[SDAG] Convert FSHL <--> FSHR if the target only supports one of them

[foad]

Summary

D77152 tried to do this but got it wrong in the shift-by-zero case.
D86430 reverted the wrong code. Reimplement the optimization with
different code depending on whether the shift amount is known to be
non-zero (modulo bitwidth).

This improves code quality for fshl tests on AMDGPU, which only has an
fshr instruction.

[bjope]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

6181

I find these comments a bit confusing, as the full transform is

// fshl X, Y, Z -> fshr (srl X, 1), (fshr (X, Y, 1), ~Z
// fshr X, Y, Z -> fshl (fshl X, Y, 1), (shl (Y, 1), ~Z

I suggest that we either show the full transform like above, or for example add some primes to X and Y on the right hand side to show that it isn't the same X and Y as on the left hand side:

// fshl X, Y, Z -> fshr X', Y', ~Z
// fshr X, Y, Z -> fshl X', Y', ~Z

[bjope]

LGTM (except the earlier nit about the comments with different X and Y values on left and right hand side)

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

bool TargetLowering::expandFunnelShift(SDNode *Node, SDValue &Result,
   SDValue Z = Node->getOperand(2);
 
   unsigned BW = VT.getScalarSizeInBits();
   bool IsFSHL = Node->getOpcode() == ISD::FSHL;
   SDLoc DL(SDValue(Node, 0));
 
   EVT ShVT = Z.getValueType();
 
+  // If a funnel shift in the other direction is more supported, use it.
+  unsigned RevOpcode = IsFSHL ? ISD::FSHR : ISD::FSHL;
+  if (!isOperationLegalOrCustom(Node->getOpcode(), VT) &&
+      isOperationLegalOrCustom(RevOpcode, VT) && isPowerOf2_32(BW)) {
+    if (isNonZeroModBitWidthOrUndef(Z, BW)) {
+      // fshl X, Y, Z -> fshr X, Y, -Z
+      // fshr X, Y, Z -> fshl X, Y, -Z
+      SDValue Zero = DAG.getConstant(0, DL, ShVT);
+      Z = DAG.getNode(ISD::SUB, DL, VT, Zero, Z);
+    } else {
+      // First do a full double-word shift of { X : Y } by 1 and then:
+      // fshl X, Y, Z -> fshr X, Y, ~Z
+      // fshr X, Y, Z -> fshl X, Y, ~Z
+      SDValue One = DAG.getConstant(1, DL, ShVT);
+      if (IsFSHL) {
+        Y = DAG.getNode(RevOpcode, DL, VT, X, Y, One);
+        X = DAG.getNode(ISD::SRL, DL, VT, X, One);
+      } else {
+        X = DAG.getNode(RevOpcode, DL, VT, X, Y, One);
+        Y = DAG.getNode(ISD::SHL, DL, VT, Y, One);
+      }
+      Z = DAG.getNOT(DL, Z, ShVT);
+    }
+    Result = DAG.getNode(RevOpcode, DL, VT, X, Y, Z);
+    return true;
+  }
+
   SDValue ShX, ShY;
   SDValue ShAmt, InvShAmt;
   if (isNonZeroModBitWidthOrUndef(Z, BW)) {
     // fshl: X << C | Y >> (BW - C)
     // fshr: X << (BW - C) | Y >> C
     // where C = Z % BW is not zero
     SDValue BitWidthC = DAG.getConstant(BW, DL, ShVT);
     ShAmt = DAG.getNode(ISD::UREM, DL, ShVT, Z, BitWidthC);

