[AMDGPU] Select build_vector DAG nodes according to the divergence

[alex-t]

Summary

This change enables divergence-driven instruction selection for the build_vector DAG nodes.
It also enables packed i16 instructions for GFX9.

[rampitec]

LGTM

[foad]

llvm/lib/Target/AMDGPU/SIInstructions.td

2463

Does this pattern need to be marked as Uniform?

2468

Does this pattern need to be marked as Divergent?

llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll

2

This file should not be executable.

llvm/lib/Target/AMDGPU/SIInstructions.td

(fma (f64 (VOP3NoMods f64:$src0)),
        (f64 (VOP3NoMods f64:$src2))),
   (V_FMAC_F64_e64 SRCMODS.NONE, $src0, SRCMODS.NONE, $src1,
                   SRCMODS.NONE, $src2)
 >;
 
 // COPY is workaround tablegen bug from multiple outputs
 // from S_LSHL_B32's multiple outputs from implicit scc def.
 def : GCNPat <
-  (v2i16 (build_vector (i16 0), (i16 SReg_32:$src1))),
+  (v2i16 (UniformBinFrag<build_vector> (i16 0), (i16 SReg_32:$src1))),
   (S_LSHL_B32 SReg_32:$src1, (i16 16))
 >;
 
 def : GCNPat <
-  (v2i16 (build_vector (i16 SReg_32:$src1), (i16 0))),
+  (v2i16 (DivergentBinFrag<build_vector> (i16 0), (i16 SReg_32:$src1))),
+  (v2i16 (V_LSHLREV_B32_e64 (i16 16), SReg_32:$src1))
+>;
+
+
+def : GCNPat <
+  (v2i16 (UniformBinFrag<build_vector> (i16 SReg_32:$src1), (i16 0))),
   (S_AND_B32 (S_MOV_B32 (i32 0xffff)), SReg_32:$src1)
 >;
 
 def : GCNPat <
-  (v2f16 (build_vector (f16 SReg_32:$src1), (f16 FP_ZERO))),
+  (v2i16 (DivergentBinFrag<build_vector> (i16 SReg_32:$src1), (i16 0))),
+  (v2i16 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), SReg_32:$src1))
+>;
+
+def : GCNPat <
+  (v2f16 (UniformBinFrag<build_vector> (f16 SReg_32:$src1), (f16 FP_ZERO))),
   (S_AND_B32 (S_MOV_B32 (i32 0xffff)), SReg_32:$src1)
 >;
 
 def : GCNPat <
+  (v2f16 (DivergentBinFrag<build_vector> (f16 SReg_32:$src1), (f16 FP_ZERO))),
+  (v2f16 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), SReg_32:$src1))
+>;
+
+def : GCNPat <
   (v2i16 (build_vector (i16 SReg_32:$src0), (i16 undef))),
   (COPY_TO_REGCLASS SReg_32:$src0, SReg_32)
 >;
 
 def : GCNPat <
   (v2i16 (build_vector (i16 VGPR_32:$src0), (i16 undef))),
   (COPY_TO_REGCLASS VGPR_32:$src0, VGPR_32)
 >;
 
 def : GCNPat <
   (v2f16 (build_vector f16:$src0, (f16 undef))),
   (COPY $src0)
 >;
 
 def : GCNPat <
-  (v2i16 (build_vector (i16 undef), (i16 SReg_32:$src1))),
+  (v2i16 (UniformBinFrag<build_vector> (i16 undef), (i16 SReg_32:$src1))),
   (S_LSHL_B32 SReg_32:$src1, (i32 16))
 >;
 
 def : GCNPat <
-  (v2f16 (build_vector (f16 undef), (f16 SReg_32:$src1))),
+  (v2i16 (DivergentBinFrag<build_vector> (i16 undef), (i16 SReg_32:$src1))),
+  (v2i16 (V_LSHLREV_B32_e64 (i32 16), SReg_32:$src1))
+>;
+
+
+def : GCNPat <
+  (v2f16 (UniformBinFrag<build_vector> (f16 undef), (f16 SReg_32:$src1))),
   (S_LSHL_B32 SReg_32:$src1, (i32 16))
 >;
 
+def : GCNPat <
+  (v2f16 (DivergentBinFrag<build_vector> (f16 undef), (f16 SReg_32:$src1))),
+  (v2f16 (V_LSHLREV_B32_e64 (i32 16), SReg_32:$src1))
+>;
+
 let SubtargetPredicate = HasVOP3PInsts in {
 def : GCNPat <
-  (v2i16 (build_vector (i16 SReg_32:$src0), (i16 SReg_32:$src1))),
+  (v2i16 (UniformBinFrag<build_vector> (i16 SReg_32:$src0), (i16 SReg_32:$src1))),
   (S_PACK_LL_B32_B16 SReg_32:$src0, SReg_32:$src1)
 >;
 
