[GlobalISel] Check if ShiftAmt is greater then size of operand

[Acim-Maravic]

Summary

matchCombineShlOfExtend did not check if the size of new shift would be
wider then a size of operand. Current condition did not work if the value
being shifted was zero. Updated to support vector splat.

[arsenm]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1747–1748

Broken for vectors, getScalarSizeInBits

llvm/test/CodeGen/AMDGPU/GlobalISel/combine-shl-from-extend-narrow.postlegal.mir

332

Add a vector version. Also would be good to have one that couldn't be trivially constant folded

[Acim-Maravic]

• added new tests with vector types
• updated to support vector splat

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/sext_inreg.ll

968–971 ↗

(On Diff #525220)

This looks like it was broken before but is differently broken now

1001 ↗

(On Diff #525220)

Ah, this test should be using ashr not shl

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/sext_inreg.ll

1001 ↗

(On Diff #525220)

I've fixed the test so you may need to regenerate this

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

bool CombinerHelper::matchCombineShlOfExtend(MachineInstr &MI,
   Register LHS = MI.getOperand(1).getReg();
 
   Register ExtSrc;
   if (!mi_match(LHS, MRI, m_GAnyExt(m_Reg(ExtSrc))) &&
       !mi_match(LHS, MRI, m_GZExt(m_Reg(ExtSrc))) &&
       !mi_match(LHS, MRI, m_GSExt(m_Reg(ExtSrc))))
     return false;
 
-  // TODO: Should handle vector splat.
   Register RHS = MI.getOperand(2).getReg();
-  auto MaybeShiftAmtVal = getIConstantVRegValWithLookThrough(RHS, MRI);
+  MachineInstr *MIShiftAmt = MRI.getVRegDef(RHS);
+  auto MaybeShiftAmtVal = isConstantOrConstantSplatVector(*MIShiftAmt, MRI);
   if (!MaybeShiftAmtVal)
     return false;
 
   if (LI) {
     LLT SrcTy = MRI.getType(ExtSrc);
 
     // We only really care about the legality with the shifted value. We can
     // pick any type the constant shift amount, so ask the target what to
     // use. Otherwise we would have to guess and hope it is reported as legal.
     LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(SrcTy);
     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_SHL, {SrcTy, ShiftAmtTy}}))
       return false;
   }
 
-  int64_t ShiftAmt = MaybeShiftAmtVal->Value.getSExtValue();
+  int64_t ShiftAmt = MaybeShiftAmtVal->getSExtValue();
   MatchData.Reg = ExtSrc;
   MatchData.Imm = ShiftAmt;
 
   unsigned MinLeadingZeros = KB->getKnownZeroes(ExtSrc).countl_one();
-  return MinLeadingZeros >= ShiftAmt;
+  unsigned SrcTySize = MRI.getType(ExtSrc).getScalarSizeInBits();
+  return MinLeadingZeros >= ShiftAmt && ShiftAmt < SrcTySize;
 }
 
 void CombinerHelper::applyCombineShlOfExtend(MachineInstr &MI,
                                              const RegisterImmPair &MatchData) {
   Register ExtSrcReg = MatchData.Reg;
   int64_t ShiftAmtVal = MatchData.Imm;
 
   LLT ExtSrcTy = MRI.getType(ExtSrcReg);

llvm/test/CodeGen/AArch64/GlobalISel/prelegalizer-combiner-divrem-insertpt-crash.mir

body:             |
   ; CHECK-LABEL: name: divrem_use_before_def
   ; CHECK: bb.0:
   ; CHECK-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
   ; CHECK-NEXT:   liveins: $x0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
   ; CHECK-NEXT:   [[DEF:%[0-9]+]]:_(s1) = G_IMPLICIT_DEF
   ; CHECK-NEXT:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; CHECK-NEXT:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
   ; CHECK-NEXT:   G_BRCOND [[DEF]](s1), %bb.2
   ; CHECK-NEXT:   G_BR %bb.1
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.1:
   ; CHECK-NEXT:   successors: %bb.2(0x80000000)
   ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
-  ; CHECK-NEXT:   [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[C1]](s32)
-  ; CHECK-NEXT:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1
-  ; CHECK-NEXT:   [[FREEZE:%[0-9]+]]:_(s64) = G_FREEZE [[C2]]
+  ; CHECK-NEXT:   [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+  ; CHECK-NEXT:   [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[C2]](s32)
+  ; CHECK-NEXT:   [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1
+  ; CHECK-NEXT:   [[FREEZE:%[0-9]+]]:_(s64) = G_FREEZE [[C3]]
   ; CHECK-NEXT:   [[UDIV:%[0-9]+]]:_(s64) = G_UDIV [[FREEZE]], [[C]]
-  ; CHECK-NEXT:   [[ADD:%[0-9]+]]:_(s64) = G_ADD [[ZEXT]], [[UDIV]]
+  ; CHECK-NEXT:   [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ZEXT]], [[C1]](s64)
+  ; CHECK-NEXT:   [[ADD:%[0-9]+]]:_(s64) = G_ADD [[SHL]], [[UDIV]]
   ; CHECK-NEXT:   G_STORE [[ADD]](s64), [[COPY]](p0) :: (store (s64))
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.2:
   bb.1:
     liveins: $x0
 