llvm/test/CodeGen/AMDGPU/fshl.ll

 define amdgpu_kernel void @fshl_i32(i32 addrspace(1)* %in, i32 %x, i32 %y, i32 %z) {
 ; SI-LABEL: fshl_i32:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0xb
 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 ; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_andn2_b32 s3, 31, s2
-; SI-NEXT:    s_lshr_b32 s1, s1, 1
-; SI-NEXT:    s_and_b32 s2, s2, 31
-; SI-NEXT:    s_lshr_b32 s1, s1, s3
-; SI-NEXT:    s_lshl_b32 s0, s0, s2
-; SI-NEXT:    s_or_b32 s0, s0, s1
-; SI-NEXT:    v_mov_b32_e32 v0, s0
+; SI-NEXT:    v_mov_b32_e32 v0, s1
+; SI-NEXT:    s_lshr_b32 s1, s0, 1
+; SI-NEXT:    v_alignbit_b32 v0, s0, v0, 1
+; SI-NEXT:    s_not_b32 s0, s2
+; SI-NEXT:    v_mov_b32_e32 v1, s0
+; SI-NEXT:    v_alignbit_b32 v0, s1, v0, v1
 ; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: fshl_i32:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s1
+; VI-NEXT:    s_not_b32 s2, s2
+; VI-NEXT:    s_lshr_b32 s1, s0, 1
+; VI-NEXT:    v_alignbit_b32 v0, s0, v0, 1
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    v_alignbit_b32 v2, s1, v0, v1
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_andn2_b32 s3, 31, s2
-; VI-NEXT:    s_lshr_b32 s1, s1, 1
-; VI-NEXT:    s_and_b32 s2, s2, 31
-; VI-NEXT:    s_lshr_b32 s1, s1, s3
-; VI-NEXT:    s_lshl_b32 s0, s0, s2
-; VI-NEXT:    s_or_b32 s0, s0, s1
 ; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    v_mov_b32_e32 v2, s0
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: fshl_i32:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x2c
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s1
+; GFX9-NEXT:    s_not_b32 s2, s2
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 1
+; GFX9-NEXT:    v_alignbit_b32 v0, s0, v0, 1
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_alignbit_b32 v2, s1, v0, v1
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    s_andn2_b32 s3, 31, s2
-; GFX9-NEXT:    s_lshr_b32 s1, s1, 1
-; GFX9-NEXT:    s_and_b32 s2, s2, 31
-; GFX9-NEXT:    s_lshr_b32 s1, s1, s3
-; GFX9-NEXT:    s_lshl_b32 s0, s0, s2
-; GFX9-NEXT:    s_or_b32 s0, s0, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX9-NEXT:    global_store_dword v[0:1], v2, off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; R600-LABEL: fshl_i32:
 ; R600:       ; %bb.0: ; %entry
-; R600-NEXT:    ALU 9, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    ALU 5, @4, KC0[CB0:0-32], KC1[]
 ; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
 ; R600-NEXT:    CF_END
 ; R600-NEXT:    PAD
 ; R600-NEXT:    ALU clause starting at 4:
-; R600-NEXT:     NOT_INT * T0.W, KC0[3].X,
-; R600-NEXT:     AND_INT T0.Z, KC0[3].X, literal.x,
-; R600-NEXT:     AND_INT T0.W, PV.W, literal.x,
-; R600-NEXT:     LSHR * T1.W, KC0[2].W, 1,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     LSHR T0.W, PS, PV.W,
-; R600-NEXT:     LSHL * T1.W, KC0[2].Z, PV.Z,
-; R600-NEXT:     OR_INT T0.X, PS, PV.W,
+; R600-NEXT:     LSHR T0.Z, KC0[2].Z, 1,
+; R600-NEXT:     BIT_ALIGN_INT T0.W, KC0[2].Z, KC0[2].W, 1,
+; R600-NEXT:     NOT_INT * T1.W, KC0[3].X,
+; R600-NEXT:     BIT_ALIGN_INT T0.X, PV.Z, PV.W, PS,
 ; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
 ; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 entry:
   %0 = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %z)
   store i32 %0, i32 addrspace(1)* %in
   ret void
 }
 