+def : GCNPat <
+  (v2i16 (DivergentBinFrag<build_vector> (i16 SReg_32:$src0), (i16 SReg_32:$src1))),
+  (v2i16 (V_LSHL_OR_B32_e64 $src1, (i32 16), (i32 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), $src0))))
+>;
+
 // With multiple uses of the shift, this will duplicate the shift and
 // increase register pressure.
 def : GCNPat <
-  (v2i16 (build_vector (i16 SReg_32:$src0), (i16 (trunc (srl_oneuse SReg_32:$src1, (i32 16)))))),
+  (v2i16 (UniformBinFrag<build_vector> (i16 SReg_32:$src0), (i16 (trunc (srl_oneuse SReg_32:$src1, (i32 16)))))),
   (v2i16 (S_PACK_LH_B32_B16 SReg_32:$src0, SReg_32:$src1))
 >;
 
+def : GCNPat <
+  (v2i16 (DivergentBinFrag<build_vector> (i16 SReg_32:$src0), (i16 (trunc (srl_oneuse SReg_32:$src1, (i32 16)))))),
+  (v2i16 (V_BFI_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), SReg_32:$src0, SReg_32:$src1))
+>;
+
 
 def : GCNPat <
-  (v2i16 (build_vector (i16 (trunc (srl_oneuse SReg_32:$src0, (i32 16)))),
+  (v2i16 (UniformBinFrag<build_vector> (i16 (trunc (srl_oneuse SReg_32:$src0, (i32 16)))),
                        (i16 (trunc (srl_oneuse SReg_32:$src1, (i32 16)))))),
   (S_PACK_HH_B32_B16 SReg_32:$src0, SReg_32:$src1)
 >;
 
+def : GCNPat <
+  (v2i16 (DivergentBinFrag<build_vector> (i16 (trunc (srl_oneuse SReg_32:$src0, (i32 16)))),
+                       (i16 (trunc (srl_oneuse SReg_32:$src1, (i32 16)))))),
+  (v2i16 (V_AND_OR_B32_e64 SReg_32:$src1, (i32 (V_MOV_B32_e32 (i32 0xffff0000))), (i32 (V_LSHRREV_B32_e64 (i32 16), SReg_32:$src0))))
+>;
+
 // TODO: Should source modifiers be matched to v_pack_b32_f16?
 def : GCNPat <
-  (v2f16 (build_vector (f16 SReg_32:$src0), (f16 SReg_32:$src1))),
+  (v2f16 (UniformBinFrag<build_vector> (f16 SReg_32:$src0), (f16 SReg_32:$src1))),
   (S_PACK_LL_B32_B16 SReg_32:$src0, SReg_32:$src1)
 >;
 
 def : GCNPat <
+  (v2f16 (DivergentBinFrag<build_vector> (f16 SReg_32:$src0), (f16 SReg_32:$src1))),
+  (v2f16 (V_LSHL_OR_B32_e64 SReg_32:$src1, (i32 16), (i32 (V_AND_B32_e64 (i32 (V_MOV_B32_e32 (i32 0xffff))), SReg_32:$src0))))
+>;
+
+
+def : GCNPat <
   (v2f16 (is_canonicalized<build_vector> (f16 (VOP3Mods (f16 VGPR_32:$src0), i32:$src0_mods)),
                                          (f16 (VOP3Mods (f16 VGPR_32:$src1), i32:$src1_mods)))),
   (V_PACK_B32_F16_e64 $src0_mods, VGPR_32:$src0, $src1_mods, VGPR_32:$src1)
 >;
 } // End SubtargetPredicate = HasVOP3PInsts
 
