[AMDGPU] Support horizontal vectorization.

Author: FarhanaAleen

Reviewed By: rampitec, arsenm

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D46213

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@331313 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index bf560a9..d9bee12 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -468,6 +468,21 @@ unsigned AMDGPUTTIImpl::getCFInstrCost(unsigned Opcode) {
   }
 }
 
+int AMDGPUTTIImpl::getArithmeticReductionCost(unsigned Opcode, Type *Ty,
+                                              bool IsPairwise) {
+  EVT OrigTy = TLI->getValueType(DL, Ty);
+
+  // Computes cost on targets that have packed math instructions(which support
+  // 16-bit types only).
+  if (IsPairwise ||
+      !ST->hasVOP3PInsts() ||
+      OrigTy.getScalarSizeInBits() != 16)
+    return BaseT::getArithmeticReductionCost(Opcode, Ty, IsPairwise);
+
+  std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
+  return LT.first * getFullRateInstrCost();
+}
+
 int AMDGPUTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
                                       unsigned Index) {
   switch (Opcode) {

diff --git a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
index a112757..744093d 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
+++ b/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h
@@ -172,6 +172,10 @@ public:
                            const Function *Callee) const;
 
   unsigned getInliningThresholdMultiplier() { return 9; }
+
+  int getArithmeticReductionCost(unsigned Opcode,
+                                 Type *Ty,
+                                 bool IsPairwise);
 };
 
 } // end namespace llvm

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index 4186e22..1ed6dd9 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6616,6 +6616,30 @@ SDValue SITargetLowering::performExtractVectorEltCombine(
     return DAG.getNode(Vec.getOpcode(), SL, EltVT, Elt);
   }
 
+  // ScalarRes = EXTRACT_VECTOR_ELT ((vector-BINOP Vec1, Vec2), Idx)
+  //    =>
+  // Vec1Elt = EXTRACT_VECTOR_ELT(Vec1, Idx)
+  // Vec2Elt = EXTRACT_VECTOR_ELT(Vec2, Idx)
+  // ScalarRes = scalar-BINOP Vec1Elt, Vec2Elt
+  if (Vec.hasOneUse()) {
+    SDLoc SL(N);
+    EVT EltVT = N->getValueType(0);
+    SDValue Idx = N->getOperand(1);
+    unsigned Opc = Vec.getOpcode();
+
+    switch(Opc) {
+    default:
+      return SDValue();
+      // TODO: Support other binary operations.
+    case ISD::FADD:
+    case ISD::ADD:
+      return DAG.getNode(Opc, SL, EltVT,
+                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
+                                     Vec.getOperand(0), Idx),
+                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
+                                     Vec.getOperand(1), Idx));
+    }
+  }
   return SDValue();
 }
 

diff --git a/test/CodeGen/AMDGPU/reduction.ll b/test/CodeGen/AMDGPU/reduction.ll
new file mode 100644 (file)
index 0000000..4db7375
--- /dev/null
+++ b/test/CodeGen/AMDGPU/reduction.ll
@@ -0,0 +1,128 @@
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: llc -march=amdgcn -mcpu=fiji  -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,VI %s
+
+; GCN-LABEL: {{^}}reduction_half4:
+; GFX9:      v_pk_add_f16 [[ADD:v[0-9]+]], [[ADD:v[0-9]+]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_add_f16_sdwa [[ADD]], [[ADD]], [[ADD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; VI:      v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+define half @reduction_half4(<4 x half> %vec4) {
+entry:
+  %rdx.shuf = shufflevector <4 x half> %vec4, <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %bin.rdx = fadd fast <4 x half> %vec4, %rdx.shuf
+  %rdx.shuf1 = shufflevector <4 x half> %bin.rdx, <4 x half> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+  %bin.rdx2 = fadd fast <4 x half> %bin.rdx, %rdx.shuf1
+  %res = extractelement <4 x half> %bin.rdx2, i32 0
+  ret half %res
+}
+
+; GCN-LABEL: {{^}}reduction_v4i16:
+; GFX9:      v_pk_add_u16 [[ADD:v[0-9]+]], [[ADD:v[0-9]+]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_add_u16_sdwa [[ADD]], [[ADD]], [[ADD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; VI:      v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+define i16 @reduction_v4i16(<4 x i16> %vec4) {
+entry:
+  %rdx.shuf = shufflevector <4 x i16> %vec4, <4 x i16> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %bin.rdx = add <4 x i16> %vec4, %rdx.shuf
+  %rdx.shuf1 = shufflevector <4 x i16> %bin.rdx, <4 x i16> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+  %bin.rdx2 = add <4 x i16> %bin.rdx, %rdx.shuf1
+  %res = extractelement <4 x i16> %bin.rdx2, i32 0
+  ret i16 %res
+}
+
+; GCN-LABEL: {{^}}reduction_half8:
+; GFX9:      v_pk_add_f16 [[ADD1:v[0-9]+]], [[ADD1:v[0-9]+]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 [[ADD:v[0-9]+]], [[ADD:v[0-9]+]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 [[ADD:v[0-9]+]], [[ADD]], [[ADD1]]{{$}}
+; GFX9-NEXT: v_add_f16_sdwa [[ADD]], [[ADD]], [[ADD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; VI:      v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+
+define half @reduction_half8(<8 x half> %vec8) {
+entry:
+  %rdx.shuf = shufflevector <8 x half> %vec8, <8 x half> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx = fadd fast <8 x half> %vec8, %rdx.shuf
+  %rdx.shuf1 = shufflevector <8 x half> %bin.rdx, <8 x half> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx2 = fadd fast <8 x half> %bin.rdx, %rdx.shuf1
+  %rdx.shuf3 = shufflevector <8 x half> %bin.rdx2, <8 x half> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx4 = fadd fast <8 x half> %bin.rdx2, %rdx.shuf3
+  %res = extractelement <8 x half> %bin.rdx4, i32 0
+  ret half %res
+}
+
+; GCN-LABEL: {{^}}reduction_v8i16:
+; GFX9:      v_pk_add_u16 [[ADD1]], [[ADD1:v[0-9]+]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_u16 [[ADD]], [[ADD]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_u16 [[ADD]], [[ADD]], [[ADD1]]{{$}}
+; GFX9-NEXT: v_add_u16_sdwa [[ADD]], [[ADD]], [[ADD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; VI:      v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+; VI-NEXT: v_add_u16_e32
+
+define i16 @reduction_v8i16(<8 x i16> %vec8) {
+entry:
+  %rdx.shuf = shufflevector <8 x i16> %vec8, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx = add <8 x i16> %vec8, %rdx.shuf
