[DAG] PromoteIntRes_ADDSUBSHLSAT - use promoted ISD::USUBSAT directly

[RKSimon]

Summary

As discussed on D96413, as long as the promoted bits of the args are zero we can use the basic ISD::USUBSAT pattern directly, without the shifting like we do for other ops.

I think something similar should be possible for ISD::UADDSAT as well, which I'll look at later.

Also, create a ISD::USUBSAT node directly - this will be expanded back by the legalizer later on if necessary.

https://alive2.llvm.org/ce/

[nikic]

LGTM

I think something similar should be possible for ISD::UADDSAT as well, which I'll look at later.

This works for usubsat because the lower bound stays the same (0), while for uaddsat it would change (from 2^bw-1 to 2^new_bw-1). So not sure a similar pattern is possible for that case.

https://alive2.llvm.org/ce/

I think you posted the wrong link there :)

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

808

Personally, I'd drop this assert. The zext is explicitly done up above, and doing a known bits query for this seems weird to me.

[nikic]

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

810

Structurally, it might make sense to move the USUBSAT check before the IsShift || TLI.isOperationLegalOrCustom(Opcode, PromotedType) branch, as we'll now unconditionally take it. In that case you also don't need to change the shift code, and can keep the llvm_unreachable in there (not sure if you switch as written may generate warnings).

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

810

We could also just dispatch USUBSAT to PromoteIntRes_ZExtIntBinOp in the main switch.

[RKSimon]

I think something similar should be possible for ISD::UADDSAT as well, which I'll look at later.

This works for usubsat because the lower bound stays the same (0), while for uaddsat it would change (from 2^bw-1 to 2^new_bw-1). So not sure a similar pattern is possible for that case.

What I noticed was that removing UADDSAT from the 'shift' codegen (and always using the default expansion further down) seemed to generate better code - I'll take another look.

We could also just dispatch USUBSAT to PromoteIntRes_ZExtIntBinOp in the main switch.

Thanks - I'd forgotten about that!

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

if (IsShift) {
     Op2Promoted = SExtPromotedInteger(Op2);
   }
   EVT PromotedType = Op1Promoted.getValueType();
   unsigned NewBits = PromotedType.getScalarSizeInBits();
 
   // Shift cannot use a min/max expansion, we can't detect overflow if all of
   // the bits have been shifted out.
   if (IsShift || TLI.isOperationLegalOrCustom(Opcode, PromotedType)) {
-    unsigned ShiftOp;
-    switch (Opcode) {
-    case ISD::SADDSAT:
-    case ISD::SSUBSAT:
-    case ISD::SSHLSAT:
-      ShiftOp = ISD::SRA;
-      break;
-    case ISD::UADDSAT:
-    case ISD::USUBSAT:
-    case ISD::USHLSAT:
-      ShiftOp = ISD::SRL;
-      break;
-    default:
-      llvm_unreachable("Expected opcode to be signed or unsigned saturation "
-                       "addition, subtraction or left shift");
-    }
-
+    auto GetShiftedOp = [&](unsigned ShiftOp) {
       unsigned SHLAmount = NewBits - OldBits;
       EVT SHVT = TLI.getShiftAmountTy(PromotedType, DAG.getDataLayout());
       SDValue ShiftAmount = DAG.getConstant(SHLAmount, dl, SHVT);
       Op1Promoted =
           DAG.getNode(ISD::SHL, dl, PromotedType, Op1Promoted, ShiftAmount);
       if (!IsShift)
         Op2Promoted =
             DAG.getNode(ISD::SHL, dl, PromotedType, Op2Promoted, ShiftAmount);
 