 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xb
 ; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xf
 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 ; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshr_b32 s9, s9, 1
-; SI-NEXT:    s_andn2_b32 s10, 31, s1
-; SI-NEXT:    s_and_b32 s1, s1, 31
-; SI-NEXT:    s_lshl_b32 s1, s3, s1
-; SI-NEXT:    s_andn2_b32 s3, 31, s0
-; SI-NEXT:    s_and_b32 s0, s0, 31
-; SI-NEXT:    s_lshr_b32 s8, s8, 1
-; SI-NEXT:    s_lshr_b32 s9, s9, s10
-; SI-NEXT:    s_lshr_b32 s3, s8, s3
-; SI-NEXT:    s_lshl_b32 s0, s2, s0
-; SI-NEXT:    s_or_b32 s1, s1, s9
-; SI-NEXT:    s_or_b32 s0, s0, s3
-; SI-NEXT:    v_mov_b32_e32 v0, s0
+; SI-NEXT:    v_mov_b32_e32 v0, s9
+; SI-NEXT:    s_not_b32 s1, s1
+; SI-NEXT:    v_alignbit_b32 v0, s3, v0, 1
 ; SI-NEXT:    v_mov_b32_e32 v1, s1
+; SI-NEXT:    s_lshr_b32 s3, s3, 1
+; SI-NEXT:    v_alignbit_b32 v1, s3, v0, v1
+; SI-NEXT:    v_mov_b32_e32 v0, s8
+; SI-NEXT:    s_not_b32 s0, s0
+; SI-NEXT:    v_alignbit_b32 v0, s2, v0, 1
+; SI-NEXT:    s_lshr_b32 s1, s2, 1
+; SI-NEXT:    v_mov_b32_e32 v2, s0
+; SI-NEXT:    v_alignbit_b32 v0, s1, v0, v2
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: fshl_v2i32:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x2c
 ; VI-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    s_not_b32 s1, s1
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    s_lshr_b32 s7, s5, 1
+; VI-NEXT:    v_alignbit_b32 v0, s5, v0, 1
+; VI-NEXT:    v_alignbit_b32 v1, s7, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v0, s6
+; VI-NEXT:    s_not_b32 s0, s0
+; VI-NEXT:    v_alignbit_b32 v0, s4, v0, 1
+; VI-NEXT:    s_lshr_b32 s1, s4, 1
+; VI-NEXT:    v_mov_b32_e32 v2, s0
+; VI-NEXT:    v_alignbit_b32 v0, s1, v0, v2
 ; VI-NEXT:    v_mov_b32_e32 v2, s2
 ; VI-NEXT:    v_mov_b32_e32 v3, s3
-; VI-NEXT:    s_lshr_b32 s7, s7, 1
-; VI-NEXT:    s_andn2_b32 s8, 31, s1
-; VI-NEXT:    s_and_b32 s1, s1, 31
-; VI-NEXT:    s_lshl_b32 s1, s5, s1
-; VI-NEXT:    s_andn2_b32 s5, 31, s0
-; VI-NEXT:    s_and_b32 s0, s0, 31
-; VI-NEXT:    s_lshr_b32 s6, s6, 1
-; VI-NEXT:    s_lshr_b32 s7, s7, s8
-; VI-NEXT:    s_lshr_b32 s5, s6, s5
-; VI-NEXT:    s_lshl_b32 s0, s4, s0
-; VI-NEXT:    s_or_b32 s1, s1, s7
-; VI-NEXT:    s_or_b32 s0, s0, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: fshl_v2i32:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    s_not_b32 s1, s1
+; GFX9-NEXT:    v_mov_b32_e32 v1, s1
+; GFX9-NEXT:    s_lshr_b32 s7, s5, 1
+; GFX9-NEXT:    v_alignbit_b32 v0, s5, v0, 1
+; GFX9-NEXT:    v_alignbit_b32 v1, s7, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, s6