+  %rdx.shuf1 = shufflevector <8 x i16> %bin.rdx, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx2 = add <8 x i16> %bin.rdx, %rdx.shuf1
+  %rdx.shuf3 = shufflevector <8 x i16> %bin.rdx2, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx4 = add <8 x i16> %bin.rdx2, %rdx.shuf3
+  %res = extractelement <8 x i16> %bin.rdx4, i32 0
+  ret i16 %res
+}
+
+; GCN-LABEL: {{^}}reduction_half16:
+; GFX9:      v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; GFX9:      v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 [[ADD1]], [[ADD1]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 [[ADD]], [[ADD]], v{{[0-9]+}}{{$}}
+; GFX9-NEXT: v_pk_add_f16 [[ADD]], [[ADD]], [[ADD1]]{{$}}
+; GFX9-NEXT: v_add_f16_sdwa [[ADD]], [[ADD]], [[ADD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; VI:      v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+; VI-NEXT: v_add_f16_e32
+
+define half @reduction_half16(<16 x half> %vec16) {
+entry:
+  %rdx.shuf = shufflevector <16 x half> %vec16, <16 x half> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx = fadd fast <16 x half> %vec16, %rdx.shuf
+  %rdx.shuf1 = shufflevector <16 x half> %bin.rdx, <16 x half> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx2 = fadd fast <16 x half> %bin.rdx, %rdx.shuf1
+  %rdx.shuf3 = shufflevector <16 x half> %bin.rdx2, <16 x half> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx4 = fadd fast <16 x half> %bin.rdx2, %rdx.shuf3
+  %rdx.shuf5 = shufflevector <16 x half> %bin.rdx4, <16 x half> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx6 = fadd fast <16 x half> %bin.rdx4, %rdx.shuf5
+  %res = extractelement <16 x half> %bin.rdx6, i32 0
+  ret half %res
+}
\ No newline at end of file

diff --git a/test/Transforms/SLPVectorizer/AMDGPU/reduction.ll b/test/Transforms/SLPVectorizer/AMDGPU/reduction.ll
new file mode 100644 (file)
index 0000000..7d64bf3
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/AMDGPU/reduction.ll
@@ -0,0 +1,346 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -slp-vectorizer -dce < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -slp-vectorizer -dce < %s | FileCheck -check-prefixes=GCN,VI %s
+
+define half @reduction_half4(<4 x half> %a) {
+; GFX9-LABEL: @reduction_half4(
+; GFX9-NEXT:  entry:
+; GFX9-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <4 x half> [[A:%.*]], <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX:%.*]] = fadd fast <4 x half> [[A]], [[RDX_SHUF]]
+; GFX9-NEXT:    [[RDX_SHUF1:%.*]] = shufflevector <4 x half> [[BIN_RDX]], <4 x half> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX2:%.*]] = fadd fast <4 x half> [[BIN_RDX]], [[RDX_SHUF1]]
+; GFX9-NEXT:    [[TMP0:%.*]] = extractelement <4 x half> [[BIN_RDX2]], i32 0
+; GFX9-NEXT:    ret half [[TMP0]]
+;
+; VI-LABEL: @reduction_half4(
+; VI-NEXT:  entry:
+; VI-NEXT:    [[ELT0:%.*]] = extractelement <4 x half> [[A:%.*]], i64 0
+; VI-NEXT:    [[ELT1:%.*]] = extractelement <4 x half> [[A]], i64 1
+; VI-NEXT:    [[ELT2:%.*]] = extractelement <4 x half> [[A]], i64 2
+; VI-NEXT:    [[ELT3:%.*]] = extractelement <4 x half> [[A]], i64 3
+; VI-NEXT:    [[ADD1:%.*]] = fadd fast half [[ELT1]], [[ELT0]]
+; VI-NEXT:    [[ADD2:%.*]] = fadd fast half [[ELT2]], [[ADD1]]