       SDValue Result =
           DAG.getNode(Opcode, dl, PromotedType, Op1Promoted, Op2Promoted);
       return DAG.getNode(ShiftOp, dl, PromotedType, Result, ShiftAmount);
+    };
+    switch (Opcode) {
+    case ISD::SADDSAT:
+    case ISD::SSUBSAT:
+    case ISD::SSHLSAT:
+      return GetShiftedOp(ISD::SRA);
+    case ISD::UADDSAT:
+    case ISD::USHLSAT:
+      return GetShiftedOp(ISD::SRL);
+    }
   }
 
   if (Opcode == ISD::USUBSAT) {
-    SDValue Max =
-        DAG.getNode(ISD::UMAX, dl, PromotedType, Op1Promoted, Op2Promoted);
-    return DAG.getNode(ISD::SUB, dl, PromotedType, Max, Op2Promoted);
+    assert(DAG.MaskedValueIsZero(Op1Promoted,
+                                 APInt::getBitsSetFrom(NewBits, OldBits)) &&
+           DAG.MaskedValueIsZero(Op2Promoted,
+                                 APInt::getBitsSetFrom(NewBits, OldBits)) &&
+           "Promoted USUBSAT args should be zero-extended");
+    return DAG.getNode(ISD::USUBSAT, dl, PromotedType, Op1Promoted,
+                       Op2Promoted);
   }
 
   if (Opcode == ISD::UADDSAT) {
     APInt MaxVal = APInt::getAllOnesValue(OldBits).zext(NewBits);
     SDValue SatMax = DAG.getConstant(MaxVal, dl, PromotedType);
     SDValue Add =
         DAG.getNode(ISD::ADD, dl, PromotedType, Op1Promoted, Op2Promoted);
     return DAG.getNode(ISD::UMIN, dl, PromotedType, Add, SatMax);

llvm/test/CodeGen/AArch64/usub_sat_vec.ll

 ; CHECK-NEXT:    mov v0.h[1], w10
 ; CHECK-NEXT:    mov v1.h[1], w11
 ; CHECK-NEXT:    ldrb w10, [x0, #3]
 ; CHECK-NEXT:    ldrb w11, [x1, #3]
 ; CHECK-NEXT:    mov v0.h[2], w8
 ; CHECK-NEXT:    mov v1.h[2], w9
 ; CHECK-NEXT:    mov v0.h[3], w10
 ; CHECK-NEXT:    mov v1.h[3], w11
-; CHECK-NEXT:    shl v1.4h, v1.4h, #8
-; CHECK-NEXT:    shl v0.4h, v0.4h, #8
 ; CHECK-NEXT:    uqsub v0.4h, v0.4h, v1.4h
-; CHECK-NEXT:    ushr v0.4h, v0.4h, #8
 ; CHECK-NEXT:    xtn v0.8b, v0.8h
 ; CHECK-NEXT:    str s0, [x2]
 ; CHECK-NEXT:    ret
   %x = load <4 x i8>, <4 x i8>* %px
   %y = load <4 x i8>, <4 x i8>* %py
   %z = call <4 x i8> @llvm.usub.sat.v4i8(<4 x i8> %x, <4 x i8> %y)
   store <4 x i8> %z, <4 x i8>* %pz
   ret void
 }
 
 define void @v2i8(<2 x i8>* %px, <2 x i8>* %py, <2 x i8>* %pz) nounwind {
 ; CHECK-LABEL: v2i8:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ldrb w8, [x0]
 ; CHECK-NEXT:    ldrb w9, [x1]
 ; CHECK-NEXT:    ldrb w10, [x0, #1]
 ; CHECK-NEXT:    ldrb w11, [x1, #1]
 ; CHECK-NEXT:    fmov s0, w8
 ; CHECK-NEXT:    fmov s1, w9
 ; CHECK-NEXT:    mov v0.s[1], w10
 ; CHECK-NEXT:    mov v1.s[1], w11
-; CHECK-NEXT:    shl v1.2s, v1.2s, #24
-; CHECK-NEXT:    shl v0.2s, v0.2s, #24
 ; CHECK-NEXT:    uqsub v0.2s, v0.2s, v1.2s
-; CHECK-NEXT:    ushr v0.2s, v0.2s, #24
 ; CHECK-NEXT:    mov w8, v0.s[1]
 ; CHECK-NEXT:    fmov w9, s0
 ; CHECK-NEXT:    strb w9, [x2]
 ; CHECK-NEXT:    strb w8, [x2, #1]
 ; CHECK-NEXT:    ret
   %x = load <2 x i8>, <2 x i8>* %px
   %y = load <2 x i8>, <2 x i8>* %py
   %z = call <2 x i8> @llvm.usub.sat.v2i8(<2 x i8> %x, <2 x i8> %y)