+; GFX9-NEXT:    s_not_b32 s0, s0
+; GFX9-NEXT:    v_alignbit_b32 v0, s4, v0, 1
+; GFX9-NEXT:    s_lshr_b32 s1, s4, 1
+; GFX9-NEXT:    v_mov_b32_e32 v2, s0
+; GFX9-NEXT:    v_alignbit_b32 v0, s1, v0, v2
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX9-NEXT:    v_mov_b32_e32 v3, s3
-; GFX9-NEXT:    s_lshr_b32 s7, s7, 1
-; GFX9-NEXT:    s_andn2_b32 s8, 31, s1
-; GFX9-NEXT:    s_and_b32 s1, s1, 31
-; GFX9-NEXT:    s_lshl_b32 s1, s5, s1
-; GFX9-NEXT:    s_andn2_b32 s5, 31, s0
-; GFX9-NEXT:    s_and_b32 s0, s0, 31
-; GFX9-NEXT:    s_lshr_b32 s6, s6, 1
-; GFX9-NEXT:    s_lshr_b32 s7, s7, s8
-; GFX9-NEXT:    s_lshr_b32 s5, s6, s5
-; GFX9-NEXT:    s_lshl_b32 s0, s4, s0
-; GFX9-NEXT:    s_or_b32 s1, s1, s7
-; GFX9-NEXT:    s_or_b32 s0, s0, s5
-; GFX9-NEXT:    v_mov_b32_e32 v0, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s1
 ; GFX9-NEXT:    global_store_dwordx2 v[2:3], v[0:1], off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; R600-LABEL: fshl_v2i32:
 ; R600:       ; %bb.0: ; %entry
-; R600-NEXT:    ALU 17, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    ALU 9, @4, KC0[CB0:0-32], KC1[]
 ; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
 ; R600-NEXT:    CF_END
 ; R600-NEXT:    PAD
 ; R600-NEXT:    ALU clause starting at 4:
-; R600-NEXT:     NOT_INT * T0.W, KC0[4].X,
-; R600-NEXT:     AND_INT T0.Y, KC0[4].X, literal.x,
-; R600-NEXT:     AND_INT T0.Z, PV.W, literal.x,
-; R600-NEXT:     LSHR T0.W, KC0[3].Z, 1,
+; R600-NEXT:     LSHR T0.Z, KC0[3].X, 1,
+; R600-NEXT:     BIT_ALIGN_INT * T0.W, KC0[3].X, KC0[3].Z, 1,
+; R600-NEXT:     NOT_INT * T1.W, KC0[4].X,
+; R600-NEXT:     BIT_ALIGN_INT T0.Y, T0.Z, T0.W, PV.W,
+; R600-NEXT:     LSHR T0.Z, KC0[2].W, 1,
+; R600-NEXT:     BIT_ALIGN_INT * T0.W, KC0[2].W, KC0[3].Y, 1,
 ; R600-NEXT:     NOT_INT * T1.W, KC0[3].W,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     AND_INT T0.X, KC0[3].W, literal.x,
-; R600-NEXT:     AND_INT T1.Y, PS, literal.x,
-; R600-NEXT:     LSHR T1.Z, KC0[3].Y, 1,
-; R600-NEXT:     LSHR T0.W, PV.W, PV.Z,
-; R600-NEXT:     LSHL * T1.W, KC0[3].X, PV.Y,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     OR_INT T0.Y, PS, PV.W,
-; R600-NEXT:     LSHR T0.W, PV.Z, PV.Y,
-; R600-NEXT:     LSHL * T1.W, KC0[2].W, PV.X,
-; R600-NEXT:     OR_INT T0.X, PS, PV.W,
+; R600-NEXT:     BIT_ALIGN_INT T0.X, T0.Z, T0.W, PV.W,
 ; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
 ; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 entry:
   %0 = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> %z)
   store <2 x i32> %0, <2 x i32> addrspace(1)* %in
   ret void
 }
 