 def : GCNPat <
   (v2f16 (scalar_to_vector f16:$src0)),

llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll

+; RUN: llc -march=amdgcn -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX9 %s
+
+; GCN-LABEL: name:            uniform_vec_0_i16
+; GCN: S_LSHL_B32
+define amdgpu_kernel void @uniform_vec_0_i16(i32 addrspace(1)* %out, i16 %a) {
+  %tmp = insertelement <2 x i16> undef, i16 0, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %a, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_0_i16
+; GCN: V_LSHLREV_B32_e64
+define i32 @divergent_vec_0_i16(i16 %a) {
+  %tmp = insertelement <2 x i16> undef, i16 0, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %a, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  ret i32 %val
+}
+
+; GCN-LABEL: name:            uniform_vec_i16_0
+; GCN: S_AND_B32
+define amdgpu_kernel void @uniform_vec_i16_0(i32 addrspace(1)* %out, i16 %a) {
+  %tmp = insertelement <2 x i16> undef, i16 %a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 0, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_i16_0
+; GCN: V_AND_B32_e64
+define i32 @divergent_vec_i16_0(i16 %a) {
+  %tmp = insertelement <2 x i16> undef, i16 %a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 0, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  ret i32 %val
+}
+
+; GCN-LABEL: name:            uniform_vec_f16_0
+; GCN: S_AND_B32
+define amdgpu_kernel void @uniform_vec_f16_0(float addrspace(1)* %out, half %a) {
+  %tmp = insertelement <2 x half> undef, half %a, i32 0
+  %vec = insertelement <2 x half> %tmp, half 0.0, i32 1
+  %val = bitcast <2 x half> %vec to float
+  store float %val, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_f16_0
+; GCN: V_CVT_F16_F32_e64 0, %0
+; GCN: COPY %1
+
+; GFX9-LABEL: name:            divergent_vec_f16_0
+; GFX9: V_AND_B32_e64
+define float @divergent_vec_f16_0(half %a) {
+  %tmp = insertelement <2 x half> undef, half %a, i32 0
+  %vec = insertelement <2 x half> %tmp, half 0.0, i32 1
+  %val = bitcast <2 x half> %vec to float
+  ret float %val
+}
+
+; GCN-LABEL: name:            uniform_vec_i16_LL
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 65535
+; GCN: %[[AND:[0-9]+]]:sreg_32 = S_AND_B32 killed %{{[0-9]+}}, killed %[[IMM]]
+; GCN: %[[SHIFT:[0-9]+]]:sreg_32 = S_MOV_B32 16
+; GCN:  %[[SHL:[0-9]+]]:sreg_32 = S_LSHL_B32 killed %{{[0-9]+}}, killed %[[SHIFT]]
+; GCN: S_OR_B32 killed %[[AND]], killed %[[SHL]]
+
+; GFX9-LABEL: name:            uniform_vec_i16_LL
+; GFX9: S_PACK_LL_B32_B16
+define amdgpu_kernel void @uniform_vec_i16_LL(i32 addrspace(4)* %in0, i32 addrspace(4)* %in1) {
+  %val0 = load volatile i32, i32 addrspace(4)* %in0
+  %val1 = load volatile i32, i32 addrspace(4)* %in1
+  %lo = trunc i32 %val0 to i16
+  %hi = trunc i32 %val1 to i16
+  %vec.0 = insertelement <2 x i16> undef, i16 %lo, i32 0
+  %vec.1 = insertelement <2 x i16> %vec.0, i16 %hi, i32 1
+  %vec.i32 = bitcast <2 x i16> %vec.1 to i32
+  call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_i16_LL
+; GCN: %[[SHIFT:[0-9]+]]:sreg_32 = S_MOV_B32 16
+; GCN: %[[SHL:[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 killed %[[SHIFT]], %1, implicit $exec
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 65535
+; GCN: %[[AND:[0-9]+]]:vgpr_32 = V_AND_B32_e64 %0, killed %[[IMM]], implicit $exec
+; GCN: V_OR_B32_e64 killed %[[AND]], killed %[[SHL]], implicit $exec
+
+; GFX9-LABEL: name:            divergent_vec_i16_LL
+; GFX9: %[[IMM:[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65535
+; GFX9: %[[AND:[0-9]+]]:vgpr_32 = V_AND_B32_e64 killed %[[IMM]]
+; GFX9: V_LSHL_OR_B32_e64 %{{[0-9]+}}, 16, killed %[[AND]]
+define i32 @divergent_vec_i16_LL(i16 %a, i16 %b) {
+  %tmp = insertelement <2 x i16> undef, i16 %a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %b, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  ret i32 %val
+}
+
+; GCN-LABEL: name:            uniform_vec_i16_LH
+; GCN-DAG: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 65535
+; GCN-DAG: %[[AND:[0-9]+]]:sreg_32 = S_AND_B32 killed %{{[0-9]+}}, killed %[[IMM]]
+; GCN-DAG: %[[NEG:[0-9]+]]:sreg_32 = S_MOV_B32 -65536
+; GCN-DAG: %[[ANDN:[0-9]+]]:sreg_32 = S_AND_B32 killed %{{[0-9]+}}, killed %[[NEG]]
+; GCN: S_OR_B32 killed %[[AND]], killed %[[ANDN]]
+
+; GFX9-LABEL: name:            uniform_vec_i16_LH
