[clang] [llvm] [X86] LowerBITREVERSE - use AND+CMPEQ+MOVMSK trick to lower scalar types (PR #92236)

Simon Pilgrim via llvm-commits llvm-commits at lists.llvm.org
Fri Nov 8 04:03:49 PST 2024


https://github.com/RKSimon updated https://github.com/llvm/llvm-project/pull/92236

>From 37fcf8300243fa6236e119743fd6d7067ac515ca Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Thu, 7 Nov 2024 17:25:49 +0000
Subject: [PATCH 1/2] [clang][x86] _mm_movpi64_epi64 - convert to shufflevector
 pattern instead of bitcasting to i64

Don't bitcast a v1i64 to i64 as constant expressions will struggle to handle this - convert to a shufflevector concat pattern like _mm_move_epi64 instead
---
 clang/lib/Headers/emmintrin.h          | 2 +-
 clang/test/CodeGen/X86/sse2-builtins.c | 4 +---
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/clang/lib/Headers/emmintrin.h b/clang/lib/Headers/emmintrin.h
index 947ad891d2c85c..de453cc1a38b99 100644
--- a/clang/lib/Headers/emmintrin.h
+++ b/clang/lib/Headers/emmintrin.h
@@ -4627,7 +4627,7 @@ _mm_movepi64_pi64(__m128i __a) {
 /// \returns A 128-bit integer vector. The lower 64 bits contain the value from
 ///    the operand. The upper 64 bits are assigned zeros.
 static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_movpi64_epi64(__m64 __a) {
-  return __extension__(__m128i)(__v2di){(long long)__a, 0};
+  return __builtin_shufflevector((__v1di)__a, _mm_setzero_si64(), 0, 1);
 }
 
 /// Moves the lower 64 bits of a 128-bit integer vector to a 128-bit
diff --git a/clang/test/CodeGen/X86/sse2-builtins.c b/clang/test/CodeGen/X86/sse2-builtins.c
index 1db40587be466d..6defd94c5d5073 100644
--- a/clang/test/CodeGen/X86/sse2-builtins.c
+++ b/clang/test/CodeGen/X86/sse2-builtins.c
@@ -878,9 +878,7 @@ TEST_CONSTEXPR(match_m64(_mm_movepi64_pi64((__m128i){8, -8}), 8ULL));
 __m128i test_mm_movpi64_epi64(__m64 A)
 {
   // CHECK-LABEL: test_mm_movpi64_epi64
-  // CHECK: [[CAST:%.*]] = bitcast <1 x i64> %{{.*}} to i64
-  // CHECK: [[INS:%.*]] = insertelement <2 x i64> poison, i64 [[CAST]], i32 0
-  // CHECK: insertelement <2 x i64> [[INS]], i64 0, i32 1
+  // CHECK: shufflevector <1 x i64> %{{.*}}, <1 x i64> %{{.*}}, <2 x i32> <i32 0, i32 1>
   return _mm_movpi64_epi64(A);
 }
 

>From e44ca60fa9b416d8ff79c233503b0798a96d5080 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Wed, 15 May 2024 10:28:36 +0100
Subject: [PATCH 2/2] [X86] LowerBITREVERSE - use AND+CMPEQ+MOVMSK trick to
 lower i8/i16/i32/i64 scalar types

By splitting each byte of a scalar type into 8 copies, in reverse order, we can then test each one to see if the bit is set and then use PMOVMSKB to pack them back together to get the bit reversal.

So far I've only managed to make this worth it for i8/i16 with SSSE3 (PSHUFB) and i32/i64 with AVX2 - AVX512BW can use VPTESTMB which is even nicer.

Fixes #79794
---
 llvm/lib/Target/X86/X86ISelLowering.cpp       |  76 ++++
 .../lib/Target/X86/X86TargetTransformInfo.cpp |   2 +
 .../CostModel/X86/bitreverse-codesize.ll      |  48 +-
 .../CostModel/X86/bitreverse-latency.ll       |  48 +-
 .../CostModel/X86/bitreverse-sizelatency.ll   |  48 +-
 .../test/Analysis/CostModel/X86/bitreverse.ll |  48 +-
 llvm/test/CodeGen/X86/vector-bitreverse.ll    | 429 +++++++++++++-----
 7 files changed, 529 insertions(+), 170 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 1ab879e9036b3e..1ced8b5546a995 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -1344,6 +1344,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
       setOperationAction(ISD::BITREVERSE,       VT, Custom);
       setOperationAction(ISD::CTLZ,             VT, Custom);
     }
+    setOperationAction(ISD::BITREVERSE,         MVT::i8,    Custom);
+    setOperationAction(ISD::BITREVERSE,         MVT::i16,   Custom);
 