+; VI-NEXT:    [[ADD3:%.*]] = fadd fast half [[ELT3]], [[ADD2]]
+; VI-NEXT:    ret half [[ADD3]]
+;
+entry:
+  %elt0 = extractelement <4 x half> %a, i64 0
+  %elt1 = extractelement <4 x half> %a, i64 1
+  %elt2 = extractelement <4 x half> %a, i64 2
+  %elt3 = extractelement <4 x half> %a, i64 3
+
+  %add1 = fadd fast half %elt1, %elt0
+  %add2 = fadd fast half %elt2, %add1
+  %add3 = fadd fast half %elt3, %add2
+
+  ret half %add3
+}
+
+define half @reduction_half8(<8 x half> %vec8) {
+; GFX9-LABEL: @reduction_half8(
+; GFX9-NEXT:  entry:
+; GFX9-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <8 x half> [[VEC8:%.*]], <8 x half> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX:%.*]] = fadd fast <8 x half> [[VEC8]], [[RDX_SHUF]]
+; GFX9-NEXT:    [[RDX_SHUF1:%.*]] = shufflevector <8 x half> [[BIN_RDX]], <8 x half> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX2:%.*]] = fadd fast <8 x half> [[BIN_RDX]], [[RDX_SHUF1]]
+; GFX9-NEXT:    [[RDX_SHUF3:%.*]] = shufflevector <8 x half> [[BIN_RDX2]], <8 x half> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX4:%.*]] = fadd fast <8 x half> [[BIN_RDX2]], [[RDX_SHUF3]]
+; GFX9-NEXT:    [[TMP0:%.*]] = extractelement <8 x half> [[BIN_RDX4]], i32 0
+; GFX9-NEXT:    ret half [[TMP0]]
+;
+; VI-LABEL: @reduction_half8(
+; VI-NEXT:  entry:
+; VI-NEXT:    [[ELT0:%.*]] = extractelement <8 x half> [[VEC8:%.*]], i64 0
+; VI-NEXT:    [[ELT1:%.*]] = extractelement <8 x half> [[VEC8]], i64 1
+; VI-NEXT:    [[ELT2:%.*]] = extractelement <8 x half> [[VEC8]], i64 2
+; VI-NEXT:    [[ELT3:%.*]] = extractelement <8 x half> [[VEC8]], i64 3
+; VI-NEXT:    [[ELT4:%.*]] = extractelement <8 x half> [[VEC8]], i64 4
+; VI-NEXT:    [[ELT5:%.*]] = extractelement <8 x half> [[VEC8]], i64 5
+; VI-NEXT:    [[ELT6:%.*]] = extractelement <8 x half> [[VEC8]], i64 6
+; VI-NEXT:    [[ELT7:%.*]] = extractelement <8 x half> [[VEC8]], i64 7
+; VI-NEXT:    [[ADD1:%.*]] = fadd fast half [[ELT1]], [[ELT0]]
+; VI-NEXT:    [[ADD2:%.*]] = fadd fast half [[ELT2]], [[ADD1]]
+; VI-NEXT:    [[ADD3:%.*]] = fadd fast half [[ELT3]], [[ADD2]]
+; VI-NEXT:    [[ADD4:%.*]] = fadd fast half [[ELT4]], [[ADD3]]
+; VI-NEXT:    [[ADD5:%.*]] = fadd fast half [[ELT5]], [[ADD4]]
+; VI-NEXT:    [[ADD6:%.*]] = fadd fast half [[ELT6]], [[ADD5]]
+; VI-NEXT:    [[ADD7:%.*]] = fadd fast half [[ELT7]], [[ADD6]]
+; VI-NEXT:    ret half [[ADD7]]
+;
+entry:
+  %elt0 = extractelement <8 x half> %vec8, i64 0
+  %elt1 = extractelement <8 x half> %vec8, i64 1
+  %elt2 = extractelement <8 x half> %vec8, i64 2
+  %elt3 = extractelement <8 x half> %vec8, i64 3
+  %elt4 = extractelement <8 x half> %vec8, i64 4
+  %elt5 = extractelement <8 x half> %vec8, i64 5
+  %elt6 = extractelement <8 x half> %vec8, i64 6
+  %elt7 = extractelement <8 x half> %vec8, i64 7
+
+  %add1 = fadd fast half %elt1, %elt0
+  %add2 = fadd fast half %elt2, %add1
+  %add3 = fadd fast half %elt3, %add2
+  %add4 = fadd fast half %elt4, %add3
+  %add5 = fadd fast half %elt5, %add4
+  %add6 = fadd fast half %elt6, %add5
+  %add7 = fadd fast half %elt7, %add6
+
+  ret half %add7
+}
+
+define half @reduction_half16(<16 x half> %vec16) {
+; GFX9-LABEL: @reduction_half16(
+; GFX9-NEXT:  entry:
+; GFX9-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <16 x half> [[VEC16:%.*]], <16 x half> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX:%.*]] = fadd fast <16 x half> [[VEC16]], [[RDX_SHUF]]
+; GFX9-NEXT:    [[RDX_SHUF1:%.*]] = shufflevector <16 x half> [[BIN_RDX]], <16 x half> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX2:%.*]] = fadd fast <16 x half> [[BIN_RDX]], [[RDX_SHUF1]]
+; GFX9-NEXT:    [[RDX_SHUF3:%.*]] = shufflevector <16 x half> [[BIN_RDX2]], <16 x half> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX4:%.*]] = fadd fast <16 x half> [[BIN_RDX2]], [[RDX_SHUF3]]
+; GFX9-NEXT:    [[RDX_SHUF5:%.*]] = shufflevector <16 x half> [[BIN_RDX4]], <16 x half> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX6:%.*]] = fadd fast <16 x half> [[BIN_RDX4]], [[RDX_SHUF5]]
+; GFX9-NEXT:    [[TMP0:%.*]] = extractelement <16 x half> [[BIN_RDX6]], i32 0
+; GFX9-NEXT:    ret half [[TMP0]]
+;
+; VI-LABEL: @reduction_half16(
+; VI-NEXT:  entry:
+; VI-NEXT:    [[ELT0:%.*]] = extractelement <16 x half> [[VEC16:%.*]], i64 0
+; VI-NEXT:    [[ELT1:%.*]] = extractelement <16 x half> [[VEC16]], i64 1
+; VI-NEXT:    [[ELT2:%.*]] = extractelement <16 x half> [[VEC16]], i64 2
+; VI-NEXT:    [[ELT3:%.*]] = extractelement <16 x half> [[VEC16]], i64 3
+; VI-NEXT:    [[ELT4:%.*]] = extractelement <16 x half> [[VEC16]], i64 4
+; VI-NEXT:    [[ELT5:%.*]] = extractelement <16 x half> [[VEC16]], i64 5
+; VI-NEXT:    [[ELT6:%.*]] = extractelement <16 x half> [[VEC16]], i64 6
+; VI-NEXT:    [[ELT7:%.*]] = extractelement <16 x half> [[VEC16]], i64 7
+; VI-NEXT:    [[ELT8:%.*]] = extractelement <16 x half> [[VEC16]], i64 8
+; VI-NEXT:    [[ELT9:%.*]] = extractelement <16 x half> [[VEC16]], i64 9
+; VI-NEXT:    [[ELT10:%.*]] = extractelement <16 x half> [[VEC16]], i64 10
+; VI-NEXT:    [[ELT11:%.*]] = extractelement <16 x half> [[VEC16]], i64 11
+; VI-NEXT:    [[ELT12:%.*]] = extractelement <16 x half> [[VEC16]], i64 12
+; VI-NEXT:    [[ELT13:%.*]] = extractelement <16 x half> [[VEC16]], i64 13
+; VI-NEXT:    [[ELT14:%.*]] = extractelement <16 x half> [[VEC16]], i64 14
+; VI-NEXT:    [[ELT15:%.*]] = extractelement <16 x half> [[VEC16]], i64 15
+; VI-NEXT:    [[ADD1:%.*]] = fadd fast half [[ELT1]], [[ELT0]]
+; VI-NEXT:    [[ADD2:%.*]] = fadd fast half [[ELT2]], [[ADD1]]
+; VI-NEXT:    [[ADD3:%.*]] = fadd fast half [[ELT3]], [[ADD2]]
+; VI-NEXT:    [[ADD4:%.*]] = fadd fast half [[ELT4]], [[ADD3]]
+; VI-NEXT:    [[ADD5:%.*]] = fadd fast half [[ELT5]], [[ADD4]]
+; VI-NEXT:    [[ADD6:%.*]] = fadd fast half [[ELT6]], [[ADD5]]
+; VI-NEXT:    [[ADD7:%.*]] = fadd fast half [[ELT7]], [[ADD6]]
+; VI-NEXT:    [[ADD8:%.*]] = fadd fast half [[ELT8]], [[ADD7]]
+; VI-NEXT:    [[ADD9:%.*]] = fadd fast half [[ELT9]], [[ADD8]]
+; VI-NEXT:    [[ADD10:%.*]] = fadd fast half [[ELT10]], [[ADD9]]
+; VI-NEXT:    [[ADD11:%.*]] = fadd fast half [[ELT11]], [[ADD10]]
+; VI-NEXT:    [[ADD12:%.*]] = fadd fast half [[ELT12]], [[ADD11]]
+; VI-NEXT:    [[ADD13:%.*]] = fadd fast half [[ELT13]], [[ADD12]]
+; VI-NEXT:    [[ADD14:%.*]] = fadd fast half [[ELT14]], [[ADD13]]
+; VI-NEXT:    [[ADD15:%.*]] = fadd fast half [[ELT15]], [[ADD14]]
+; VI-NEXT:    ret half [[ADD15]]
+;
+entry:
+  %elt0 = extractelement <16 x half> %vec16, i64 0