+; GFX9: S_PACK_LH_B32_B16
+define amdgpu_kernel void @uniform_vec_i16_LH(i32 addrspace(1)* %out, i16 %a, i32 %b) {
+  %shift = lshr i32 %b, 16
+  %tr = trunc i32 %shift to i16
+  %tmp = insertelement <2 x i16> undef, i16 %a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %tr, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_i16_LH
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 65535
+; GCN: V_BFI_B32_e64 killed %[[IMM]]
+define i32 @divergent_vec_i16_LH(i16 %a, i32 %b) {
+  %shift = lshr i32 %b, 16
+  %tr = trunc i32 %shift to i16
+  %tmp = insertelement <2 x i16> undef, i16 %a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %tr, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  ret i32 %val
+}
+
+; GCN-LABEL: name:            uniform_vec_i16_HH
+; GCN: %[[SHIFT:[0-9]+]]:sreg_32 = S_MOV_B32 16
+; GCN:  %[[SHR:[0-9]+]]:sreg_32 = S_LSHR_B32 killed %{{[0-9]+}}, killed %[[SHIFT]]
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 -65536
+; GCN: %[[AND:[0-9]+]]:sreg_32 = S_AND_B32 killed %{{[0-9]+}}, killed %[[IMM]]
+; GCN: S_OR_B32 killed %[[SHR]], killed %[[AND]]
+
+; GFX9-LABEL: name:            uniform_vec_i16_HH
+; GFX9: S_PACK_HH_B32_B16
+define amdgpu_kernel void @uniform_vec_i16_HH(i32 addrspace(1)* %out, i32 %a, i32 %b) {
+  %shift_a = lshr i32 %a, 16
+  %tr_a = trunc i32 %shift_a to i16
+  %shift_b = lshr i32 %b, 16
+  %tr_b = trunc i32 %shift_b to i16
+  %tmp = insertelement <2 x i16> undef, i16 %tr_a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %tr_b, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_i16_HH
+; GCN: %[[SHR:[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 killed %{{[0-9]+}}, %0, implicit $exec
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 -65536
+; GCN: %[[AND:[0-9]+]]:vgpr_32 = V_AND_B32_e64 %1, killed %[[IMM]], implicit $exec
+; GCN: V_OR_B32_e64 killed %[[SHR]], killed %[[AND]], implicit $exec
+
+; GFX9-LABEL: name:            divergent_vec_i16_HH
+; GFX9: %[[SHR:[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, %0
+; GFX9: %[[IMM:[0-9]+]]:vgpr_32 = V_MOV_B32_e32 -65536, implicit $exec
+; GFX9: V_AND_OR_B32_e64 %1, killed %[[IMM]], killed %[[SHR]]
+define i32 @divergent_vec_i16_HH(i32 %a, i32 %b) {
+  %shift_a = lshr i32 %a, 16
+  %tr_a = trunc i32 %shift_a to i16
+  %shift_b = lshr i32 %b, 16
+  %tr_b = trunc i32 %shift_b to i16
+  %tmp = insertelement <2 x i16> undef, i16 %tr_a, i32 0
+  %vec = insertelement <2 x i16> %tmp, i16 %tr_b, i32 1
+  %val = bitcast <2 x i16> %vec to i32
+  ret i32 %val
+}
+
+; GCN-LABEL: name:            uniform_vec_f16_LL
+; GCN: %[[IMM:[0-9]+]]:sreg_32 = S_MOV_B32 65535
+; GCN: %[[AND:[0-9]+]]:sreg_32 = S_AND_B32 killed %{{[0-9]+}}, killed %[[IMM]]
+; GCN: %[[SHIFT:[0-9]+]]:sreg_32 = S_MOV_B32 16
+; GCN:  %[[SHL:[0-9]+]]:sreg_32 = S_LSHL_B32 killed %{{[0-9]+}}, killed %[[SHIFT]]
+; GCN: S_OR_B32 killed %[[AND]], killed %[[SHL]]
+
+; GFX9-LABEL: name:            uniform_vec_f16_LL
+; GFX9: S_PACK_LL_B32_B16
+define amdgpu_kernel void @uniform_vec_f16_LL(i32 addrspace(4)* %in0, i32 addrspace(4)* %in1) {
+  %val0 = load volatile i32, i32 addrspace(4)* %in0
+  %val1 = load volatile i32, i32 addrspace(4)* %in1
+  %lo.i = trunc i32 %val0 to i16
+  %hi.i = trunc i32 %val1 to i16
+  %lo = bitcast i16 %lo.i to half
+  %hi = bitcast i16 %hi.i to half
+  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
+  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
+  %vec.i32 = bitcast <2 x half> %vec.1 to i32
+
+  call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0
+  ret void
+}
+
+; GCN-LABEL: name:            divergent_vec_f16_LL
+; GCN: %[[SHIFT:[0-9]+]]:sreg_32 = S_MOV_B32 16
+; GCN: %[[SHL:[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 killed %[[SHIFT]]
+; GCN: V_OR_B32_e64 killed %{{[0-9]+}}, killed %[[SHL]], implicit $exec
+
+; GFX9-LABEL: name:            divergent_vec_f16_LL
+; GFX9: %[[IMM:[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65535
+; GFX9: %[[AND:[0-9]+]]:vgpr_32 = V_AND_B32_e64 killed %[[IMM]]
+; GFX9: V_LSHL_OR_B32_e64 %{{[0-9]+}}, 16, killed %[[AND]]
+define float @divergent_vec_f16_LL(half %a, half %b) {
+  %tmp = insertelement <2 x half> undef, half %a, i32 0
+  %vec = insertelement <2 x half> %tmp, half %b, i32 1
+  %val = bitcast <2 x half> %vec to float
+  ret float %val
+}