     // These might be better off as horizontal vector ops.
     setOperationAction(ISD::ADD,                MVT::i16, Custom);
@@ -1566,6 +1568,12 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
       setCondCodeAction(ISD::SETLE, VT, Custom);
     }
 
+    // TODO: Support AVX1 once i32 bitreverse codegen is better.
+    if (Subtarget.hasInt256()) {
+      setOperationAction(ISD::BITREVERSE, MVT::i32, Custom);
+      setOperationAction(ISD::BITREVERSE, MVT::i64, Custom);
+    }
+
     setOperationAction(ISD::SETCC,          MVT::v4f64, Custom);
     setOperationAction(ISD::SETCC,          MVT::v8f32, Custom);
     setOperationAction(ISD::STRICT_FSETCC,  MVT::v4f64, Custom);
@@ -32089,6 +32097,68 @@ static SDValue LowerBITREVERSE_XOP(SDValue Op, SelectionDAG &DAG) {
   return DAG.getBitcast(VT, Res);
 }
 
+// Lower i8/i16/i32 with vXi8 AND+CMPEQ+MOVMSK trick.
+// Broadcast each byte across 8 bytes in byteswaped order, test if each bit
+// is set and use MOVMSK/VPTESTMB to pack the bit results together.
+static SDValue LowerBITREVERSE_MOVMSK(SDValue Op, const X86Subtarget &Subtarget,
+                                      SelectionDAG &DAG) {
+  MVT VT = Op.getSimpleValueType();
+  assert(!VT.isVector() && "Scalar type expected");
+  SDValue In = Op.getOperand(0);
+  SDLoc DL(Op);
+
+  unsigned MaxBits =
+      Subtarget.hasBWI() ? 512 : (Subtarget.hasInt256() ? 256 : 128);
+  unsigned AllBits = std::max<unsigned>(VT.getSizeInBits() * 8, 128);
+  unsigned VecBits = std::min<unsigned>(AllBits, MaxBits);
+  unsigned MaxBytes = MaxBits / 8;
+  unsigned NumBytes = VecBits / 8;
+
+  MVT ConstVT = MVT::getVectorVT(MVT::i64, VecBits / 64);
+  MVT ByteVT = MVT::getVectorVT(MVT::i8, VecBits / 8);
+  MVT SrcVT = MVT::getVectorVT(VT, VecBits / VT.getScalarSizeInBits());
+
+  SDValue Splat = DAG.getBitcast(ByteVT, DAG.getSplat(SrcVT, DL, In));
+  SDValue Mask = DAG.getBitcast(
+      ByteVT, DAG.getConstant(0x102040810204080ULL, DL, ConstVT));
+
+  // Repeat each byte 8 times in reverse order.
+  // Create a big shuffle mask and then
+  SmallVector<int, 32> ByteSplatMask;
+  for (unsigned I = 0, E = VT.getSizeInBits() / 8; I != E; ++I)
+    ByteSplatMask.append(8, (E - 1) - I);
+
+  if (ByteSplatMask.size() < 16)
+    ByteSplatMask.append(16 - ByteSplatMask.size(), -1);
+
+  SDValue Res = DAG.getConstant(0, DL, VT);
+  for (unsigned Ofs = 0; Ofs < VT.getSizeInBits(); Ofs += MaxBytes) {
+    // Repeat each byte 8 times in reverse order.
+    ArrayRef SubMask = ArrayRef(ByteSplatMask).slice(Ofs, NumBytes);
+    SDValue Bytes = DAG.getVectorShuffle(ByteVT, DL, Splat, Splat, SubMask);
+
+    // Isolate one bit per byte.
+    SDValue And = DAG.getNode(ISD::AND, DL, ByteVT, Bytes, Mask);
+
+    SDValue Bits;
+    if (Subtarget.hasBWI() && (VT == MVT::i64 || Subtarget.hasVLX())) {
+      MVT CmpVT = MVT::getVectorVT(MVT::i1, NumBytes);
+      SDValue Zero = DAG.getConstant(0, DL, ByteVT);
+      SDValue Cmp = DAG.getSetCC(DL, CmpVT, And, Zero, ISD::CondCode::SETNE);
+      SDValue Msk = DAG.getBitcast(MVT::getIntegerVT(NumBytes), Cmp);
+      Bits = DAG.getZExtOrTrunc(Msk, DL, VT);
+    } else {
+      SDValue Cmp = DAG.getSetCC(DL, ByteVT, And, Mask, ISD::CondCode::SETEQ);
+      SDValue Msk = getPMOVMSKB(DL, Cmp, DAG, Subtarget);
+      Bits = DAG.getZExtOrTrunc(Msk, DL, VT);
+    }
+    Bits = DAG.getNode(ISD::SHL, DL, VT, Bits,
+                       DAG.getShiftAmountConstant(Ofs, VT, DL));
+    Res = DAG.getNode(ISD::OR, DL, VT, Res, Bits);
+  }
+  return Res;
+}
+
 static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget,
                                SelectionDAG &DAG) {