     %0:_(p0) = COPY $x0
     %1:_(s1) = G_IMPLICIT_DEF

llvm/test/CodeGen/AMDGPU/GlobalISel/combine-shl-from-extend-narrow.postlegal.mir

bb.0:
     %narrow:_(s16) = G_TRUNC %argument
     %masklow14:_(s16) = G_CONSTANT i16 16383
     %masked:_(s16) = G_AND %narrow, %masklow14
     %extend:_(s64) = G_ZEXT %masked
     %shiftamt:_(s32) = G_CONSTANT i32 2
     %shl:_(s64) = G_SHL %extend, %shiftamt
     $vgpr0_vgpr1 = COPY %shl
 ...
+
+---
+name: do_not_shl_s32_zero_by_16_from_zext_s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0
+    ; GFX6-LABEL: name: do_not_shl_s32_zero_by_16_from_zext_s16
+    ; GFX6: liveins: $vgpr0
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX6-NEXT: %extend:_(s32) = G_ZEXT %zero(s16)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX6-NEXT: $vgpr0 = COPY %shl(s32)
+    ; GFX9-LABEL: name: do_not_shl_s32_zero_by_16_from_zext_s16
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX9-NEXT: %extend:_(s32) = G_ZEXT %zero(s16)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX9-NEXT: $vgpr0 = COPY %shl(s32)
+    %zero:_(s16) = G_CONSTANT i16 0
+    %extend:_(s32) = G_ZEXT %zero:_(s16)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    $vgpr0 = COPY %shl
+...
+
+---
+name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0
+    ; GFX6-LABEL: name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+    ; GFX6: liveins: $vgpr0
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX6-NEXT: %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero(s16), %zero(s16)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt(s16), %shiftamt(s16)
+    ; GFX6-NEXT: %extend:_(<2 x s32>) = G_ZEXT %zerovector(<2 x s16>)
+    ; GFX6-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector(<2 x s16>)
+    ; GFX6-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
+    ; GFX9-LABEL: name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX9-NEXT: %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero(s16), %zero(s16)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt(s16), %shiftamt(s16)
+    ; GFX9-NEXT: %extend:_(<2 x s32>) = G_ZEXT %zerovector(<2 x s16>)
+    ; GFX9-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector(<2 x s16>)
+    ; GFX9-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
+    %zero:_(s16) = G_CONSTANT i16 0
+    %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero, %zero:_(s16)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt, %shiftamt:_(s16)
+    %extend:_(<2 x s32>) = G_ZEXT %zerovector:_(<2 x s16>)
+    %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector
+    $vgpr0_vgpr1 = COPY %shl
+...
+
+---
+name: do_not_shl_s32_by_16_from_zext_s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0
+    ; GFX6-LABEL: name: do_not_shl_s32_by_16_from_zext_s16
+    ; GFX6: liveins: $vgpr0
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %argument:_(s32) = COPY $vgpr0
+    ; GFX6-NEXT: %truncate:_(s16) = G_TRUNC %argument(s32)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %extend:_(s32) = G_ZEXT %truncate(s16)
+    ; GFX6-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX6-NEXT: $vgpr0 = COPY %shl(s32)
+    ; GFX9-LABEL: name: do_not_shl_s32_by_16_from_zext_s16
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %argument:_(s32) = COPY $vgpr0
+    ; GFX9-NEXT: %truncate:_(s16) = G_TRUNC %argument(s32)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %extend:_(s32) = G_ZEXT %truncate(s16)
+    ; GFX9-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX9-NEXT: $vgpr0 = COPY %shl(s32)
+    %argument:_(s32) = COPY $vgpr0
+    %truncate:_(s16) = G_TRUNC %argument:_(s32)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %extend:_(s32) = G_ZEXT %truncate:_(s16)
+    %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    $vgpr0 = COPY %shl
+...

llvm/test/CodeGen/AMDGPU/GlobalISel/combine-shl-from-extend-narrow.prelegal.mir

bb.0:
 