llvm/test/CodeGen/AMDGPU/idot8s.ll

 ; GFX9-NEXT:    v_mov_b32_e32 v4, 0xffff
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:    global_load_ushort v3, v0, s[2:3]
 ; GFX9-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_and_b32_e32 v11, 15, v1
+; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 28, v1
+; GFX9-NEXT:    v_bfe_u32 v6, v1, 24, 4
+; GFX9-NEXT:    v_bfe_u32 v7, v1, 20, 4
+; GFX9-NEXT:    v_bfe_u32 v8, v1, 16, 4
+; GFX9-NEXT:    v_bfe_u32 v9, v1, 12, 4
+; GFX9-NEXT:    v_bfe_u32 v10, v1, 8, 4
+; GFX9-NEXT:    v_bfe_u32 v11, v1, 4, 4
+; GFX9-NEXT:    v_and_b32_e32 v1, 15, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_and_b32_e32 v18, 15, v2
-; GFX9-NEXT:    v_bfe_u32 v5, v1, 24, 4
-; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 28, v1
-; GFX9-NEXT:    v_bfe_u32 v7, v1, 16, 4
-; GFX9-NEXT:    v_bfe_u32 v8, v1, 20, 4
-; GFX9-NEXT:    v_bfe_u32 v9, v1, 8, 4
-; GFX9-NEXT:    v_bfe_u32 v10, v1, 12, 4
-; GFX9-NEXT:    v_bfe_u32 v1, v1, 4, 4
-; GFX9-NEXT:    v_bfe_u32 v12, v2, 24, 4
-; GFX9-NEXT:    v_lshrrev_b32_e32 v13, 28, v2
-; GFX9-NEXT:    v_bfe_u32 v14, v2, 16, 4
-; GFX9-NEXT:    v_bfe_u32 v15, v2, 20, 4
-; GFX9-NEXT:    v_bfe_u32 v16, v2, 8, 4
-; GFX9-NEXT:    v_bfe_u32 v17, v2, 12, 4
-; GFX9-NEXT:    v_bfe_u32 v2, v2, 4, 4
-; GFX9-NEXT:    v_and_b32_e32 v11, v4, v11
-; GFX9-NEXT:    v_and_b32_e32 v18, v4, v18
-; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v11
-; GFX9-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX9-NEXT:    v_and_b32_e32 v9, v4, v9
-; GFX9-NEXT:    v_and_b32_e32 v5, v4, v5
-; GFX9-NEXT:    v_and_b32_e32 v16, v4, v16
+; GFX9-NEXT:    v_lshrrev_b32_e32 v12, 28, v2
+; GFX9-NEXT:    v_bfe_u32 v13, v2, 24, 4
+; GFX9-NEXT:    v_bfe_u32 v14, v2, 20, 4
+; GFX9-NEXT:    v_bfe_u32 v15, v2, 16, 4
+; GFX9-NEXT:    v_bfe_u32 v16, v2, 12, 4
+; GFX9-NEXT:    v_bfe_u32 v17, v2, 8, 4
+; GFX9-NEXT:    v_bfe_u32 v18, v2, 4, 4
+; GFX9-NEXT:    v_and_b32_e32 v2, 15, v2
+; GFX9-NEXT:    v_and_b32_e32 v1, v4, v1
+; GFX9-NEXT:    v_and_b32_e32 v2, v4, v2
+; GFX9-NEXT:    v_lshl_or_b32 v1, v11, 16, v1
+; GFX9-NEXT:    v_lshl_or_b32 v2, v18, 16, v2
+; GFX9-NEXT:    v_and_b32_e32 v10, v4, v10
+; GFX9-NEXT:    v_and_b32_e32 v6, v4, v6
+; GFX9-NEXT:    v_and_b32_e32 v17, v4, v17
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v2, 12, v2 op_sel_hi:[0,1]
-; GFX9-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
-; GFX9-NEXT:    v_lshl_or_b32 v6, v17, 16, v16
+; GFX9-NEXT:    v_lshl_or_b32 v9, v9, 16, v10
+; GFX9-NEXT:    v_lshl_or_b32 v5, v5, 16, v6
+; GFX9-NEXT:    v_lshl_or_b32 v6, v16, 16, v17
 ; GFX9-NEXT:    v_pk_ashrrev_i16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_ashrrev_i16 v2, 12, v2 op_sel_hi:[0,1]
-; GFX9-NEXT:    v_and_b32_e32 v7, v4, v7
-; GFX9-NEXT:    v_and_b32_e32 v14, v4, v14
-; GFX9-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX9-NEXT:    v_and_b32_e32 v8, v4, v8
+; GFX9-NEXT:    v_and_b32_e32 v15, v4, v15
+; GFX9-NEXT:    v_and_b32_e32 v4, v4, v13
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v6, 12, v6 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_mul_lo_u16 v1, v1, v2
-; GFX9-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX9-NEXT:    v_lshl_or_b32 v8, v15, 16, v14
-; GFX9-NEXT:    v_lshl_or_b32 v4, v13, 16, v4
+; GFX9-NEXT:    v_lshl_or_b32 v7, v7, 16, v8
+; GFX9-NEXT:    v_lshl_or_b32 v8, v14, 16, v15
+; GFX9-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
 ; GFX9-NEXT:    v_pk_ashrrev_i16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_ashrrev_i16 v6, 12, v6 op_sel_hi:[0,1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_u16_e32 v2, v1, v3
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v5, 12, v5 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v8, 12, v8 op_sel_hi:[0,1]
 ; GFX9-NEXT:    v_pk_lshlrev_b16 v4, 12, v4 op_sel_hi:[0,1]