   MVT VT = Op.getSimpleValueType();
@@ -32114,6 +32184,12 @@ static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget,
     assert(
         (VT == MVT::i32 || VT == MVT::i64 || VT == MVT::i16 || VT == MVT::i8) &&
         "Only tested for i8/i16/i32/i64");
+
+    if (!Subtarget.hasGFNI())
+      if (SDValue R = LowerBITREVERSE_MOVMSK(Op, Subtarget, DAG))
+        return R;
+
+    // Lower i8/i16/i32/i64 as vXi8 BITREVERSE + BSWAP
     MVT VecVT = MVT::getVectorVT(VT, 128 / VT.getSizeInBits());
     SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, In);
     Res = DAG.getNode(ISD::BITREVERSE, DL, MVT::v16i8,
diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index 7f362f48c2e78f..befe3233d233f7 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -4193,6 +4193,8 @@ X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
     { ISD::BITREVERSE, MVT::v4i32,   { 16, 20, 11, 21 } },
     { ISD::BITREVERSE, MVT::v8i16,   { 16, 20, 11, 21 } },
     { ISD::BITREVERSE, MVT::v16i8,   { 11, 12, 10, 16 } },
+    { ISD::BITREVERSE, MVT::i16,     {  5,  6,  6, 10 } }, // AND+PCMPEQB+PMOVMSKB
+    { ISD::BITREVERSE, MVT::i8,      {  4,  4,  7,  8 } }, // AND+PCMPEQB+PMOVMSKB
     { ISD::BSWAP,      MVT::v2i64,   {  2,  3,  1,  5 } },
     { ISD::BSWAP,      MVT::v4i32,   {  2,  3,  1,  5 } },
     { ISD::BSWAP,      MVT::v8i16,   {  2,  3,  1,  5 } },
diff --git a/llvm/test/Analysis/CostModel/X86/bitreverse-codesize.ll b/llvm/test/Analysis/CostModel/X86/bitreverse-codesize.ll
index e02ba761aa8d29..460bb60eddd68b 100644
--- a/llvm/test/Analysis/CostModel/X86/bitreverse-codesize.ll
+++ b/llvm/test/Analysis/CostModel/X86/bitreverse-codesize.ll
@@ -101,13 +101,25 @@ define i32 @var_bitreverse_i32(i32 %a) {
 }
 
 define i16 @var_bitreverse_i16(i16 %a) {
-; X86-LABEL: 'var_bitreverse_i16'
-; X86-NEXT:  Cost Model: Found an estimated cost of 17 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i16'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 17 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i16'
-; X64-NEXT:  Cost Model: Found an estimated cost of 17 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE42-LABEL: 'var_bitreverse_i16'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i16'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i16'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX512-LABEL: 'var_bitreverse_i16'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i16'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
@@ -138,13 +150,25 @@ define i16 @var_bitreverse_i16(i16 %a) {
 }
 
 define i8 @var_bitreverse_i8(i8 %a) {
-; X86-LABEL: 'var_bitreverse_i8'
-; X86-NEXT:  Cost Model: Found an estimated cost of 13 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i8'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 13 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; SSE42-LABEL: 'var_bitreverse_i8'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i8'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i8'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i8'
-; X64-NEXT:  Cost Model: Found an estimated cost of 13 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; AVX512-LABEL: 'var_bitreverse_i8'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i8'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
diff --git a/llvm/test/Analysis/CostModel/X86/bitreverse-latency.ll b/llvm/test/Analysis/CostModel/X86/bitreverse-latency.ll
index ba231b985c26bf..6e7e97cecc99e2 100644
--- a/llvm/test/Analysis/CostModel/X86/bitreverse-latency.ll
+++ b/llvm/test/Analysis/CostModel/X86/bitreverse-latency.ll
@@ -101,13 +101,25 @@ define i32 @var_bitreverse_i32(i32 %a) {
 }
 