+  %elt1 = extractelement <16 x half> %vec16, i64 1
+  %elt2 = extractelement <16 x half> %vec16, i64 2
+  %elt3 = extractelement <16 x half> %vec16, i64 3
+  %elt4 = extractelement <16 x half> %vec16, i64 4
+  %elt5 = extractelement <16 x half> %vec16, i64 5
+  %elt6 = extractelement <16 x half> %vec16, i64 6
+  %elt7 = extractelement <16 x half> %vec16, i64 7
+  %elt8 = extractelement <16 x half> %vec16, i64 8
+  %elt9 = extractelement <16 x half> %vec16, i64 9
+  %elt10 = extractelement <16 x half> %vec16, i64 10
+  %elt11 = extractelement <16 x half> %vec16, i64 11
+  %elt12 = extractelement <16 x half> %vec16, i64 12
+  %elt13 = extractelement <16 x half> %vec16, i64 13
+  %elt14 = extractelement <16 x half> %vec16, i64 14
+  %elt15 = extractelement <16 x half> %vec16, i64 15
+
+  %add1 = fadd fast half %elt1, %elt0
+  %add2 = fadd fast half %elt2, %add1
+  %add3 = fadd fast half %elt3, %add2
+  %add4 = fadd fast half %elt4, %add3
+  %add5 = fadd fast half %elt5, %add4
+  %add6 = fadd fast half %elt6, %add5
+  %add7 = fadd fast half %elt7, %add6
+  %add8 = fadd fast half %elt8, %add7
+  %add9 = fadd fast half %elt9, %add8
+  %add10 = fadd fast half %elt10, %add9
+  %add11 = fadd fast half %elt11, %add10
+  %add12 = fadd fast half %elt12, %add11
+  %add13 = fadd fast half %elt13, %add12
+  %add14 = fadd fast half %elt14, %add13
+  %add15 = fadd fast half %elt15, %add14
+
+  ret half %add15
+}
+
+; FIXME: support vectorization;
+define half @reduction_sub_half4(<4 x half> %a) {
+; GCN-LABEL: @reduction_sub_half4(
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[ELT0:%.*]] = extractelement <4 x half> [[A:%.*]], i64 0
+; GCN-NEXT:    [[ELT1:%.*]] = extractelement <4 x half> [[A]], i64 1
+; GCN-NEXT:    [[ELT2:%.*]] = extractelement <4 x half> [[A]], i64 2
+; GCN-NEXT:    [[ELT3:%.*]] = extractelement <4 x half> [[A]], i64 3
+; GCN-NEXT:    [[ADD1:%.*]] = fsub fast half [[ELT1]], [[ELT0]]
+; GCN-NEXT:    [[ADD2:%.*]] = fsub fast half [[ELT2]], [[ADD1]]
+; GCN-NEXT:    [[ADD3:%.*]] = fsub fast half [[ELT3]], [[ADD2]]
+; GCN-NEXT:    ret half [[ADD3]]
+;
+entry:
+  %elt0 = extractelement <4 x half> %a, i64 0
+  %elt1 = extractelement <4 x half> %a, i64 1
+  %elt2 = extractelement <4 x half> %a, i64 2
+  %elt3 = extractelement <4 x half> %a, i64 3
+
+  %add1 = fsub fast half %elt1, %elt0
+  %add2 = fsub fast half %elt2, %add1
+  %add3 = fsub fast half %elt3, %add2
+
+  ret half %add3
+}
+
+define i16 @reduction_v4i16(<4 x i16> %a) {
+; GFX9-LABEL: @reduction_v4i16(
+; GFX9-NEXT:  entry:
+; GFX9-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <4 x i16> [[A:%.*]], <4 x i16> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX:%.*]] = add <4 x i16> [[A]], [[RDX_SHUF]]
+; GFX9-NEXT:    [[RDX_SHUF1:%.*]] = shufflevector <4 x i16> [[BIN_RDX]], <4 x i16> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX2:%.*]] = add <4 x i16> [[BIN_RDX]], [[RDX_SHUF1]]
+; GFX9-NEXT:    [[TMP0:%.*]] = extractelement <4 x i16> [[BIN_RDX2]], i32 0
+; GFX9-NEXT:    ret i16 [[TMP0]]
+;
+; VI-LABEL: @reduction_v4i16(
+; VI-NEXT:  entry:
+; VI-NEXT:    [[ELT0:%.*]] = extractelement <4 x i16> [[A:%.*]], i64 0
+; VI-NEXT:    [[ELT1:%.*]] = extractelement <4 x i16> [[A]], i64 1