     ; GFX6-LABEL: name: narrow_shl_v2s32_by_2_from_zext_v2s16
     ; GFX6: liveins: $vgpr0
     ; GFX6-NEXT: {{  $}}
     ; GFX6-NEXT: %narrow:_(<2 x s16>) = COPY $vgpr0
     ; GFX6-NEXT: %masklow14:_(s16) = G_CONSTANT i16 16383
     ; GFX6-NEXT: %masklow14vec:_(<2 x s16>) = G_BUILD_VECTOR %masklow14(s16), %masklow14(s16)
     ; GFX6-NEXT: %masked:_(<2 x s16>) = G_AND %narrow, %masklow14vec
-    ; GFX6-NEXT: %extend:_(<2 x s32>) = G_ZEXT %masked(<2 x s16>)
-    ; GFX6-NEXT: %shiftamt:_(s32) = G_CONSTANT i32 2
-    ; GFX6-NEXT: %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt(s32), %shiftamt(s32)
-    ; GFX6-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvec(<2 x s32>)
+    ; GFX6-NEXT: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 2
+    ; GFX6-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16)
+    ; GFX6-NEXT: [[SHL:%[0-9]+]]:_(<2 x s16>) = G_SHL %masked, [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX6-NEXT: %shl:_(<2 x s32>) = G_ZEXT [[SHL]](<2 x s16>)
     ; GFX6-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
     ; GFX9-LABEL: name: narrow_shl_v2s32_by_2_from_zext_v2s16
     ; GFX9: liveins: $vgpr0
     ; GFX9-NEXT: {{  $}}
     ; GFX9-NEXT: %narrow:_(<2 x s16>) = COPY $vgpr0
     ; GFX9-NEXT: %masklow14:_(s16) = G_CONSTANT i16 16383
     ; GFX9-NEXT: %masklow14vec:_(<2 x s16>) = G_BUILD_VECTOR %masklow14(s16), %masklow14(s16)
     ; GFX9-NEXT: %masked:_(<2 x s16>) = G_AND %narrow, %masklow14vec
-    ; GFX9-NEXT: %extend:_(<2 x s32>) = G_ZEXT %masked(<2 x s16>)
-    ; GFX9-NEXT: %shiftamt:_(s32) = G_CONSTANT i32 2
-    ; GFX9-NEXT: %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt(s32), %shiftamt(s32)
-    ; GFX9-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvec(<2 x s32>)
+    ; GFX9-NEXT: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 2
+    ; GFX9-NEXT: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16)
+    ; GFX9-NEXT: [[SHL:%[0-9]+]]:_(<2 x s16>) = G_SHL %masked, [[BUILD_VECTOR]](<2 x s16>)
+    ; GFX9-NEXT: %shl:_(<2 x s32>) = G_ZEXT [[SHL]](<2 x s16>)
     ; GFX9-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
     %narrow:_(<2 x s16>) = COPY $vgpr0
     %masklow14:_(s16) = G_CONSTANT i16 16383
     %masklow14vec:_(<2 x s16>) = G_BUILD_VECTOR %masklow14, %masklow14
     %masked:_(<2 x s16>) = G_AND %narrow, %masklow14vec
     %extend:_(<2 x s32>) = G_ZEXT %masked
     %shiftamt:_(s32) = G_CONSTANT i32 2
     %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt, %shiftamt
bb.0:
 