 ; GFX9-DL-NEXT:    v_mov_b32_e32 v4, 0xffff
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_and_b32_e32 v11, 15, v1
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 28, v1
+; GFX9-DL-NEXT:    v_bfe_u32 v6, v1, 24, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v7, v1, 20, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v8, v1, 16, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v9, v1, 12, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v10, v1, 8, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v11, v1, 4, 4
+; GFX9-DL-NEXT:    v_and_b32_e32 v1, 15, v1
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_and_b32_e32 v18, 15, v2
-; GFX9-DL-NEXT:    v_bfe_u32 v5, v1, 24, 4
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 28, v1
-; GFX9-DL-NEXT:    v_bfe_u32 v7, v1, 16, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v8, v1, 20, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v9, v1, 8, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v10, v1, 12, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v1, v1, 4, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v12, v2, 24, 4
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v13, 28, v2
-; GFX9-DL-NEXT:    v_bfe_u32 v14, v2, 16, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v15, v2, 20, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v16, v2, 8, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v17, v2, 12, 4
-; GFX9-DL-NEXT:    v_bfe_u32 v2, v2, 4, 4
-; GFX9-DL-NEXT:    v_and_b32_e32 v11, v4, v11
-; GFX9-DL-NEXT:    v_and_b32_e32 v18, v4, v18
-; GFX9-DL-NEXT:    v_lshl_or_b32 v1, v1, 16, v11
-; GFX9-DL-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX9-DL-NEXT:    v_and_b32_e32 v9, v4, v9
-; GFX9-DL-NEXT:    v_and_b32_e32 v5, v4, v5
-; GFX9-DL-NEXT:    v_and_b32_e32 v16, v4, v16
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v12, 28, v2
+; GFX9-DL-NEXT:    v_bfe_u32 v13, v2, 24, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v14, v2, 20, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v15, v2, 16, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v16, v2, 12, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v17, v2, 8, 4
+; GFX9-DL-NEXT:    v_bfe_u32 v18, v2, 4, 4
+; GFX9-DL-NEXT:    v_and_b32_e32 v2, 15, v2
+; GFX9-DL-NEXT:    v_and_b32_e32 v1, v4, v1
+; GFX9-DL-NEXT:    v_and_b32_e32 v2, v4, v2
+; GFX9-DL-NEXT:    v_lshl_or_b32 v1, v11, 16, v1
+; GFX9-DL-NEXT:    v_lshl_or_b32 v2, v18, 16, v2
+; GFX9-DL-NEXT:    v_and_b32_e32 v10, v4, v10
+; GFX9-DL-NEXT:    v_and_b32_e32 v6, v4, v6
+; GFX9-DL-NEXT:    v_and_b32_e32 v17, v4, v17
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v2, 12, v2 op_sel_hi:[0,1]
-; GFX9-DL-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX9-DL-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
-; GFX9-DL-NEXT:    v_lshl_or_b32 v6, v17, 16, v16
+; GFX9-DL-NEXT:    v_lshl_or_b32 v9, v9, 16, v10
+; GFX9-DL-NEXT:    v_lshl_or_b32 v5, v5, 16, v6
+; GFX9-DL-NEXT:    v_lshl_or_b32 v6, v16, 16, v17
 ; GFX9-DL-NEXT:    v_pk_ashrrev_i16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_ashrrev_i16 v2, 12, v2 op_sel_hi:[0,1]
-; GFX9-DL-NEXT:    v_and_b32_e32 v7, v4, v7
-; GFX9-DL-NEXT:    v_and_b32_e32 v14, v4, v14
-; GFX9-DL-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX9-DL-NEXT:    v_and_b32_e32 v8, v4, v8
+; GFX9-DL-NEXT:    v_and_b32_e32 v15, v4, v15
+; GFX9-DL-NEXT:    v_and_b32_e32 v4, v4, v13
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v6, 12, v6 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_mul_lo_u16 v1, v1, v2
-; GFX9-DL-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX9-DL-NEXT:    v_lshl_or_b32 v8, v15, 16, v14
-; GFX9-DL-NEXT:    v_lshl_or_b32 v4, v13, 16, v4
+; GFX9-DL-NEXT:    v_lshl_or_b32 v7, v7, 16, v8
+; GFX9-DL-NEXT:    v_lshl_or_b32 v8, v14, 16, v15
+; GFX9-DL-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
 ; GFX9-DL-NEXT:    v_pk_ashrrev_i16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_ashrrev_i16 v6, 12, v6 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_add_u16_e32 v2, v1, v3
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v5, 12, v5 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v8, 12, v8 op_sel_hi:[0,1]
 ; GFX9-DL-NEXT:    v_pk_lshlrev_b16 v4, 12, v4 op_sel_hi:[0,1]