 ; CHECK-NEXT:    ldrh w8, [x0]
 ; CHECK-NEXT:    ldrh w9, [x1]
 ; CHECK-NEXT:    ldrh w10, [x0, #2]
 ; CHECK-NEXT:    ldrh w11, [x1, #2]
 ; CHECK-NEXT:    fmov s0, w8
 ; CHECK-NEXT:    fmov s1, w9
 ; CHECK-NEXT:    mov v0.s[1], w10
 ; CHECK-NEXT:    mov v1.s[1], w11
-; CHECK-NEXT:    shl v1.2s, v1.2s, #16
-; CHECK-NEXT:    shl v0.2s, v0.2s, #16
 ; CHECK-NEXT:    uqsub v0.2s, v0.2s, v1.2s
-; CHECK-NEXT:    ushr v0.2s, v0.2s, #16
 ; CHECK-NEXT:    mov w8, v0.s[1]
 ; CHECK-NEXT:    fmov w9, s0
 ; CHECK-NEXT:    strh w9, [x2]
 ; CHECK-NEXT:    strh w8, [x2, #2]
 ; CHECK-NEXT:    ret
   %x = load <2 x i16>, <2 x i16>* %px
   %y = load <2 x i16>, <2 x i16>* %py
   %z = call <2 x i16> @llvm.usub.sat.v2i16(<2 x i16> %x, <2 x i16> %y)
; CHECK-NEXT:    ret
   store <1 x i16> %z, <1 x i16>* %pz
   ret void
 }
 
 define <16 x i4> @v16i4(<16 x i4> %x, <16 x i4> %y) nounwind {
 ; CHECK-LABEL: v16i4:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    movi v2.16b, #15
-; CHECK-NEXT:    and v0.16b, v0.16b, v2.16b
 ; CHECK-NEXT:    and v1.16b, v1.16b, v2.16b
-; CHECK-NEXT:    shl v1.16b, v1.16b, #4
-; CHECK-NEXT:    shl v0.16b, v0.16b, #4
+; CHECK-NEXT:    and v0.16b, v0.16b, v2.16b
 ; CHECK-NEXT:    uqsub v0.16b, v0.16b, v1.16b
-; CHECK-NEXT:    ushr v0.16b, v0.16b, #4
 ; CHECK-NEXT:    ret
   %z = call <16 x i4> @llvm.usub.sat.v16i4(<16 x i4> %x, <16 x i4> %y)
   ret <16 x i4> %z
 }
 
 define <16 x i1> @v16i1(<16 x i1> %x, <16 x i1> %y) nounwind {
 ; CHECK-LABEL: v16i1:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    movi v2.16b, #1
-; CHECK-NEXT:    and v0.16b, v0.16b, v2.16b
 ; CHECK-NEXT:    and v1.16b, v1.16b, v2.16b
-; CHECK-NEXT:    shl v1.16b, v1.16b, #7
-; CHECK-NEXT:    shl v0.16b, v0.16b, #7
+; CHECK-NEXT:    and v0.16b, v0.16b, v2.16b
 ; CHECK-NEXT:    uqsub v0.16b, v0.16b, v1.16b
-; CHECK-NEXT:    ushr v0.16b, v0.16b, #7
 ; CHECK-NEXT:    ret
   %z = call <16 x i1> @llvm.usub.sat.v16i1(<16 x i1> %x, <16 x i1> %y)
   ret <16 x i1> %z
 }
 
 define <2 x i32> @v2i32(<2 x i32> %x, <2 x i32> %y) nounwind {
 ; CHECK-LABEL: v2i32:
 ; CHECK:       // %bb.0:

llvm/test/CodeGen/AMDGPU/usubsat.ll

 ; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
 ; GFX6-NEXT:    v_max_u32_e32 v0, v0, v1
 ; GFX6-NEXT:    v_sub_i32_e32 v0, vcc, v0, v1
 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX8-LABEL: v_usubsat_i8:
 ; GFX8:       ; %bb.0:
 ; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
-; GFX8-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX8-NEXT:    v_sub_u16_e64 v0, v0, v1 clamp
-; GFX8-NEXT:    v_lshrrev_b16_e32 v0, 8, v0
+; GFX8-NEXT:    v_sub_u16_sdwa v0, v0, v1 clamp dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_usubsat_i8:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
-; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
-; GFX9-NEXT:    v_sub_u16_e64 v0, v0, v1 clamp
-; GFX9-NEXT:    v_lshrrev_b16_e32 v0, 8, v0
+; GFX9-NEXT:    v_sub_u16_sdwa v0, v0, v1 clamp dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   %result = call i8 @llvm.usub.sat.i8(i8 %lhs, i8 %rhs)
   ret i8 %result
 }
 
 define i16 @v_usubsat_i16(i16 %lhs, i16 %rhs) {
 ; GFX6-LABEL: v_usubsat_i16:
 ; GFX6:       ; %bb.0:

llvm/test/CodeGen/ARM/usub_sat_plus.ll

; CHECK-ARM-NEXT:    bx lr
   %tmp = call i64 @llvm.usub.sat.i64(i64 %x, i64 %z)
   ret i64 %tmp
 }
 
 define zeroext i16 @func16(i16 zeroext %x, i16 zeroext %y, i16 zeroext %z) nounwind {
 ; CHECK-T1-LABEL: func16:
 ; CHECK-T1:       @ %bb.0:
 ; CHECK-T1-NEXT:    muls r1, r2, r1
-; CHECK-T1-NEXT:    uxth r2, r1
-; CHECK-T1-NEXT:    cmp r0, r2
-; CHECK-T1-NEXT:    bhi .LBB2_2
+; CHECK-T1-NEXT:    uxth r1, r1
+; CHECK-T1-NEXT:    subs r0, r0, r1
+; CHECK-T1-NEXT:    bhs .LBB2_2
 ; CHECK-T1-NEXT:  @ %bb.1:
-; CHECK-T1-NEXT:    mov r0, r2
+; CHECK-T1-NEXT:    movs r0, #0
 ; CHECK-T1-NEXT:  .LBB2_2:
-; CHECK-T1-NEXT:    subs r0, r0, r1
 ; CHECK-T1-NEXT:    uxth r0, r0
 ; CHECK-T1-NEXT:    bx lr
 ;
 ; CHECK-T2-LABEL: func16:
 ; CHECK-T2:       @ %bb.0:
-; CHECK-T2-NEXT:    mul r3, r1, r2
-; CHECK-T2-NEXT:    uxth r3, r3
-; CHECK-T2-NEXT:    cmp r0, r3
-; CHECK-T2-NEXT:    it hi
-; CHECK-T2-NEXT:    movhi r3, r0
-; CHECK-T2-NEXT:    mls r0, r1, r2, r3
+; CHECK-T2-NEXT:    muls r1, r2, r1
+; CHECK-T2-NEXT:    uxth r1, r1
+; CHECK-T2-NEXT:    subs r0, r0, r1
+; CHECK-T2-NEXT:    it lo
+; CHECK-T2-NEXT:    movlo r0, #0
 ; CHECK-T2-NEXT:    uxth r0, r0
 ; CHECK-T2-NEXT:    bx lr
 ;
 ; CHECK-ARM-LABEL: func16:
 ; CHECK-ARM:       @ %bb.0:
-; CHECK-ARM-NEXT:    mul r3, r1, r2
-; CHECK-ARM-NEXT:    uxth r3, r3
-; CHECK-ARM-NEXT:    cmp r0, r3
-; CHECK-ARM-NEXT:    movhi r3, r0
-; CHECK-ARM-NEXT:    mls r0, r1, r2, r3
+; CHECK-ARM-NEXT:    mul r1, r1, r2
+; CHECK-ARM-NEXT:    uxth r1, r1
+; CHECK-ARM-NEXT:    subs r0, r0, r1
+; CHECK-ARM-NEXT:    movlo r0, #0
 ; CHECK-ARM-NEXT:    uxth r0, r0
 ; CHECK-ARM-NEXT:    bx lr
   %a = mul i16 %y, %z
   %tmp = call i16 @llvm.usub.sat.i16(i16 %x, i16 %a)
   ret i16 %tmp
 }
 