     ; GFX6-LABEL: name: narrow_shl_v2s64_by_2_from_anyext_v2s32
     ; GFX6: liveins: $vgpr0_vgpr1
     ; GFX6-NEXT: {{  $}}
     ; GFX6-NEXT: %narrow:_(<2 x s32>) = COPY $vgpr0_vgpr1
     ; GFX6-NEXT: %masklow30:_(s32) = G_CONSTANT i32 1073741823
     ; GFX6-NEXT: %masklow30vec:_(<2 x s32>) = G_BUILD_VECTOR %masklow30(s32), %masklow30(s32)
     ; GFX6-NEXT: %masked:_(<2 x s32>) = G_AND %narrow, %masklow30vec
-    ; GFX6-NEXT: %extend:_(<2 x s64>) = G_ANYEXT %masked(<2 x s32>)
     ; GFX6-NEXT: %shiftamt:_(s32) = G_CONSTANT i32 2
     ; GFX6-NEXT: %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt(s32), %shiftamt(s32)
-    ; GFX6-NEXT: %shl:_(<2 x s64>) = G_SHL %extend, %shiftamtvec(<2 x s32>)
+    ; GFX6-NEXT: [[SHL:%[0-9]+]]:_(<2 x s32>) = G_SHL %masked, %shiftamtvec(<2 x s32>)
+    ; GFX6-NEXT: %shl:_(<2 x s64>) = G_ZEXT [[SHL]](<2 x s32>)
     ; GFX6-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %shl(<2 x s64>)
     ; GFX9-LABEL: name: narrow_shl_v2s64_by_2_from_anyext_v2s32
     ; GFX9: liveins: $vgpr0_vgpr1
     ; GFX9-NEXT: {{  $}}
     ; GFX9-NEXT: %narrow:_(<2 x s32>) = COPY $vgpr0_vgpr1
     ; GFX9-NEXT: %masklow30:_(s32) = G_CONSTANT i32 1073741823
     ; GFX9-NEXT: %masklow30vec:_(<2 x s32>) = G_BUILD_VECTOR %masklow30(s32), %masklow30(s32)
     ; GFX9-NEXT: %masked:_(<2 x s32>) = G_AND %narrow, %masklow30vec
-    ; GFX9-NEXT: %extend:_(<2 x s64>) = G_ANYEXT %masked(<2 x s32>)
     ; GFX9-NEXT: %shiftamt:_(s32) = G_CONSTANT i32 2
     ; GFX9-NEXT: %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt(s32), %shiftamt(s32)
-    ; GFX9-NEXT: %shl:_(<2 x s64>) = G_SHL %extend, %shiftamtvec(<2 x s32>)
+    ; GFX9-NEXT: [[SHL:%[0-9]+]]:_(<2 x s32>) = G_SHL %masked, %shiftamtvec(<2 x s32>)
+    ; GFX9-NEXT: %shl:_(<2 x s64>) = G_ZEXT [[SHL]](<2 x s32>)
     ; GFX9-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %shl(<2 x s64>)
     %narrow:_(<2 x s32>) = COPY $vgpr0_vgpr1
     %masklow30:_(s32) = G_CONSTANT i32 1073741823
     %masklow30vec:_(<2 x s32>) = G_BUILD_VECTOR %masklow30, %masklow30
     %masked:_(<2 x s32>) = G_AND %narrow, %masklow30vec
     %extend:_(<2 x s64>) = G_ANYEXT %masked
     %shiftamt:_(s32) = G_CONSTANT i32 2
     %shiftamtvec:_(<2 x s32>) = G_BUILD_VECTOR %shiftamt, %shiftamt
     %shl:_(<2 x s64>) = G_SHL %extend, %shiftamtvec
     $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %shl
 ...
+
+---
+name: do_not_shl_s32_zero_by_16_from_zext_s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0
+    ; GFX6-LABEL: name: do_not_shl_s32_zero_by_16_from_zext_s16
+    ; GFX6: liveins: $vgpr0
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX6-NEXT: %extend:_(s32) = G_ZEXT %zero(s16)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX6-NEXT: $vgpr0 = COPY %shl(s32)
+    ; GFX9-LABEL: name: do_not_shl_s32_zero_by_16_from_zext_s16
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX9-NEXT: %extend:_(s32) = G_ZEXT %zero(s16)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX9-NEXT: $vgpr0 = COPY %shl(s32)
+    %zero:_(s16) = G_CONSTANT i16 0
+    %extend:_(s32) = G_ZEXT %zero:_(s16)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    $vgpr0 = COPY %shl
+...
+
+---
+name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+    ; GFX6-LABEL: name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+    ; GFX6: liveins: $vgpr0, $vgpr1
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX6-NEXT: %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero(s16), %zero(s16)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt(s16), %shiftamt(s16)
+    ; GFX6-NEXT: %extend:_(<2 x s32>) = G_ZEXT %zerovector(<2 x s16>)
+    ; GFX6-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector(<2 x s16>)
+    ; GFX6-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
+    ; GFX9-LABEL: name: do_not_shl_v2s32_zero_by_16_from_zext_v2s16
+    ; GFX9: liveins: $vgpr0, $vgpr1
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %zero:_(s16) = G_CONSTANT i16 0
+    ; GFX9-NEXT: %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero(s16), %zero(s16)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt(s16), %shiftamt(s16)
+    ; GFX9-NEXT: %extend:_(<2 x s32>) = G_ZEXT %zerovector(<2 x s16>)
+    ; GFX9-NEXT: %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector(<2 x s16>)
+    ; GFX9-NEXT: $vgpr0_vgpr1 = COPY %shl(<2 x s32>)
+    %zero:_(s16) = G_CONSTANT i16 0
+    %zerovector:_(<2 x s16>) = G_BUILD_VECTOR %zero, %zero:_(s16)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %shiftamtvector:_(<2 x s16>) = G_BUILD_VECTOR %shiftamt, %shiftamt:_(s16)
+    %extend:_(<2 x s32>) = G_ZEXT %zerovector:_(<2 x s16>)
+    %shl:_(<2 x s32>) = G_SHL %extend, %shiftamtvector
+    $vgpr0_vgpr1 = COPY %shl
+...
+
+---
+name: do_not_shl_s32_by_16_from_zext_s16
+tracksRegLiveness: true
+legalized: true
+
+body:             |
+  bb.0:
+    liveins: $vgpr0
+    ; GFX6-LABEL: name: do_not_shl_s32_by_16_from_zext_s16
+    ; GFX6: liveins: $vgpr0
+    ; GFX6-NEXT: {{  $}}
+    ; GFX6-NEXT: %argument:_(s32) = COPY $vgpr0
+    ; GFX6-NEXT: %truncate:_(s16) = G_TRUNC %argument(s32)
+    ; GFX6-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX6-NEXT: %extend:_(s32) = G_ZEXT %truncate(s16)
+    ; GFX6-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX6-NEXT: $vgpr0 = COPY %shl(s32)
+    ; GFX9-LABEL: name: do_not_shl_s32_by_16_from_zext_s16
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: %argument:_(s32) = COPY $vgpr0
+    ; GFX9-NEXT: %truncate:_(s16) = G_TRUNC %argument(s32)
+    ; GFX9-NEXT: %shiftamt:_(s16) = G_CONSTANT i16 16
+    ; GFX9-NEXT: %extend:_(s32) = G_ZEXT %truncate(s16)
+    ; GFX9-NEXT: %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    ; GFX9-NEXT: $vgpr0 = COPY %shl(s32)
+    %argument:_(s32) = COPY $vgpr0
+    %truncate:_(s16) = G_TRUNC %argument:_(s32)
+    %shiftamt:_(s16) = G_CONSTANT i16 16
+    %extend:_(s32) = G_ZEXT %truncate:_(s16)
+    %shl:_(s32) = G_SHL %extend, %shiftamt(s16)
+    $vgpr0 = COPY %shl
+...