 define i16 @var_bitreverse_i16(i16 %a) {
-; X86-LABEL: 'var_bitreverse_i16'
-; X86-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i16'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i16'
-; X64-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE42-LABEL: 'var_bitreverse_i16'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i16'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i16'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX512-LABEL: 'var_bitreverse_i16'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i16'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
@@ -138,13 +150,25 @@ define i16 @var_bitreverse_i16(i16 %a) {
 }
 
 define i8 @var_bitreverse_i8(i8 %a) {
-; X86-LABEL: 'var_bitreverse_i8'
-; X86-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i8'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; SSE42-LABEL: 'var_bitreverse_i8'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i8'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i8'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i8'
-; X64-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; AVX512-LABEL: 'var_bitreverse_i8'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i8'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
diff --git a/llvm/test/Analysis/CostModel/X86/bitreverse-sizelatency.ll b/llvm/test/Analysis/CostModel/X86/bitreverse-sizelatency.ll
index d60fac228fc069..a2dbe0837d7735 100644
--- a/llvm/test/Analysis/CostModel/X86/bitreverse-sizelatency.ll
+++ b/llvm/test/Analysis/CostModel/X86/bitreverse-sizelatency.ll
@@ -101,13 +101,25 @@ define i32 @var_bitreverse_i32(i32 %a) {
 }
 
 define i16 @var_bitreverse_i16(i16 %a) {
-; X86-LABEL: 'var_bitreverse_i16'
-; X86-NEXT:  Cost Model: Found an estimated cost of 19 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i16'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 19 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i16'
-; X64-NEXT:  Cost Model: Found an estimated cost of 19 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+; SSE42-LABEL: 'var_bitreverse_i16'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i16'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i16'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
+;
+; AVX512-LABEL: 'var_bitreverse_i16'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i16 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i16'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
@@ -138,13 +150,25 @@ define i16 @var_bitreverse_i16(i16 %a) {
 }
 
 define i8 @var_bitreverse_i8(i8 %a) {
-; X86-LABEL: 'var_bitreverse_i8'
-; X86-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i8'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; SSE42-LABEL: 'var_bitreverse_i8'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i8'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i8'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i8'
-; X64-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
+; AVX512-LABEL: 'var_bitreverse_i8'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret i8 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i8'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
diff --git a/llvm/test/Analysis/CostModel/X86/bitreverse.ll b/llvm/test/Analysis/CostModel/X86/bitreverse.ll
index a890147fee465a..3d9a518a26e862 100644
--- a/llvm/test/Analysis/CostModel/X86/bitreverse.ll
+++ b/llvm/test/Analysis/CostModel/X86/bitreverse.ll
@@ -101,13 +101,25 @@ define i32 @var_bitreverse_i32(i32 %a) {
 }
 
 define i16 @var_bitreverse_i16(i16 %a) {
-; X86-LABEL: 'var_bitreverse_i16'
-; X86-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i16'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i16'
-; X64-NEXT:  Cost Model: Found an estimated cost of 9 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
+; SSE42-LABEL: 'var_bitreverse_i16'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i16'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i16'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
+;
+; AVX512-LABEL: 'var_bitreverse_i16'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i16 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i16'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %bitreverse = call i16 @llvm.bitreverse.i16(i16 %a)
@@ -138,13 +150,25 @@ define i16 @var_bitreverse_i16(i16 %a) {
 }
 
