamed_addr #0 {
+; GFX908-LABEL: half2:
+; GFX908:       ; %bb.0:
+; GFX908-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX908-NEXT:    v_mov_b32_e32 v0, 0
+; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX908-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX908-NEXT:    s_waitcnt vmcnt(0)
+; GFX908-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX908-NEXT:    s_endpgm
+;
+; GFX90A-LABEL: half2:
+; GFX90A:       ; %bb.0:
+; GFX90A-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX90A-NEXT:    v_mov_b32_e32 v0, 0
+; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX90A-NEXT:    s_waitcnt vmcnt(0)
+; GFX90A-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX90A-NEXT:    s_endpgm
+;
+; GFX1030-LABEL: half2:
+; GFX1030:       ; %bb.0:
+; GFX1030-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX1030-NEXT:    v_mov_b32_e32 v0, 0
+; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX1030-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX1030-NEXT:    s_waitcnt vmcnt(0)
+; GFX1030-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX1030-NEXT:    s_endpgm
+  %gep0 = getelementptr half, ptr addrspace(1) %0, i64 0
+  %gep1 = getelementptr half, ptr addrspace(1) %0, i64 1
+  %l0 = load half, ptr addrspace(1) %gep0
+  %l1 = load half, ptr addrspace(1) %gep1
+  %sgep0 = getelementptr half, ptr addrspace(1) %1, i64 0
+  %sgep1 = getelementptr half, ptr addrspace(1) %1, i64 1
+  store half %l0, ptr addrspace(1) %sgep0
+  store half %l1, ptr addrspace(1) %sgep1
+  ret void
+}
+
+

llvm/test/CodeGen/AMDGPU/load-constant-i16.ll

 ; GCN-NOHSA-SI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[0:3], 0
 ; GCN-NOHSA-SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 ; GCN-NOHSA-SI-NEXT:    s_endpgm
 ;
 ; GCN-HSA-LABEL: constant_load_v16i16_align2:
 ; GCN-HSA:       ; %bb.0: ; %entry
 ; GCN-HSA-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; GCN-HSA-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-HSA-NEXT:    s_add_u32 s2, s0, 16
-; GCN-HSA-NEXT:    s_addc_u32 s3, s1, 0
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v0, s0
-; GCN-HSA-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v1, s1
-; GCN-HSA-NEXT:    v_mov_b32_e32 v4, s2
+; GCN-HSA-NEXT:    s_add_u32 s0, s0, 16
+; GCN-HSA-NEXT:    s_addc_u32 s1, s1, 0
+; GCN-HSA-NEXT:    v_mov_b32_e32 v5, s1
+; GCN-HSA-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN-HSA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]
 ; GCN-HSA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]
 ; GCN-HSA-NEXT:    s_waitcnt vmcnt(1)
 ; GCN-HSA-NEXT:    flat_store_dwordx4 v[0:1], v[0:3]
 ; GCN-HSA-NEXT:    s_waitcnt vmcnt(1)
 ; GCN-HSA-NEXT:    flat_store_dwordx4 v[0:1], v[4:7]
 ; GCN-HSA-NEXT:    s_endpgm
 ;