llvm/test/CodeGen/AMDGPU/GlobalISel/shl-ext-reduce.ll

 }
 
 define amdgpu_ps <2 x i64> @s_shl_v2i64_zext_v2i32(<2 x i32> inreg %x) {
 ; GCN-LABEL: s_shl_v2i64_zext_v2i32:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_brev_b32 s2, -4
 ; GCN-NEXT:    s_mov_b32 s3, s2
 ; GCN-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
+; GCN-NEXT:    s_lshl_b32 s0, s0, 2
+; GCN-NEXT:    s_lshl_b32 s2, s1, 2
+; GCN-NEXT:    s_mov_b32 s1, 0
 ; GCN-NEXT:    s_mov_b32 s3, 0
-; GCN-NEXT:    s_mov_b32 s2, s0
-; GCN-NEXT:    s_mov_b32 s4, s1
-; GCN-NEXT:    s_mov_b32 s5, s3
-; GCN-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN-NEXT:    s_lshl_b64 s[2:3], s[4:5], 2
 ; GCN-NEXT:    ; return to shader part epilog
 ;
 ; GFX10PLUS-LABEL: s_shl_v2i64_zext_v2i32:
 ; GFX10PLUS:       ; %bb.0:
 ; GFX10PLUS-NEXT:    s_brev_b32 s2, -4
 ; GFX10PLUS-NEXT:    s_mov_b32 s3, s2
 ; GFX10PLUS-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
 ; GFX10PLUS-NEXT:    s_mov_b32 s3, 0
-; GFX10PLUS-NEXT:    s_mov_b32 s2, s0
-; GFX10PLUS-NEXT:    s_mov_b32 s4, s1
-; GFX10PLUS-NEXT:    s_mov_b32 s5, s3
-; GFX10PLUS-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX10PLUS-NEXT:    s_lshl_b64 s[2:3], s[4:5], 2
+; GFX10PLUS-NEXT:    s_lshl_b32 s0, s0, 2
+; GFX10PLUS-NEXT:    s_lshl_b32 s2, s1, 2
+; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
 ; GFX10PLUS-NEXT:    ; return to shader part epilog
   %and = and <2 x i32> %x, <i32 1073741823, i32 1073741823>
   %ext = zext <2 x i32> %and to <2 x i64>
   %shl = shl <2 x i64> %ext, <i64 2, i64 2>
   ret <2 x i64> %shl
 }
 