 ; GFX10-DL-XNACK-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-DL-XNACK-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-XNACK-NEXT:    s_clause 0x1
 ; GFX10-DL-XNACK-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-XNACK-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-XNACK-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-DL-XNACK-NEXT:    global_load_ushort v3, v0, s[0:1]
 ; GFX10-DL-XNACK-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v11, 15, v1
+; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v5, 28, v1
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v6, v1, 24, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v7, v1, 20, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v8, v1, 16, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v9, v1, 12, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v10, v1, 8, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v11, v1, 4, 4
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v1, 15, v1
 ; GFX10-DL-XNACK-NEXT:    s_waitcnt vmcnt(1)
 ; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v13, 15, v2
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v5, v1, 24, 4
-; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v6, 28, v1
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v7, v1, 16, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v8, v1, 20, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v9, v1, 8, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v10, v1, 12, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v1, v1, 4, 4
 ; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v16, v2, 4, 4
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v11, v4, v11
+; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v12, 28, v2
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v14, v2, 24, 4
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v1, v4, v1
 ; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v13, v4, v13
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v18, v2, 8, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v12, v2, 24, 4
-; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v14, 28, v2
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v1, v1, 16, v11
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v15, v2, 20, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v17, v2, 16, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v18, v2, 12, 4
+; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v2, v2, 8, 4
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v1, v11, 16, v1
 ; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v11, v16, 16, v13
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v15, v2, 16, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v17, v2, 20, 4
-; GFX10-DL-XNACK-NEXT:    v_bfe_u32 v2, v2, 12, 4
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v9, v4, v9
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v7, v4, v7
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v13, v4, v18
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v10, v4, v10
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v8, v4, v8
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v2, v4, v2
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v11, 12, v11 op_sel_hi:[0,1]
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v2, v2, 16, v13
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v9, v9, 16, v10
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v7, v7, 16, v8
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v2, v18, 16, v2
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v8, 12, v11 op_sel_hi:[0,1]
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v10, v4, v15
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v10, v4, v17
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v2, 12, v2 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_mul_lo_u16 v1, v1, v8
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v8, v17, 16, v10
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v8, v15, 16, v10
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v2, 12, v2 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
 ; GFX10-DL-XNACK-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-XNACK-NEXT:    v_add_nc_u16 v1, v1, v3
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v3, v4, v5
-; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v5, 12, v8 op_sel_hi:[0,1]
-; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v3, v4, v6
+; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v6, 12, v8 op_sel_hi:[0,1]
+; GFX10-DL-XNACK-NEXT:    v_and_b32_e32 v4, v4, v14
 ; GFX10-DL-XNACK-NEXT:    v_pk_mul_lo_u16 v2, v9, v2
 ; GFX10-DL-XNACK-NEXT:    v_add_nc_u16 v1, v1, v10
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v3, v6, 16, v3
-; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v5, 12, v5 op_sel_hi:[0,1]
-; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v4, v14, 16, v4
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v3, v5, 16, v3
+; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v5, 12, v6 op_sel_hi:[0,1]
+; GFX10-DL-XNACK-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
 ; GFX10-DL-XNACK-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
 ; GFX10-DL-XNACK-NEXT:    v_add_nc_u16 v1, v1, v2
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v2, 12, v3 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_lshlrev_b16 v3, 12, v4 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_mul_lo_u16 v4, v7, v5
 ; GFX10-DL-XNACK-NEXT:    v_add_nc_u16 v1, v1, v6
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v2, 12, v2 op_sel_hi:[0,1]
 ; GFX10-DL-XNACK-NEXT:    v_pk_ashrrev_i16 v3, 12, v3 op_sel_hi:[0,1]

 ; GFX10-DL-NOXNACK-NEXT:    s_add_u32 s8, s8, s3
 ; GFX10-DL-NOXNACK-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-DL-NOXNACK-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NOXNACK-NEXT:    s_clause 0x1
 ; GFX10-DL-NOXNACK-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NOXNACK-NEXT:    global_load_dword v0, v0, s[6:7]
 ; GFX10-DL-NOXNACK-NEXT:    global_load_ushort v3, v2, s[0:1]
 ; GFX10-DL-NOXNACK-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v11, 15, v1
+; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v5, 28, v1
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v6, v1, 24, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v7, v1, 20, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v8, v1, 16, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v9, v1, 12, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v10, v1, 8, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v11, v1, 4, 4
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v1, 15, v1
 ; GFX10-DL-NOXNACK-NEXT:    s_waitcnt vmcnt(1)
 ; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v13, 15, v0
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v5, v1, 24, 4
-; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v6, 28, v1
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v7, v1, 16, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v8, v1, 20, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v9, v1, 8, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v10, v1, 12, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v1, v1, 4, 4
 ; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v16, v0, 4, 4
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v11, v4, v11
+; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v12, 28, v0
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v14, v0, 24, 4
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v1, v4, v1
 ; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v13, v4, v13
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v18, v0, 8, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v12, v0, 24, 4
-; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v14, 28, v0
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v1, v1, 16, v11
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v15, v0, 20, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v17, v0, 16, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v18, v0, 12, 4
+; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v0, v0, 8, 4
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v1, v11, 16, v1
 ; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v11, v16, 16, v13
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v15, v0, 16, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v17, v0, 20, 4
-; GFX10-DL-NOXNACK-NEXT:    v_bfe_u32 v0, v0, 12, 4
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v9, v4, v9
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v7, v4, v7
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v13, v4, v18
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v10, v4, v10
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v8, v4, v8
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v0, v4, v0
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v11, 12, v11 op_sel_hi:[0,1]
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v0, v0, 16, v13
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v9, v9, 16, v10
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v7, v7, 16, v8
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v8, 12, v11 op_sel_hi:[0,1]
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v10, v4, v15
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v10, v4, v17
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v0, 12, v0 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_mul_lo_u16 v1, v1, v8
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v8, v17, 16, v10
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v8, v15, 16, v10
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v9, 12, v9 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v0, 12, v0 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v7, 12, v7 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
 ; GFX10-DL-NOXNACK-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NOXNACK-NEXT:    v_add_nc_u16 v1, v1, v3
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v3, v4, v5
-; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v5, 12, v8 op_sel_hi:[0,1]
-; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v3, v4, v6
+; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v6, 12, v8 op_sel_hi:[0,1]
+; GFX10-DL-NOXNACK-NEXT:    v_and_b32_e32 v4, v4, v14
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_mul_lo_u16 v0, v9, v0
 ; GFX10-DL-NOXNACK-NEXT:    v_add_nc_u16 v1, v1, v10
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v3, v6, 16, v3
-; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v5, 12, v5 op_sel_hi:[0,1]
-; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v4, v14, 16, v4
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v3, v5, 16, v3
+; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v5, 12, v6 op_sel_hi:[0,1]
+; GFX10-DL-NOXNACK-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
 ; GFX10-DL-NOXNACK-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
 ; GFX10-DL-NOXNACK-NEXT:    v_add_nc_u16 v0, v1, v0
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v1, 12, v3 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_lshlrev_b16 v3, 12, v4 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_mul_lo_u16 v4, v7, v5
 ; GFX10-DL-NOXNACK-NEXT:    v_add_nc_u16 v0, v0, v6
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v1, 12, v1 op_sel_hi:[0,1]
 ; GFX10-DL-NOXNACK-NEXT:    v_pk_ashrrev_i16 v3, 12, v3 op_sel_hi:[0,1]