 define i8 @var_bitreverse_i8(i8 %a) {
-; X86-LABEL: 'var_bitreverse_i8'
-; X86-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X86-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
+; SSE2-LABEL: 'var_bitreverse_i8'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
+;
+; SSE42-LABEL: 'var_bitreverse_i8'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
+;
+; AVX1-LABEL: 'var_bitreverse_i8'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
+;
+; AVX2-LABEL: 'var_bitreverse_i8'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
 ;
-; X64-LABEL: 'var_bitreverse_i8'
-; X64-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
-; X64-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
+; AVX512-LABEL: 'var_bitreverse_i8'
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
+; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i8 %bitreverse
 ;
 ; XOP-LABEL: 'var_bitreverse_i8'
 ; XOP-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %bitreverse = call i8 @llvm.bitreverse.i8(i8 %a)
diff --git a/llvm/test/CodeGen/X86/vector-bitreverse.ll b/llvm/test/CodeGen/X86/vector-bitreverse.ll
index 5dcf19013f0b7c..a7b7e9397b4d25 100644
--- a/llvm/test/CodeGen/X86/vector-bitreverse.ll
+++ b/llvm/test/CodeGen/X86/vector-bitreverse.ll
@@ -4,8 +4,8 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=ALL,AVX,AVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=ALL,AVX,AVX2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512F
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512BW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512BW
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512BW,AVX512BWNOVL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512BW,AVX512BWVL
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx | FileCheck %s --check-prefixes=ALL,XOP,XOPAVX1
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx2 | FileCheck %s --check-prefixes=ALL,XOP,XOPAVX2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3,+gfni | FileCheck %s --check-prefixes=ALL,GFNISSE
@@ -18,39 +18,89 @@
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx512bw
 
 define i8 @test_bitreverse_i8(i8 %a) nounwind {
-; SSE-LABEL: test_bitreverse_i8:
-; SSE:       # %bb.0:
-; SSE-NEXT:    rolb $4, %dil
-; SSE-NEXT:    movl %edi, %eax
-; SSE-NEXT:    andb $51, %al
-; SSE-NEXT:    shlb $2, %al
-; SSE-NEXT:    shrb $2, %dil
-; SSE-NEXT:    andb $51, %dil
-; SSE-NEXT:    orb %dil, %al
-; SSE-NEXT:    movl %eax, %ecx
-; SSE-NEXT:    andb $85, %cl
-; SSE-NEXT:    addb %cl, %cl
-; SSE-NEXT:    shrb %al
-; SSE-NEXT:    andb $85, %al
-; SSE-NEXT:    orb %cl, %al
-; SSE-NEXT:    retq
+; SSE2-LABEL: test_bitreverse_i8:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    rolb $4, %dil
+; SSE2-NEXT:    movl %edi, %eax
+; SSE2-NEXT:    andb $51, %al
+; SSE2-NEXT:    shlb $2, %al
+; SSE2-NEXT:    shrb $2, %dil
+; SSE2-NEXT:    andb $51, %dil
+; SSE2-NEXT:    orb %dil, %al
+; SSE2-NEXT:    movl %eax, %ecx