 define zeroext i8 @func8(i8 zeroext %x, i8 zeroext %y, i8 zeroext %z) nounwind {
 ; CHECK-T1-LABEL: func8:
 ; CHECK-T1:       @ %bb.0:
 ; CHECK-T1-NEXT:    muls r1, r2, r1
-; CHECK-T1-NEXT:    uxtb r2, r1
-; CHECK-T1-NEXT:    cmp r0, r2
-; CHECK-T1-NEXT:    bhi .LBB3_2
+; CHECK-T1-NEXT:    uxtb r1, r1
+; CHECK-T1-NEXT:    subs r0, r0, r1
+; CHECK-T1-NEXT:    bhs .LBB3_2
 ; CHECK-T1-NEXT:  @ %bb.1:
-; CHECK-T1-NEXT:    mov r0, r2
+; CHECK-T1-NEXT:    movs r0, #0
 ; CHECK-T1-NEXT:  .LBB3_2:
-; CHECK-T1-NEXT:    subs r0, r0, r1
 ; CHECK-T1-NEXT:    uxtb r0, r0
 ; CHECK-T1-NEXT:    bx lr
 ;
 ; CHECK-T2-LABEL: func8:
 ; CHECK-T2:       @ %bb.0:
-; CHECK-T2-NEXT:    mul r3, r1, r2
-; CHECK-T2-NEXT:    uxtb r3, r3
-; CHECK-T2-NEXT:    cmp r0, r3
-; CHECK-T2-NEXT:    it hi
-; CHECK-T2-NEXT:    movhi r3, r0
-; CHECK-T2-NEXT:    mls r0, r1, r2, r3
+; CHECK-T2-NEXT:    muls r1, r2, r1
+; CHECK-T2-NEXT:    uxtb r1, r1
+; CHECK-T2-NEXT:    subs r0, r0, r1
+; CHECK-T2-NEXT:    it lo
+; CHECK-T2-NEXT:    movlo r0, #0
 ; CHECK-T2-NEXT:    uxtb r0, r0
 ; CHECK-T2-NEXT:    bx lr
 ;
 ; CHECK-ARM-LABEL: func8:
 ; CHECK-ARM:       @ %bb.0:
-; CHECK-ARM-NEXT:    smulbb r3, r1, r2
-; CHECK-ARM-NEXT:    uxtb r3, r3
-; CHECK-ARM-NEXT:    cmp r0, r3
-; CHECK-ARM-NEXT:    movhi r3, r0
-; CHECK-ARM-NEXT:    mls r0, r1, r2, r3
+; CHECK-ARM-NEXT:    smulbb r1, r1, r2
+; CHECK-ARM-NEXT:    uxtb r1, r1
+; CHECK-ARM-NEXT:    subs r0, r0, r1
+; CHECK-ARM-NEXT:    movlo r0, #0
 ; CHECK-ARM-NEXT:    uxtb r0, r0
 ; CHECK-ARM-NEXT:    bx lr
   %a = mul i8 %y, %z
   %tmp = call i8 @llvm.usub.sat.i8(i8 %x, i8 %a)
   ret i8 %tmp
 }
 