+; VI-NEXT:    [[ELT2:%.*]] = extractelement <4 x i16> [[A]], i64 2
+; VI-NEXT:    [[ELT3:%.*]] = extractelement <4 x i16> [[A]], i64 3
+; VI-NEXT:    [[ADD1:%.*]] = add i16 [[ELT1]], [[ELT0]]
+; VI-NEXT:    [[ADD2:%.*]] = add i16 [[ELT2]], [[ADD1]]
+; VI-NEXT:    [[ADD3:%.*]] = add i16 [[ELT3]], [[ADD2]]
+; VI-NEXT:    ret i16 [[ADD3]]
+;
+entry:
+  %elt0 = extractelement <4 x i16> %a, i64 0
+  %elt1 = extractelement <4 x i16> %a, i64 1
+  %elt2 = extractelement <4 x i16> %a, i64 2
+  %elt3 = extractelement <4 x i16> %a, i64 3
+
+  %add1 = add i16 %elt1, %elt0
+  %add2 = add i16 %elt2, %add1
+  %add3 = add i16 %elt3, %add2
+
+  ret i16 %add3
+}
+
+define i16 @reduction_v8i16(<8 x i16> %vec8) {
+; GFX9-LABEL: @reduction_v8i16(
+; GFX9-NEXT:  entry:
+; GFX9-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <8 x i16> [[VEC8:%.*]], <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX:%.*]] = add <8 x i16> [[VEC8]], [[RDX_SHUF]]
+; GFX9-NEXT:    [[RDX_SHUF1:%.*]] = shufflevector <8 x i16> [[BIN_RDX]], <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX2:%.*]] = add <8 x i16> [[BIN_RDX]], [[RDX_SHUF1]]
+; GFX9-NEXT:    [[RDX_SHUF3:%.*]] = shufflevector <8 x i16> [[BIN_RDX2]], <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+; GFX9-NEXT:    [[BIN_RDX4:%.*]] = add <8 x i16> [[BIN_RDX2]], [[RDX_SHUF3]]
+; GFX9-NEXT:    [[TMP0:%.*]] = extractelement <8 x i16> [[BIN_RDX4]], i32 0
+; GFX9-NEXT:    ret i16 [[TMP0]]
+;
+; VI-LABEL: @reduction_v8i16(
+; VI-NEXT:  entry:
+; VI-NEXT:    [[ELT0:%.*]] = extractelement <8 x i16> [[VEC8:%.*]], i64 0
+; VI-NEXT:    [[ELT1:%.*]] = extractelement <8 x i16> [[VEC8]], i64 1
+; VI-NEXT:    [[ELT2:%.*]] = extractelement <8 x i16> [[VEC8]], i64 2
+; VI-NEXT:    [[ELT3:%.*]] = extractelement <8 x i16> [[VEC8]], i64 3
+; VI-NEXT:    [[ELT4:%.*]] = extractelement <8 x i16> [[VEC8]], i64 4
+; VI-NEXT:    [[ELT5:%.*]] = extractelement <8 x i16> [[VEC8]], i64 5
+; VI-NEXT:    [[ELT6:%.*]] = extractelement <8 x i16> [[VEC8]], i64 6
+; VI-NEXT:    [[ELT7:%.*]] = extractelement <8 x i16> [[VEC8]], i64 7
+; VI-NEXT:    [[ADD1:%.*]] = add i16 [[ELT1]], [[ELT0]]
+; VI-NEXT:    [[ADD2:%.*]] = add i16 [[ELT2]], [[ADD1]]
+; VI-NEXT:    [[ADD3:%.*]] = add i16 [[ELT3]], [[ADD2]]
+; VI-NEXT:    [[ADD4:%.*]] = add i16 [[ELT4]], [[ADD3]]
+; VI-NEXT:    [[ADD5:%.*]] = add i16 [[ELT5]], [[ADD4]]
+; VI-NEXT:    [[ADD6:%.*]] = add i16 [[ELT6]], [[ADD5]]
+; VI-NEXT:    [[ADD7:%.*]] = add i16 [[ELT7]], [[ADD6]]
+; VI-NEXT:    ret i16 [[ADD7]]
+;
+entry:
+  %elt0 = extractelement <8 x i16> %vec8, i64 0
+  %elt1 = extractelement <8 x i16> %vec8, i64 1
+  %elt2 = extractelement <8 x i16> %vec8, i64 2
+  %elt3 = extractelement <8 x i16> %vec8, i64 3
+  %elt4 = extractelement <8 x i16> %vec8, i64 4
+  %elt5 = extractelement <8 x i16> %vec8, i64 5
+  %elt6 = extractelement <8 x i16> %vec8, i64 6
+  %elt7 = extractelement <8 x i16> %vec8, i64 7
+
+  %add1 = add i16 %elt1, %elt0
+  %add2 = add i16 %elt2, %add1
+  %add3 = add i16 %elt3, %add2
+  %add4 = add i16 %elt4, %add3
+  %add5 = add i16 %elt5, %add4
+  %add6 = add i16 %elt6, %add5
+  %add7 = add i16 %elt7, %add6
+
+  ret i16 %add7
+}
+
+; FIXME: This should be vectorized on GFX9.