llvm/test/CodeGen/AMDGPU/vector_shuffle.packed.ll

; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 3, i32 undef, i32 undef, i32 7>
   ret <4 x half> %shuffle
 }
 
 define <4 x half> @shuffle_v4f16_35u5(<4 x half> addrspace(1)* %arg0, <4 x half> addrspace(1)* %arg1) {
 ; GFX9-LABEL: shuffle_v4f16_35u5:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_dword v4, v[2:3], off
 ; GFX9-NEXT:    global_load_dword v5, v[0:1], off offset:4
+; GFX9-NEXT:    global_load_dword v4, v[2:3], off
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_and_b32_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: shuffle_v4f16_35u5:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_load_dword v4, v[2:3], off
 ; GFX10-NEXT:    global_load_dword v5, v[0:1], off offset:4
+; GFX10-NEXT:    global_load_dword v4, v[2:3], off
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
-; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_and_b32_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
 ; GFX10-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <4 x half>, <4 x half> addrspace(1)* %arg0
   %val1 = load <4 x half>, <4 x half> addrspace(1)* %arg1
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 3, i32 5, i32 undef, i32 5>
   ret <4 x half> %shuffle
 }
 
 define <4 x half> @shuffle_v4f16_357u(<4 x half> addrspace(1)* %arg0, <4 x half> addrspace(1)* %arg1) {
 ; GFX9-LABEL: shuffle_v4f16_357u:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX9-NEXT:    global_load_dword v6, v[0:1], off offset:4
+; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_and_b32_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
 ; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: shuffle_v4f16_357u:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX10-NEXT:    global_load_dword v6, v[0:1], off offset:4
+; GFX10-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
-; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_and_b32_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
 ; GFX10-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <4 x half>, <4 x half> addrspace(1)* %arg0
   %val1 = load <4 x half>, <4 x half> addrspace(1)* %arg1
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 3, i32 5, i32 7, i32 undef>
   ret <4 x half> %shuffle
 }

 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    global_load_dword v6, v[0:1], off offset:4
 ; GFX9-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_and_b32_sdwa v1, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
 ; GFX9-NEXT:    v_and_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
 ; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 16, v1
 ; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: shuffle_v4f16_5734:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    global_load_dwordx2 v[4:5], v[2:3], off
 ; GFX10-NEXT:    global_load_dword v6, v[0:1], off offset:4
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
-; GFX10-NEXT:    v_and_b32_sdwa v2, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT:    v_and_b32_sdwa v1, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_and_b32_sdwa v3, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX10-NEXT:    v_lshl_or_b32 v0, v1, 16, v2
+; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
 ; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <4 x half>, <4 x half> addrspace(1)* %arg0
   %val1 = load <4 x half>, <4 x half> addrspace(1)* %arg1
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 5, i32 7, i32 3, i32 4>
   ret <4 x half> %shuffle
 }
 
; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 1, i32 1, i32 0, i32 0>
   ret <4 x half> %shuffle
 }
 
 define <4 x half> @shuffle_v4f16_6161(<4 x half> addrspace(1)* %arg0, <4 x half> addrspace(1)* %arg1) {
 ; GFX9-LABEL: shuffle_v4f16_6161:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_dword v4, v[0:1], off
-; GFX9-NEXT:    global_load_dword v5, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_dword v4, v[2:3], off offset:4
+; GFX9-NEXT:    global_load_dword v5, v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
+; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v4
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v5
-; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
+; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: shuffle_v4f16_6161:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_load_dword v4, v[0:1], off
-; GFX10-NEXT:    global_load_dword v5, v[2:3], off offset:4
+; GFX10-NEXT:    global_load_dword v4, v[2:3], off offset:4
+; GFX10-NEXT:    global_load_dword v5, v[0:1], off
 ; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
+; GFX10-NEXT:    v_and_b32_e32 v0, 0xffff, v4
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff, v5
-; GFX10-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
+; GFX10-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <4 x half>, <4 x half> addrspace(1)* %arg0
   %val1 = load <4 x half>, <4 x half> addrspace(1)* %arg1
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 6, i32 1, i32 6, i32 1>
   ret <4 x half> %shuffle
 }