 define zeroext i4 @func4(i4 zeroext %x, i4 zeroext %y, i4 zeroext %z) nounwind {
 ; CHECK-T1-LABEL: func4:
 ; CHECK-T1:       @ %bb.0:
 ; CHECK-T1-NEXT:    muls r1, r2, r1
 ; CHECK-T1-NEXT:    movs r2, #15
-; CHECK-T1-NEXT:    mov r3, r1
-; CHECK-T1-NEXT:    ands r3, r2
-; CHECK-T1-NEXT:    cmp r0, r3
-; CHECK-T1-NEXT:    bhi .LBB4_2
+; CHECK-T1-NEXT:    ands r1, r2
+; CHECK-T1-NEXT:    subs r0, r0, r1
+; CHECK-T1-NEXT:    bhs .LBB4_2
 ; CHECK-T1-NEXT:  @ %bb.1:
-; CHECK-T1-NEXT:    mov r0, r3
+; CHECK-T1-NEXT:    movs r0, #0
 ; CHECK-T1-NEXT:  .LBB4_2:
-; CHECK-T1-NEXT:    subs r0, r0, r1
 ; CHECK-T1-NEXT:    ands r0, r2
 ; CHECK-T1-NEXT:    bx lr
 ;
 ; CHECK-T2-LABEL: func4:
 ; CHECK-T2:       @ %bb.0:
-; CHECK-T2-NEXT:    mul r3, r1, r2
-; CHECK-T2-NEXT:    and r3, r3, #15
-; CHECK-T2-NEXT:    cmp r0, r3
-; CHECK-T2-NEXT:    it hi
-; CHECK-T2-NEXT:    movhi r3, r0
-; CHECK-T2-NEXT:    mls r0, r1, r2, r3
+; CHECK-T2-NEXT:    muls r1, r2, r1
+; CHECK-T2-NEXT:    and r1, r1, #15
+; CHECK-T2-NEXT:    subs r0, r0, r1
+; CHECK-T2-NEXT:    it lo
+; CHECK-T2-NEXT:    movlo r0, #0
 ; CHECK-T2-NEXT:    and r0, r0, #15
 ; CHECK-T2-NEXT:    bx lr
 ;
 ; CHECK-ARM-LABEL: func4:
 ; CHECK-ARM:       @ %bb.0:
-; CHECK-ARM-NEXT:    smulbb r3, r1, r2
-; CHECK-ARM-NEXT:    and r3, r3, #15
-; CHECK-ARM-NEXT:    cmp r0, r3
-; CHECK-ARM-NEXT:    movhi r3, r0
-; CHECK-ARM-NEXT:    mls r0, r1, r2, r3
+; CHECK-ARM-NEXT:    smulbb r1, r1, r2
+; CHECK-ARM-NEXT:    and r1, r1, #15
+; CHECK-ARM-NEXT:    subs r0, r0, r1
+; CHECK-ARM-NEXT:    movlo r0, #0
 ; CHECK-ARM-NEXT:    and r0, r0, #15
 ; CHECK-ARM-NEXT:    bx lr
   %a = mul i4 %y, %z
   %tmp = call i4 @llvm.usub.sat.i4(i4 %x, i4 %a)
   ret i4 %tmp
 }

llvm/test/CodeGen/X86/usub_sat_plus.ll

; X64-NEXT:    retq
   %a = mul i8 %y, %z
   %tmp = call i8 @llvm.usub.sat.i8(i8 %x, i8 %a)
   ret i8 %tmp
 }
 