+; SSE2-NEXT:    andb $85, %cl
+; SSE2-NEXT:    addb %cl, %cl
+; SSE2-NEXT:    shrb %al
+; SSE2-NEXT:    andb $85, %al
+; SSE2-NEXT:    orb %cl, %al
+; SSE2-NEXT:    retq
 ;
-; AVX-LABEL: test_bitreverse_i8:
-; AVX:       # %bb.0:
-; AVX-NEXT:    rolb $4, %dil
-; AVX-NEXT:    movl %edi, %eax
-; AVX-NEXT:    andb $51, %al
-; AVX-NEXT:    shlb $2, %al
-; AVX-NEXT:    shrb $2, %dil
-; AVX-NEXT:    andb $51, %dil
-; AVX-NEXT:    orb %dil, %al
-; AVX-NEXT:    movl %eax, %ecx
-; AVX-NEXT:    andb $85, %cl
-; AVX-NEXT:    addb %cl, %cl
-; AVX-NEXT:    shrb %al
-; AVX-NEXT:    andb $85, %al
-; AVX-NEXT:    orb %cl, %al
-; AVX-NEXT:    retq
+; SSSE3-LABEL: test_bitreverse_i8:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    movd %edi, %xmm0
+; SSSE3-NEXT:    pxor %xmm1, %xmm1
+; SSSE3-NEXT:    pshufb %xmm1, %xmm0
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSSE3-NEXT:    pmovmskb %xmm0, %eax
+; SSSE3-NEXT:    # kill: def $al killed $al killed $eax
+; SSSE3-NEXT:    retq
+;
+; AVX1-LABEL: test_bitreverse_i8:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vmovd %edi, %xmm0
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vmovddup {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX1-NEXT:    # xmm1 = mem[0,0]
+; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    # kill: def $al killed $al killed $eax
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_bitreverse_i8:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpmovmskb %xmm0, %eax
+; AVX2-NEXT:    # kill: def $al killed $al killed $eax
+; AVX2-NEXT:    retq
+;
+; AVX512F-LABEL: test_bitreverse_i8:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vmovd %edi, %xmm0
+; AVX512F-NEXT:    vpbroadcastb %xmm0, %xmm0
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512F-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT:    vpmovmskb %xmm0, %eax
+; AVX512F-NEXT:    # kill: def $al killed $al killed $eax
+; AVX512F-NEXT:    retq
+;
+; AVX512BWNOVL-LABEL: test_bitreverse_i8:
+; AVX512BWNOVL:       # %bb.0:
+; AVX512BWNOVL-NEXT:    vmovd %edi, %xmm0
+; AVX512BWNOVL-NEXT:    vpbroadcastb %xmm0, %xmm0
+; AVX512BWNOVL-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512BWNOVL-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX512BWNOVL-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX512BWNOVL-NEXT:    vpmovmskb %xmm0, %eax
+; AVX512BWNOVL-NEXT:    # kill: def $al killed $al killed $eax
+; AVX512BWNOVL-NEXT:    retq
+;
+; AVX512BWVL-LABEL: test_bitreverse_i8:
+; AVX512BWVL:       # %bb.0:
+; AVX512BWVL-NEXT:    vpbroadcastb %edi, %xmm0
+; AVX512BWVL-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512BWVL-NEXT:    vptestmb %xmm1, %xmm0, %k0
+; AVX512BWVL-NEXT:    kmovd %k0, %eax
+; AVX512BWVL-NEXT:    # kill: def $al killed $al killed $eax
+; AVX512BWVL-NEXT:    retq
 ;
 ; XOP-LABEL: test_bitreverse_i8:
 ; XOP:       # %bb.0:
@@ -80,51 +130,94 @@ define i8 @test_bitreverse_i8(i8 %a) nounwind {
 }
 