+
+define i16 @reduction_icmp_v4i16(<4 x i16> %vec4) {
+; GCN-LABEL: @reduction_icmp_v4i16(
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[ELT0:%.*]] = extractelement <4 x i16> [[VEC4:%.*]], i64 0
+; GCN-NEXT:    [[ELT1:%.*]] = extractelement <4 x i16> [[VEC4]], i64 1
+; GCN-NEXT:    [[ELT2:%.*]] = extractelement <4 x i16> [[VEC4]], i64 2
+; GCN-NEXT:    [[ELT3:%.*]] = extractelement <4 x i16> [[VEC4]], i64 3
+; GCN-NEXT:    [[CMP1:%.*]] = icmp ult i16 [[ELT1]], [[ELT0]]
+; GCN-NEXT:    [[MIN1:%.*]] = select i1 [[CMP1]], i16 [[ELT1]], i16 [[ELT0]]
+; GCN-NEXT:    [[CMP2:%.*]] = icmp ult i16 [[ELT2]], [[MIN1]]
+; GCN-NEXT:    [[MIN2:%.*]] = select i1 [[CMP2]], i16 [[ELT2]], i16 [[MIN1]]
+; GCN-NEXT:    [[CMP3:%.*]] = icmp ult i16 [[ELT3]], [[MIN2]]
+; GCN-NEXT:    [[MIN3:%.*]] = select i1 [[CMP3]], i16 [[ELT3]], i16 [[MIN2]]
+; GCN-NEXT:    ret i16 [[MIN3]]
+;
+entry:
+  %elt0 = extractelement <4 x i16> %vec4, i64 0
+  %elt1 = extractelement <4 x i16> %vec4, i64 1
+  %elt2 = extractelement <4 x i16> %vec4, i64 2
+  %elt3 = extractelement <4 x i16> %vec4, i64 3
+
+  %cmp1 = icmp ult i16 %elt1, %elt0
+  %min1 = select i1 %cmp1, i16 %elt1, i16 %elt0
+  %cmp2 = icmp ult i16 %elt2, %min1
+  %min2 = select i1 %cmp2, i16 %elt2, i16 %min1
+  %cmp3 = icmp ult i16 %elt3, %min2
+  %min3 = select i1 %cmp3, i16 %elt3, i16 %min2
+
+  ret i16 %min3
+}
+
+; Tests to make sure reduction does not kick in. vega does not support packed math for types larger than 16 bits.
+define float @reduction_v4float(<4 x float> %a) {
+; GCN-LABEL: @reduction_v4float(
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[ELT0:%.*]] = extractelement <4 x float> [[A:%.*]], i64 0
+; GCN-NEXT:    [[ELT1:%.*]] = extractelement <4 x float> [[A]], i64 1
+; GCN-NEXT:    [[ELT2:%.*]] = extractelement <4 x float> [[A]], i64 2
+; GCN-NEXT:    [[ELT3:%.*]] = extractelement <4 x float> [[A]], i64 3
+; GCN-NEXT:    [[ADD1:%.*]] = fadd fast float [[ELT1]], [[ELT0]]
+; GCN-NEXT:    [[ADD2:%.*]] = fadd fast float [[ELT2]], [[ADD1]]
+; GCN-NEXT:    [[ADD3:%.*]] = fadd fast float [[ELT3]], [[ADD2]]
+; GCN-NEXT:    ret float [[ADD3]]
+;
+entry:
+  %elt0 = extractelement <4 x float> %a, i64 0
+  %elt1 = extractelement <4 x float> %a, i64 1
+  %elt2 = extractelement <4 x float> %a, i64 2
+  %elt3 = extractelement <4 x float> %a, i64 3
+
+  %add1 = fadd fast float %elt1, %elt0
+  %add2 = fadd fast float %elt2, %add1
+  %add3 = fadd fast float %elt3, %add2
+
+  ret float %add3
+}
\ No newline at end of file