llvm/test/CodeGen/AMDGPU/load-global-i16.ll

 ; GCN-NOHSA-SI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[8:11], 0 offset:16
 ; GCN-NOHSA-SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[8:11], 0
 ; GCN-NOHSA-SI-NEXT:    s_endpgm
 ;
 ; GCN-HSA-LABEL: global_load_v16i16_align2:
 ; GCN-HSA:       ; %bb.0: ; %entry
 ; GCN-HSA-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GCN-HSA-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-HSA-NEXT:    s_add_u32 s4, s2, 16
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v5, s1
+; GCN-HSA-NEXT:    s_addc_u32 s5, s3, 0
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN-HSA-NEXT:    s_add_u32 s0, s0, 16
 ; GCN-HSA-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN-HSA-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]
 ; GCN-HSA-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]
-; GCN-HSA-NEXT:    s_add_u32 s0, s2, 16
-; GCN-HSA-NEXT:    s_addc_u32 s1, s3, 0
-; GCN-HSA-NEXT:    v_mov_b32_e32 v11, s1
+; GCN-HSA-NEXT:    v_mov_b32_e32 v11, s5
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v9, s3
-; GCN-HSA-NEXT:    v_mov_b32_e32 v10, s0
+; GCN-HSA-NEXT:    v_mov_b32_e32 v10, s4
 ; GCN-HSA-NEXT:    v_mov_b32_e32 v8, s2
 ; GCN-HSA-NEXT:    s_waitcnt vmcnt(1)
 ; GCN-HSA-NEXT:    flat_store_dwordx4 v[10:11], v[0:3]
 ; GCN-HSA-NEXT:    s_waitcnt vmcnt(1)
 ; GCN-HSA-NEXT:    flat_store_dwordx4 v[8:9], v[4:7]
 ; GCN-HSA-NEXT:    s_endpgm
 ;
 ; GCN-NOHSA-VI-LABEL: global_load_v16i16_align2:

llvm/test/CodeGen/AMDGPU/udiv.ll

 ; VI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GCN-LABEL: v_udiv_i16:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_add_u32 s4, s2, 2
-; GCN-NEXT:    s_addc_u32 s5, s3, 0
-; GCN-NEXT:    v_mov_b32_e32 v0, s4
-; GCN-NEXT:    v_mov_b32_e32 v1, s5
-; GCN-NEXT:    flat_load_ushort v2, v[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN-NEXT:    v_mov_b32_e32 v1, s3
-; GCN-NEXT:    flat_load_ushort v0, v[0:1]
+; GCN-NEXT:    flat_load_dword v0, v[0:1]
 ; GCN-NEXT:    v_mov_b32_e32 v1, s1
-; GCN-NEXT:    s_waitcnt vmcnt(1)
-; GCN-NEXT:    v_cvt_f32_u32_e32 v2, v2
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_cvt_f32_u32_e32 v3, v0
-; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v2
+; GCN-NEXT:    v_cvt_f32_u32_sdwa v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GCN-NEXT:    v_cvt_f32_u32_sdwa v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
 ; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    v_rcp_iflag_f32_e32 v4, v2
 ; GCN-NEXT:    v_mul_f32_e32 v4, v3, v4
 ; GCN-NEXT:    v_trunc_f32_e32 v4, v4
 ; GCN-NEXT:    v_cvt_u32_f32_e32 v5, v4
 ; GCN-NEXT:    v_mad_f32 v3, -v4, v2, v3
 ; GCN-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, v2
 ; GCN-NEXT:    v_addc_u32_e32 v2, vcc, 0, v5, vcc
 ; GCN-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GCN-NEXT:    flat_store_dword v[0:1], v2
 ; GCN-NEXT:    s_endpgm
 ;
 ; GFX1030-LABEL: v_udiv_i16:
 ; GFX1030:       ; %bb.0:
 ; GFX1030-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX1030-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX1030-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX1030-NEXT:    s_clause 0x1
-; GFX1030-NEXT:    global_load_ushort v1, v0, s[2:3] offset:2
-; GFX1030-NEXT:    global_load_ushort v2, v0, s[2:3]
-; GFX1030-NEXT:    s_waitcnt vmcnt(1)
-; GFX1030-NEXT:    v_cvt_f32_u32_e32 v1, v1
+; GFX1030-NEXT:    global_load_dword v1, v0, s[2:3]
 ; GFX1030-NEXT:    s_waitcnt vmcnt(0)
-; GFX1030-NEXT:    v_cvt_f32_u32_e32 v2, v2
-; GFX1030-NEXT:    v_rcp_iflag_f32_e32 v3, v1
-; GFX1030-NEXT:    v_mul_f32_e32 v3, v2, v3
+; GFX1030-NEXT:    v_cvt_f32_u32_sdwa v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GFX1030-NEXT:    v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX1030-NEXT:    v_rcp_iflag_f32_e32 v3, v2
+; GFX1030-NEXT:    v_mul_f32_e32 v3, v1, v3
 ; GFX1030-NEXT:    v_trunc_f32_e32 v3, v3
-; GFX1030-NEXT:    v_fma_f32 v2, -v3, v1, v2
+; GFX1030-NEXT:    v_fma_f32 v1, -v3, v2, v1
 ; GFX1030-NEXT:    v_cvt_u32_f32_e32 v3, v3
-; GFX1030-NEXT:    v_cmp_ge_f32_e64 vcc_lo, |v2|, v1
+; GFX1030-NEXT:    v_cmp_ge_f32_e64 vcc_lo, |v1|, v2
 ; GFX1030-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v3, vcc_lo
 ; GFX1030-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX1030-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX1030-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_udiv_i16:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]