 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0xd
 ; SI-NEXT:    s_load_dwordx4 s[12:15], s[0:1], 0x11
 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x15
 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 ; SI-NEXT:    s_mov_b32 s6, -1
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshr_b32 s14, s14, 1
-; SI-NEXT:    s_andn2_b32 s16, 31, s3
-; SI-NEXT:    s_and_b32 s3, s3, 31
-; SI-NEXT:    s_lshl_b32 s3, s11, s3
-; SI-NEXT:    s_andn2_b32 s11, 31, s2
-; SI-NEXT:    s_and_b32 s2, s2, 31
-; SI-NEXT:    s_lshl_b32 s2, s10, s2
-; SI-NEXT:    s_lshr_b32 s11, s14, s11
-; SI-NEXT:    s_or_b32 s2, s2, s11
-; SI-NEXT:    s_andn2_b32 s10, 31, s1
-; SI-NEXT:    s_and_b32 s1, s1, 31
-; SI-NEXT:    s_lshr_b32 s11, s13, 1
-; SI-NEXT:    s_lshl_b32 s1, s9, s1
-; SI-NEXT:    s_lshr_b32 s10, s11, s10
-; SI-NEXT:    s_lshr_b32 s15, s15, 1
-; SI-NEXT:    s_or_b32 s1, s1, s10
-; SI-NEXT:    s_andn2_b32 s9, 31, s0
-; SI-NEXT:    s_and_b32 s0, s0, 31
-; SI-NEXT:    s_lshr_b32 s10, s12, 1
-; SI-NEXT:    s_lshr_b32 s15, s15, s16
-; SI-NEXT:    s_lshr_b32 s9, s10, s9
-; SI-NEXT:    s_lshl_b32 s0, s8, s0
-; SI-NEXT:    s_or_b32 s3, s3, s15
-; SI-NEXT:    s_or_b32 s0, s0, s9
-; SI-NEXT:    v_mov_b32_e32 v0, s0
+; SI-NEXT:    v_mov_b32_e32 v0, s15
+; SI-NEXT:    s_not_b32 s3, s3
+; SI-NEXT:    v_alignbit_b32 v0, s11, v0, 1
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_lshr_b32 s11, s11, 1
+; SI-NEXT:    v_alignbit_b32 v3, s11, v0, v1
+; SI-NEXT:    v_mov_b32_e32 v0, s14
+; SI-NEXT:    s_not_b32 s2, s2
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    v_alignbit_b32 v0, s10, v0, 1
+; SI-NEXT:    s_lshr_b32 s3, s10, 1
+; SI-NEXT:    v_alignbit_b32 v2, s3, v0, v1
+; SI-NEXT:    v_mov_b32_e32 v0, s13
+; SI-NEXT:    s_not_b32 s1, s1
 ; SI-NEXT:    v_mov_b32_e32 v1, s1
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v3, s3
+; SI-NEXT:    v_alignbit_b32 v0, s9, v0, 1
+; SI-NEXT:    s_lshr_b32 s2, s9, 1
+; SI-NEXT:    v_alignbit_b32 v1, s2, v0, v1
+; SI-NEXT:    v_mov_b32_e32 v0, s12
+; SI-NEXT:    s_not_b32 s0, s0
+; SI-NEXT:    v_alignbit_b32 v0, s8, v0, 1
+; SI-NEXT:    s_lshr_b32 s1, s8, 1
+; SI-NEXT:    v_mov_b32_e32 v4, s0
+; SI-NEXT:    v_alignbit_b32 v0, s1, v0, v4
 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: fshl_v4i32:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[12:13], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x44
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x54
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s11
+; VI-NEXT:    s_not_b32 s3, s3
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    s_lshr_b32 s11, s7, 1
+; VI-NEXT:    v_alignbit_b32 v0, s7, v0, 1
+; VI-NEXT:    v_alignbit_b32 v3, s11, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v0, s10
+; VI-NEXT:    s_not_b32 s2, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    v_alignbit_b32 v0, s6, v0, 1
+; VI-NEXT:    s_lshr_b32 s3, s6, 1
+; VI-NEXT:    v_alignbit_b32 v2, s3, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    s_not_b32 s1, s1
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_alignbit_b32 v0, s5, v0, 1
+; VI-NEXT:    s_lshr_b32 s2, s5, 1
+; VI-NEXT:    v_alignbit_b32 v1, s2, v0, v1
+; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_not_b32 s0, s0
+; VI-NEXT:    v_alignbit_b32 v0, s4, v0, 1
+; VI-NEXT:    s_lshr_b32 s1, s4, 1
+; VI-NEXT:    v_mov_b32_e32 v4, s0