 define <2 x i64> @v_shl_v2i64_zext_v2i32(<2 x i32> %x) {
-; GFX7-LABEL: v_shl_v2i64_zext_v2i32:
-; GFX7:       ; %bb.0:
-; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT:    v_and_b32_e32 v2, 0x3fffffff, v1
-; GFX7-NEXT:    v_mov_b32_e32 v1, 0
-; GFX7-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
-; GFX7-NEXT:    v_mov_b32_e32 v3, v1
-; GFX7-NEXT:    v_lshl_b64 v[0:1], v[0:1], 2
-; GFX7-NEXT:    v_lshl_b64 v[2:3], v[2:3], 2
-; GFX7-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: v_shl_v2i64_zext_v2i32:
-; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT:    v_and_b32_e32 v2, 0x3fffffff, v1
-; GFX8-NEXT:    v_mov_b32_e32 v1, 0
-; GFX8-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
-; GFX8-NEXT:    v_mov_b32_e32 v3, v1
-; GFX8-NEXT:    v_lshlrev_b64 v[0:1], 2, v[0:1]
-; GFX8-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX8-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: v_shl_v2i64_zext_v2i32:
-; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_and_b32_e32 v2, 0x3fffffff, v1
-; GFX9-NEXT:    v_mov_b32_e32 v1, 0
-; GFX9-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
-; GFX9-NEXT:    v_mov_b32_e32 v3, v1
-; GFX9-NEXT:    v_lshlrev_b64 v[0:1], 2, v[0:1]
-; GFX9-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX9-NEXT:    s_setpc_b64 s[30:31]
+; GCN-LABEL: v_shl_v2i64_zext_v2i32:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
+; GCN-NEXT:    v_and_b32_e32 v1, 0x3fffffff, v1
+; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GCN-NEXT:    v_lshlrev_b32_e32 v2, 2, v1
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    v_mov_b32_e32 v3, 0
+; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: v_shl_v2i64_zext_v2i32:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
+; GFX10-NEXT:    v_and_b32_e32 v1, 0x3fffffff, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v3, 0
-; GFX10-NEXT:    v_and_b32_e32 v2, 0x3fffffff, v0
-; GFX10-NEXT:    v_and_b32_e32 v4, 0x3fffffff, v1
-; GFX10-NEXT:    v_mov_b32_e32 v5, v3
-; GFX10-NEXT:    v_lshlrev_b64 v[0:1], 2, v[2:3]
-; GFX10-NEXT:    v_lshlrev_b64 v[2:3], 2, v[4:5]
+; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 2, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: v_shl_v2i64_zext_v2i32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_and_b32 v2, 0x3fffffff, v0
-; GFX11-NEXT:    v_dual_mov_b32 v5, v3 :: v_dual_and_b32 v4, 0x3fffffff, v1
-; GFX11-NEXT:    v_lshlrev_b64 v[0:1], 2, v[2:3]
-; GFX11-NEXT:    v_lshlrev_b64 v[2:3], 2, v[4:5]
+; GFX11-NEXT:    v_and_b32_e32 v0, 0x3fffffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0x3fffffff, v1
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v2, 2, v1
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %and = and <2 x i32> %x, <i32 1073741823, i32 1073741823>
   %ext = zext <2 x i32> %and to <2 x i64>
   %shl = shl <2 x i64> %ext, <i64 2, i64 2>
   ret <2 x i64> %shl
 }
 
 define amdgpu_ps <2 x i64> @s_shl_v2i64_sext_v2i32(<2 x i32> inreg %x) {
 ; GCN-LABEL: s_shl_v2i64_sext_v2i32:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_brev_b32 s2, -8
 ; GCN-NEXT:    s_mov_b32 s3, s2
 ; GCN-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
-; GCN-NEXT:    s_ashr_i32 s3, s0, 31
-; GCN-NEXT:    s_mov_b32 s2, s0
-; GCN-NEXT:    s_ashr_i32 s5, s1, 31
-; GCN-NEXT:    s_mov_b32 s4, s1
-; GCN-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN-NEXT:    s_lshl_b64 s[2:3], s[4:5], 2
+; GCN-NEXT:    s_lshl_b32 s0, s0, 2
+; GCN-NEXT:    s_lshl_b32 s2, s1, 2
+; GCN-NEXT:    s_mov_b32 s1, 0
+; GCN-NEXT:    s_mov_b32 s3, 0
 ; GCN-NEXT:    ; return to shader part epilog
 ;
 ; GFX10PLUS-LABEL: s_shl_v2i64_sext_v2i32:
 ; GFX10PLUS:       ; %bb.0:
 ; GFX10PLUS-NEXT:    s_brev_b32 s2, -8
 ; GFX10PLUS-NEXT:    s_mov_b32 s3, s2
 ; GFX10PLUS-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
-; GFX10PLUS-NEXT:    s_ashr_i32 s3, s0, 31
-; GFX10PLUS-NEXT:    s_mov_b32 s2, s0
-; GFX10PLUS-NEXT:    s_ashr_i32 s5, s1, 31
-; GFX10PLUS-NEXT:    s_mov_b32 s4, s1
-; GFX10PLUS-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX10PLUS-NEXT:    s_lshl_b64 s[2:3], s[4:5], 2
+; GFX10PLUS-NEXT:    s_mov_b32 s3, 0
+; GFX10PLUS-NEXT:    s_lshl_b32 s0, s0, 2
+; GFX10PLUS-NEXT:    s_lshl_b32 s2, s1, 2
+; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
 ; GFX10PLUS-NEXT:    ; return to shader part epilog
   %and = and <2 x i32> %x, <i32 536870911, i32 536870911>
   %ext = sext <2 x i32> %and to <2 x i64>
   %shl = shl <2 x i64> %ext, <i64 2, i64 2>
   ret <2 x i64> %shl
 }
 