llvm/test/Transforms/LoadStoreVectorizer/AMDGPU/merge-stores.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -mtriple=amdgcn-amd-amdhsa --mcpu=hawaii -passes=load-store-vectorizer -S -o - %s | FileCheck %s
 ; Copy of test/CodeGen/AMDGPU/merge-stores.ll with some additions
 
 target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
 
 ; TODO: Vector element tests
 ; TODO: Non-zero base offset for load and store combinations
 ; TODO: Same base addrspacecasted
;
 
   store i16 0, ptr addrspace(1) %out.gep.1
   store i16 0, ptr addrspace(1) %out, align 4
   ret void
 }
 
 define amdgpu_kernel void @merge_global_store_2_constants_i16_natural_align(ptr addrspace(1) %out) #0 {
 ; CHECK-LABEL: @merge_global_store_2_constants_i16_natural_align(
-; CHECK-NEXT:    [[OUT_GEP_1:%.*]] = getelementptr i16, ptr addrspace(1) [[OUT:%.*]], i32 1
-; CHECK-NEXT:    store i16 123, ptr addrspace(1) [[OUT_GEP_1]], align 2
-; CHECK-NEXT:    store i16 456, ptr addrspace(1) [[OUT]], align 2
+; CHECK-NEXT:    store <2 x i16> <i16 456, i16 123>, ptr addrspace(1) [[OUT:%.*]], align 2
 ; CHECK-NEXT:    ret void
 ;
   %out.gep.1 = getelementptr i16, ptr addrspace(1) %out, i32 1
 
   store i16 123, ptr addrspace(1) %out.gep.1
   store i16 456, ptr addrspace(1) %out
   ret void
 }
 
 define amdgpu_kernel void @merge_global_store_2_constants_i16_align_1(ptr addrspace(1) %out) #0 {
 ; CHECK-LABEL: @merge_global_store_2_constants_i16_align_1(
 ; CHECK-NEXT:    store <2 x i16> <i16 456, i16 123>, ptr addrspace(1) [[OUT:%.*]], align 1
 ; CHECK-NEXT:    ret void
 ;
   %out.gep.1 = getelementptr i16, ptr addrspace(1) %out, i32 1
 
   store i16 123, ptr addrspace(1) %out.gep.1, align 1
   store i16 456, ptr addrspace(1) %out, align 1
   ret void
 }
 
 define amdgpu_kernel void @merge_global_store_2_constants_half_natural_align(ptr addrspace(1) %out) #0 {
 ; CHECK-LABEL: @merge_global_store_2_constants_half_natural_align(
-; CHECK-NEXT:    [[OUT_GEP_1:%.*]] = getelementptr half, ptr addrspace(1) [[OUT:%.*]], i32 1
-; CHECK-NEXT:    store half 0xH4000, ptr addrspace(1) [[OUT_GEP_1]], align 2
-; CHECK-NEXT:    store half 0xH3C00, ptr addrspace(1) [[OUT]], align 2
+; CHECK-NEXT:    store <2 x half> <half 0xH3C00, half 0xH4000>, ptr addrspace(1) [[OUT:%.*]], align 2
 ; CHECK-NEXT:    ret void
 ;
   %out.gep.1 = getelementptr half, ptr addrspace(1) %out, i32 1
 
   store half 2.0, ptr addrspace(1) %out.gep.1
   store half 1.0, ptr addrspace(1) %out
   ret void
 }