+; VI-NEXT:    v_alignbit_b32 v0, s1, v0, v4
 ; VI-NEXT:    v_mov_b32_e32 v4, s12
 ; VI-NEXT:    v_mov_b32_e32 v5, s13
-; VI-NEXT:    s_lshr_b32 s10, s10, 1
-; VI-NEXT:    s_andn2_b32 s14, 31, s3
-; VI-NEXT:    s_and_b32 s3, s3, 31
-; VI-NEXT:    s_lshl_b32 s3, s7, s3
-; VI-NEXT:    s_andn2_b32 s7, 31, s2
-; VI-NEXT:    s_and_b32 s2, s2, 31
-; VI-NEXT:    s_lshl_b32 s2, s6, s2
-; VI-NEXT:    s_lshr_b32 s7, s10, s7
-; VI-NEXT:    s_or_b32 s2, s2, s7
-; VI-NEXT:    s_andn2_b32 s6, 31, s1
-; VI-NEXT:    s_and_b32 s1, s1, 31
-; VI-NEXT:    s_lshr_b32 s7, s9, 1
-; VI-NEXT:    s_lshl_b32 s1, s5, s1
-; VI-NEXT:    s_lshr_b32 s6, s7, s6
-; VI-NEXT:    s_lshr_b32 s11, s11, 1
-; VI-NEXT:    s_or_b32 s1, s1, s6
-; VI-NEXT:    s_andn2_b32 s5, 31, s0
-; VI-NEXT:    s_and_b32 s0, s0, 31
-; VI-NEXT:    s_lshr_b32 s6, s8, 1
-; VI-NEXT:    s_lshr_b32 s11, s11, s14
-; VI-NEXT:    s_lshr_b32 s5, s6, s5
-; VI-NEXT:    s_lshl_b32 s0, s4, s0
-; VI-NEXT:    s_or_b32 s3, s3, s11
-; VI-NEXT:    s_or_b32 s0, s0, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s2
-; VI-NEXT:    v_mov_b32_e32 v3, s3
 ; VI-NEXT:    flat_store_dwordx4 v[4:5], v[0:3]
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: fshl_v4i32:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[12:13], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[8:11], s[0:1], 0x44
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x54
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s11
+; GFX9-NEXT:    s_not_b32 s3, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_lshr_b32 s11, s7, 1
+; GFX9-NEXT:    v_alignbit_b32 v0, s7, v0, 1
+; GFX9-NEXT:    v_alignbit_b32 v3, s11, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, s10
+; GFX9-NEXT:    s_not_b32 s2, s2
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_alignbit_b32 v0, s6, v0, 1
+; GFX9-NEXT:    s_lshr_b32 s3, s6, 1
+; GFX9-NEXT:    v_alignbit_b32 v2, s3, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, s9
+; GFX9-NEXT:    s_not_b32 s1, s1
+; GFX9-NEXT:    v_mov_b32_e32 v1, s1
+; GFX9-NEXT:    v_alignbit_b32 v0, s5, v0, 1
+; GFX9-NEXT:    s_lshr_b32 s2, s5, 1
+; GFX9-NEXT:    v_alignbit_b32 v1, s2, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    s_not_b32 s0, s0
+; GFX9-NEXT:    v_alignbit_b32 v0, s4, v0, 1
+; GFX9-NEXT:    s_lshr_b32 s1, s4, 1
+; GFX9-NEXT:    v_mov_b32_e32 v4, s0
+; GFX9-NEXT:    v_alignbit_b32 v0, s1, v0, v4
 ; GFX9-NEXT:    v_mov_b32_e32 v4, s12
 ; GFX9-NEXT:    v_mov_b32_e32 v5, s13
-; GFX9-NEXT:    s_lshr_b32 s10, s10, 1
-; GFX9-NEXT:    s_andn2_b32 s14, 31, s3
-; GFX9-NEXT:    s_and_b32 s3, s3, 31
-; GFX9-NEXT:    s_lshl_b32 s3, s7, s3
-; GFX9-NEXT:    s_andn2_b32 s7, 31, s2
-; GFX9-NEXT:    s_and_b32 s2, s2, 31
-; GFX9-NEXT:    s_lshl_b32 s2, s6, s2
-; GFX9-NEXT:    s_lshr_b32 s7, s10, s7
-; GFX9-NEXT:    s_or_b32 s2, s2, s7
-; GFX9-NEXT:    s_andn2_b32 s6, 31, s1
-; GFX9-NEXT:    s_and_b32 s1, s1, 31
-; GFX9-NEXT:    s_lshr_b32 s7, s9, 1
-; GFX9-NEXT:    s_lshl_b32 s1, s5, s1
-; GFX9-NEXT:    s_lshr_b32 s6, s7, s6
-; GFX9-NEXT:    s_lshr_b32 s11, s11, 1
-; GFX9-NEXT:    s_or_b32 s1, s1, s6
-; GFX9-NEXT:    s_andn2_b32 s5, 31, s0
-; GFX9-NEXT:    s_and_b32 s0, s0, 31
-; GFX9-NEXT:    s_lshr_b32 s6, s8, 1
-; GFX9-NEXT:    s_lshr_b32 s11, s11, s14
-; GFX9-NEXT:    s_lshr_b32 s5, s6, s5
-; GFX9-NEXT:    s_lshl_b32 s0, s4, s0
-; GFX9-NEXT:    s_or_b32 s3, s3, s11