 define <2 x i64> @v_shl_v2i64_sext_v2i32(<2 x i32> %x) {
-; GFX7-LABEL: v_shl_v2i64_sext_v2i32:
-; GFX7:       ; %bb.0:
-; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
-; GFX7-NEXT:    v_and_b32_e32 v2, 0x1fffffff, v1
-; GFX7-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GFX7-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
-; GFX7-NEXT:    v_lshl_b64 v[0:1], v[0:1], 2
-; GFX7-NEXT:    v_lshl_b64 v[2:3], v[2:3], 2
-; GFX7-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: v_shl_v2i64_sext_v2i32:
-; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
-; GFX8-NEXT:    v_and_b32_e32 v2, 0x1fffffff, v1
-; GFX8-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GFX8-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
-; GFX8-NEXT:    v_lshlrev_b64 v[0:1], 2, v[0:1]
-; GFX8-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX8-NEXT:    s_setpc_b64 s[30:31]
+; GCN-LABEL: v_shl_v2i64_sext_v2i32:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
+; GCN-NEXT:    v_and_b32_e32 v1, 0x1fffffff, v1
+; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GCN-NEXT:    v_lshlrev_b32_e32 v2, 2, v1
+; GCN-NEXT:    v_mov_b32_e32 v1, 0
+; GCN-NEXT:    v_mov_b32_e32 v3, 0
+; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX9-LABEL: v_shl_v2i64_sext_v2i32:
-; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
-; GFX9-NEXT:    v_and_b32_e32 v2, 0x1fffffff, v1
-; GFX9-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GFX9-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
-; GFX9-NEXT:    v_lshlrev_b64 v[0:1], 2, v[0:1]
-; GFX9-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX9-NEXT:    s_setpc_b64 s[30:31]
+; GFX10-LABEL: v_shl_v2i64_sext_v2i32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
+; GFX10-NEXT:    v_and_b32_e32 v1, 0x1fffffff, v1
+; GFX10-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 2, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX10PLUS-LABEL: v_shl_v2i64_sext_v2i32:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10PLUS-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v2, 0x1fffffff, v1
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v1, 31, v0
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v3, 31, v2
-; GFX10PLUS-NEXT:    v_lshlrev_b64 v[0:1], 2, v[0:1]
-; GFX10PLUS-NEXT:    v_lshlrev_b64 v[2:3], 2, v[2:3]
-; GFX10PLUS-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: v_shl_v2i64_sext_v2i32:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0x1fffffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0x1fffffff, v1
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v2, 2, v1
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %and = and <2 x i32> %x, <i32 536870911, i32 536870911>
   %ext = sext <2 x i32> %and to <2 x i64>
   %shl = shl <2 x i64> %ext, <i64 2, i64 2>
   ret <2 x i64> %shl
 }
 
 define amdgpu_ps i32 @s_shl_i32_zext_i16(i16 inreg %x) {
 ; GFX7-LABEL: s_shl_i32_zext_i16:

 define amdgpu_ps <2 x i32> @s_shl_v2i32_zext_v2i16(<2 x i16> inreg %x) {
 ; GFX7-LABEL: s_shl_v2i32_zext_v2i16:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX7-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX7-NEXT:    s_or_b32 s0, s1, s0
 ; GFX7-NEXT:    s_and_b32 s0, s0, 0x3fff3fff
 ; GFX7-NEXT:    s_lshr_b32 s1, s0, 16
-; GFX7-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX7-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX7-NEXT:    s_lshl_b32 s1, s1, 2
+; GFX7-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX7-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX7-NEXT:    ; return to shader part epilog
 ;
 ; GFX8-LABEL: s_shl_v2i32_zext_v2i16:
 ; GFX8:       ; %bb.0:
 ; GFX8-NEXT:    s_movk_i32 s2, 0x3fff
 ; GFX8-NEXT:    s_mov_b32 s3, s2
 ; GFX8-NEXT:    s_lshr_b32 s1, s0, 16
 ; GFX8-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX8-NEXT:    s_and_b64 s[0:1], s[0:1], s[2:3]
 ; GFX8-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX8-NEXT:    s_lshl_b32 s1, s1, 2
 ; GFX8-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_shl_v2i32_zext_v2i16:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_and_b32 s0, s0, 0x3fff3fff
 ; GFX9-NEXT:    s_lshr_b32 s1, s0, 16
-; GFX9-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX9-NEXT:    s_lshl_b32 s0, s0, 2
+; GFX9-NEXT:    s_lshl_b32 s0, s0, 0x20002
 ; GFX9-NEXT:    s_lshl_b32 s1, s1, 2
+; GFX9-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX9-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX9-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10PLUS-LABEL: s_shl_v2i32_zext_v2i16:
 ; GFX10PLUS:       ; %bb.0:
 ; GFX10PLUS-NEXT:    s_and_b32 s0, s0, 0x3fff3fff
-; GFX10PLUS-NEXT:    s_and_b32 s1, s0, 0xffff
-; GFX10PLUS-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX10PLUS-NEXT:    s_lshl_b32 s0, s1, 2
-; GFX10PLUS-NEXT:    s_lshl_b32 s1, s2, 2
+; GFX10PLUS-NEXT:    s_lshr_b32 s1, s0, 16
+; GFX10PLUS-NEXT:    s_lshl_b32 s0, s0, 0x20002
+; GFX10PLUS-NEXT:    s_lshl_b32 s1, s1, 2
+; GFX10PLUS-NEXT:    s_pack_ll_b32_b16 s1, s0, s1
+; GFX10PLUS-NEXT:    s_and_b32 s0, s1, 0xffff
+; GFX10PLUS-NEXT:    s_lshr_b32 s1, s1, 16
 ; GFX10PLUS-NEXT:    ; return to shader part epilog
   %and = and <2 x i16> %x, <i16 16383, i16 16383>
   %ext = zext <2 x i16> %and to <2 x i32>
   %shl = shl <2 x i32> %ext, <i32 2, i32 2>
   ret <2 x i32> %shl
 }
 
 ; FIXME: This doesn't do what we want. The pre-legalizer combiner
 ; fails to handle the vector splat. The post-legalizer sees the zext
 ; legalized into the and. This is probably not that important, since
 ; we really do this combine in the machine level for lowered
 ; getelementptrs.
 define <2 x i32> @v_shl_v2i32_zext_v2i16(<2 x i16> %x) {
 ; GFX7-LABEL: v_shl_v2i32_zext_v2i16:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX7-NEXT:    v_or_b32_e32 v0, v1, v0
 ; GFX7-NEXT:    v_and_b32_e32 v0, 0x3fff3fff, v0
 ; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
-; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
+; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX8-LABEL: v_shl_v2i32_zext_v2i16:
 ; GFX8:       ; %bb.0:
 ; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX8-NEXT:    v_and_b32_e32 v1, 0x3fff3fff, v0
 ; GFX8-NEXT:    v_mov_b32_e32 v2, 2
-; GFX8-NEXT:    v_lshlrev_b32_sdwa v0, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; GFX8-NEXT:    v_lshlrev_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT:    v_lshlrev_b16_e32 v0, 2, v1
+; GFX8-NEXT:    v_lshlrev_b16_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: v_shl_v2i32_zext_v2i16:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s4, 2
-; GFX9-NEXT:    v_and_b32_e32 v1, 0x3fff3fff, v0
-; GFX9-NEXT:    v_lshlrev_b32_sdwa v0, s4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; GFX9-NEXT:    v_lshlrev_b32_sdwa v1, s4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    v_and_b32_e32 v0, 0x3fff3fff, v0
+; GFX9-NEXT:    v_pk_lshlrev_b16 v0, 2, v0 op_sel_hi:[0,1]
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
+; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX10-LABEL: v_shl_v2i32_zext_v2i16:
-; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    v_and_b32_e32 v1, 0x3fff3fff, v0
-; GFX10-NEXT:    s_mov_b32 s4, 2
-; GFX10-NEXT:    v_lshlrev_b32_sdwa v0, s4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
-; GFX10-NEXT:    v_lshlrev_b32_sdwa v1, s4, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX10-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: v_shl_v2i32_zext_v2i16:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_and_b32_e32 v0, 0x3fff3fff, v0
-; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v0
-; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v1
-; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 2, v2
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX10PLUS-LABEL: v_shl_v2i32_zext_v2i16:
+; GFX10PLUS:       ; %bb.0:
+; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10PLUS-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 0x3fff3fff, v0
+; GFX10PLUS-NEXT:    v_pk_lshlrev_b16 v1, 2, v0 op_sel_hi:[0,1]
+; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX10PLUS-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX10PLUS-NEXT:    s_setpc_b64 s[30:31]
   %and = and <2 x i16> %x, <i16 16383, i16 16383>
   %ext = zext <2 x i16> %and to <2 x i32>
   %shl = shl <2 x i32> %ext, <i32 2, i32 2>
   ret <2 x i32> %shl
 }
 
 declare i32 @llvm.amdgcn.workitem.id.x() #0
 
 attributes #0 = { nounwind readnone speculatable willreturn }