 define zeroext i4 @func4(i4 zeroext %x, i4 zeroext %y, i4 zeroext %z) nounwind {
 ; X86-LABEL: func4:
 ; X86:       # %bb.0:
-; X86-NEXT:    pushl %ebx
-; X86-NEXT:    pushl %esi
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
 ; X86-NEXT:    mulb {{[0-9]+}}(%esp)
-; X86-NEXT:    movl %eax, %edx
-; X86-NEXT:    andb $15, %dl
-; X86-NEXT:    movzbl %dl, %esi
-; X86-NEXT:    movzbl %cl, %ebx
-; X86-NEXT:    cmpb %dl, %cl
-; X86-NEXT:    cmovbel %esi, %ebx
-; X86-NEXT:    subb %al, %bl
-; X86-NEXT:    movzbl %bl, %eax
+; X86-NEXT:    andb $15, %al
+; X86-NEXT:    xorl %edx, %edx
+; X86-NEXT:    subb %al, %cl
+; X86-NEXT:    movzbl %cl, %eax
+; X86-NEXT:    cmovbl %edx, %eax
 ; X86-NEXT:    andl $15, %eax
-; X86-NEXT:    popl %esi
-; X86-NEXT:    popl %ebx
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: func4:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %esi, %eax
 ; X64-NEXT:    # kill: def $al killed $al killed $eax
 ; X64-NEXT:    mulb %dl
-; X64-NEXT:    movl %eax, %ecx
-; X64-NEXT:    andb $15, %cl
-; X64-NEXT:    movzbl %cl, %ecx
-; X64-NEXT:    cmpb %cl, %dil
-; X64-NEXT:    cmoval %edi, %ecx
-; X64-NEXT:    subb %al, %cl
-; X64-NEXT:    movzbl %cl, %eax
+; X64-NEXT:    andb $15, %al
+; X64-NEXT:    xorl %ecx, %ecx
+; X64-NEXT:    subb %al, %dil
+; X64-NEXT:    movzbl %dil, %eax
+; X64-NEXT:    cmovbl %ecx, %eax
 ; X64-NEXT:    andl $15, %eax
 ; X64-NEXT:    retq
   %a = mul i4 %y, %z
   %tmp = call i4 @llvm.usub.sat.i4(i4 %x, i4 %a)
   ret i4 %tmp
 }

llvm/test/CodeGen/X86/usub_sat_vec.ll

; AVX-NEXT:    retq
   ret void
 }
 
 ; Promotion
 
 define <16 x i4> @v16i4(<16 x i4> %x, <16 x i4> %y) nounwind {
 ; SSE-LABEL: v16i4:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    psllw $4, %xmm1
-; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
+; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; SSE-NEXT:    pand %xmm2, %xmm1
-; SSE-NEXT:    psllw $4, %xmm0
 ; SSE-NEXT:    pand %xmm2, %xmm0
 ; SSE-NEXT:    psubusb %xmm1, %xmm0
-; SSE-NEXT:    psrlw $4, %xmm0
-; SSE-NEXT:    pand {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: v16i4:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpsllw $4, %xmm1, %xmm1
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm2 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
+; AVX-NEXT:    vmovdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX-NEXT:    vpand %xmm2, %xmm1, %xmm1
-; AVX-NEXT:    vpsllw $4, %xmm0, %xmm0
 ; AVX-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX-NEXT:    vpsubusb %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpsrlw $4, %xmm0, %xmm0
-; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %z = call <16 x i4> @llvm.usub.sat.v16i4(<16 x i4> %x, <16 x i4> %y)
   ret <16 x i4> %z
 }
 
 define <16 x i1> @v16i1(<16 x i1> %x, <16 x i1> %y) nounwind {
 ; SSE-LABEL: v16i1:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    psllw $7, %xmm1
-; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
+; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
 ; SSE-NEXT:    pand %xmm2, %xmm1
-; SSE-NEXT:    psllw $7, %xmm0
 ; SSE-NEXT:    pand %xmm2, %xmm0
 ; SSE-NEXT:    psubusb %xmm1, %xmm0
-; SSE-NEXT:    psrlw $7, %xmm0
-; SSE-NEXT:    pand {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: v16i1:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpsllw $7, %xmm1, %xmm1
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
 ; AVX1-NEXT:    vpand %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpsllw $7, %xmm0, %xmm0
 ; AVX1-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpsubusb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpsrlw $7, %xmm0, %xmm0
-; AVX1-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v16i1:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpsllw $7, %xmm1, %xmm1
-; AVX2-NEXT:    vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX2-NEXT:    vmovdqa {{.*#+}} xmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
 ; AVX2-NEXT:    vpand %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vpsllw $7, %xmm0, %xmm0
 ; AVX2-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vpsubusb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vpsrlw $7, %xmm0, %xmm0
-; AVX2-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: v16i1:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; AVX512F-NEXT:    vpslld $31, %zmm0, %zmm0
 ; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm1
 ; AVX512F-NEXT:    vpslld $31, %zmm1, %zmm1