-; GFX9-NEXT:    s_or_b32 s0, s0, s5
-; GFX9-NEXT:    v_mov_b32_e32 v0, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s1
-; GFX9-NEXT:    v_mov_b32_e32 v2, s2
-; GFX9-NEXT:    v_mov_b32_e32 v3, s3
 ; GFX9-NEXT:    global_store_dwordx4 v[4:5], v[0:3], off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; R600-LABEL: fshl_v4i32:
 ; R600:       ; %bb.0: ; %entry
-; R600-NEXT:    ALU 35, @4, KC0[CB0:0-32], KC1[]
-; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T2.XYZW, T0.X, 1
+; R600-NEXT:    ALU 17, @4, KC0[CB0:0-32], KC1[]
+; R600-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1
 ; R600-NEXT:    CF_END
 ; R600-NEXT:    PAD
 ; R600-NEXT:    ALU clause starting at 4:
-; R600-NEXT:     AND_INT * T0.W, KC0[5].Y, literal.x,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     LSHR T0.Z, KC0[4].Z, 1,
-; R600-NEXT:     NOT_INT T1.W, KC0[6].X,
-; R600-NEXT:     AND_INT * T2.W, KC0[6].X, literal.x,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     AND_INT T0.X, KC0[5].Z, literal.x,
-; R600-NEXT:     LSHL T0.Y, KC0[4].X, PS,
-; R600-NEXT:     NOT_INT T1.Z, KC0[5].W,
-; R600-NEXT:     AND_INT * T1.W, PV.W, literal.x,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     LSHR * T2.W, KC0[5].X, 1,
-; R600-NEXT:     LSHR T1.X, PV.W, T1.W,
-; R600-NEXT:     AND_INT T1.Y, KC0[5].W, literal.x,
-; R600-NEXT:     AND_INT T1.Z, T1.Z, literal.x,
-; R600-NEXT:     LSHR T1.W, KC0[4].W, 1,
+; R600-NEXT:     LSHR T0.Z, KC0[4].X, 1,
+; R600-NEXT:     BIT_ALIGN_INT * T0.W, KC0[4].X, KC0[5].X, 1,
+; R600-NEXT:     NOT_INT * T1.W, KC0[6].X,
+; R600-NEXT:     LSHR T0.Y, KC0[3].W, 1,
+; R600-NEXT:     BIT_ALIGN_INT T1.Z, KC0[3].W, KC0[4].W, 1,
+; R600-NEXT:     BIT_ALIGN_INT * T0.W, T0.Z, T0.W, PV.W,
+; R600-NEXT:     NOT_INT * T1.W, KC0[5].W,
+; R600-NEXT:     LSHR T1.Y, KC0[3].Z, 1,
+; R600-NEXT:     BIT_ALIGN_INT T0.Z, T0.Y, T1.Z, PV.W,
+; R600-NEXT:     BIT_ALIGN_INT * T1.W, KC0[3].Z, KC0[4].Z, 1,
 ; R600-NEXT:     NOT_INT * T2.W, KC0[5].Z,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     AND_INT T2.X, PS, literal.x,
-; R600-NEXT:     LSHR T2.Y, PV.W, PV.Z,
-; R600-NEXT:     LSHL T1.Z, KC0[3].W, PV.Y,
-; R600-NEXT:     NOT_INT T1.W, KC0[5].Y,
-; R600-NEXT:     OR_INT * T2.W, T0.Y, PV.X,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     AND_INT T1.X, PV.W, literal.x,
-; R600-NEXT:     LSHR T0.Y, KC0[4].Y, 1,
-; R600-NEXT:     OR_INT T2.Z, PV.Z, PV.Y,
-; R600-NEXT:     LSHR T1.W, T0.Z, PV.X,
-; R600-NEXT:     LSHL * T3.W, KC0[3].Z, T0.X,
-; R600-NEXT:    31(4.344025e-44), 0(0.000000e+00)
-; R600-NEXT:     OR_INT T2.Y, PS, PV.W,
-; R600-NEXT:     LSHR T1.W, PV.Y, PV.X,
-; R600-NEXT:     LSHL * T0.W, KC0[3].Y, T0.W,
-; R600-NEXT:     OR_INT T2.X, PS, PV.W,
-; R600-NEXT:     LSHR * T0.X, KC0[2].Y, literal.x,
+; R600-NEXT:     BIT_ALIGN_INT T0.Y, T1.Y, T1.W, PV.W,
+; R600-NEXT:     LSHR T1.Z, KC0[3].Y, 1,
+; R600-NEXT:     BIT_ALIGN_INT * T1.W, KC0[3].Y, KC0[4].Y, 1,
+; R600-NEXT:     NOT_INT * T2.W, KC0[5].Y,
+; R600-NEXT:     BIT_ALIGN_INT T0.X, T1.Z, T1.W, PV.W,
+; R600-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
 ; R600-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 entry:
   %0 = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %z)
   store <4 x i32> %0, <4 x i32> addrspace(1)* %in
   ret void
 }
 
 define amdgpu_kernel void @fshl_v4i32_imm(<4 x i32> addrspace(1)* %in, <4 x i32> %x, <4 x i32> %y) {