 define i16 @test_bitreverse_i16(i16 %a) nounwind {
-; SSE-LABEL: test_bitreverse_i16:
-; SSE:       # %bb.0:
-; SSE-NEXT:    # kill: def $edi killed $edi def $rdi
-; SSE-NEXT:    rolw $8, %di
-; SSE-NEXT:    movl %edi, %eax
-; SSE-NEXT:    andl $3855, %eax # imm = 0xF0F
-; SSE-NEXT:    shll $4, %eax
-; SSE-NEXT:    shrl $4, %edi
-; SSE-NEXT:    andl $3855, %edi # imm = 0xF0F
-; SSE-NEXT:    orl %eax, %edi
-; SSE-NEXT:    movl %edi, %eax
-; SSE-NEXT:    andl $13107, %eax # imm = 0x3333
-; SSE-NEXT:    shrl $2, %edi
-; SSE-NEXT:    andl $13107, %edi # imm = 0x3333
-; SSE-NEXT:    leal (%rdi,%rax,4), %eax
-; SSE-NEXT:    movl %eax, %ecx
-; SSE-NEXT:    andl $21845, %ecx # imm = 0x5555
-; SSE-NEXT:    shrl %eax
-; SSE-NEXT:    andl $21845, %eax # imm = 0x5555
-; SSE-NEXT:    leal (%rax,%rcx,2), %eax
-; SSE-NEXT:    # kill: def $ax killed $ax killed $eax
-; SSE-NEXT:    retq
+; SSE2-LABEL: test_bitreverse_i16:
+; SSE2:       # %bb.0:
+; SSE2-NEXT:    # kill: def $edi killed $edi def $rdi
+; SSE2-NEXT:    rolw $8, %di
+; SSE2-NEXT:    movl %edi, %eax
+; SSE2-NEXT:    andl $3855, %eax # imm = 0xF0F
+; SSE2-NEXT:    shll $4, %eax
+; SSE2-NEXT:    shrl $4, %edi
+; SSE2-NEXT:    andl $3855, %edi # imm = 0xF0F
+; SSE2-NEXT:    orl %eax, %edi
+; SSE2-NEXT:    movl %edi, %eax
+; SSE2-NEXT:    andl $13107, %eax # imm = 0x3333
+; SSE2-NEXT:    shrl $2, %edi
+; SSE2-NEXT:    andl $13107, %edi # imm = 0x3333
+; SSE2-NEXT:    leal (%rdi,%rax,4), %eax
+; SSE2-NEXT:    movl %eax, %ecx
+; SSE2-NEXT:    andl $21845, %ecx # imm = 0x5555
+; SSE2-NEXT:    shrl %eax
+; SSE2-NEXT:    andl $21845, %eax # imm = 0x5555
+; SSE2-NEXT:    leal (%rax,%rcx,2), %eax
+; SSE2-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSE2-NEXT:    retq
 ;
-; AVX-LABEL: test_bitreverse_i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX-NEXT:    rolw $8, %di
-; AVX-NEXT:    movl %edi, %eax
-; AVX-NEXT:    andl $3855, %eax # imm = 0xF0F
-; AVX-NEXT:    shll $4, %eax
-; AVX-NEXT:    shrl $4, %edi
-; AVX-NEXT:    andl $3855, %edi # imm = 0xF0F
-; AVX-NEXT:    orl %eax, %edi
-; AVX-NEXT:    movl %edi, %eax
-; AVX-NEXT:    andl $13107, %eax # imm = 0x3333
-; AVX-NEXT:    shrl $2, %edi
-; AVX-NEXT:    andl $13107, %edi # imm = 0x3333
-; AVX-NEXT:    leal (%rdi,%rax,4), %eax
-; AVX-NEXT:    movl %eax, %ecx
-; AVX-NEXT:    andl $21845, %ecx # imm = 0x5555
-; AVX-NEXT:    shrl %eax
-; AVX-NEXT:    andl $21845, %eax # imm = 0x5555
-; AVX-NEXT:    leal (%rax,%rcx,2), %eax
-; AVX-NEXT:    # kill: def $ax killed $ax killed $eax
-; AVX-NEXT:    retq
+; SSSE3-LABEL: test_bitreverse_i16:
+; SSSE3:       # %bb.0:
+; SSSE3-NEXT:    movd %edi, %xmm0
+; SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pcmpeqb %xmm1, %xmm0
+; SSSE3-NEXT:    pmovmskb %xmm0, %eax
+; SSSE3-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSSE3-NEXT:    retq
+;
+; AVX1-LABEL: test_bitreverse_i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vmovd %edi, %xmm0
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX1-NEXT:    vmovddup {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX1-NEXT:    # xmm1 = mem[0,0]
+; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpmovmskb %xmm0, %eax
+; AVX1-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_bitreverse_i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpmovmskb %xmm0, %eax
+; AVX2-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX2-NEXT:    retq
+;
+; AVX512F-LABEL: test_bitreverse_i16:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vmovd %edi, %xmm0
+; AVX512F-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512F-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT:    vpmovmskb %xmm0, %eax
+; AVX512F-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512F-NEXT:    retq
+;
+; AVX512BWNOVL-LABEL: test_bitreverse_i16:
+; AVX512BWNOVL:       # %bb.0:
+; AVX512BWNOVL-NEXT:    vmovd %edi, %xmm0
+; AVX512BWNOVL-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX512BWNOVL-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512BWNOVL-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; AVX512BWNOVL-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX512BWNOVL-NEXT:    vpmovmskb %xmm0, %eax
+; AVX512BWNOVL-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512BWNOVL-NEXT:    retq
+;
+; AVX512BWVL-LABEL: test_bitreverse_i16:
+; AVX512BWVL:       # %bb.0:
+; AVX512BWVL-NEXT:    vmovd %edi, %xmm0
+; AVX512BWVL-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX512BWVL-NEXT:    vpbroadcastq {{.*#+}} xmm1 = [72624976668147840,72624976668147840]
+; AVX512BWVL-NEXT:    vptestmb %xmm1, %xmm0, %k0
+; AVX512BWVL-NEXT:    kmovd %k0, %eax
+; AVX512BWVL-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512BWVL-NEXT:    retq
 ;
 ; XOP-LABEL: test_bitreverse_i16:
 ; XOP:       # %bb.0:
@@ -178,27 +271,73 @@ define i32 @test_bitreverse_i32(i32 %a) nounwind {
 ; SSE-NEXT:    leal (%rax,%rcx,2), %eax
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: test_bitreverse_i32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX-NEXT:    bswapl %edi
-; AVX-NEXT:    movl %edi, %eax
-; AVX-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
-; AVX-NEXT:    shll $4, %eax
-; AVX-NEXT:    shrl $4, %edi
-; AVX-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
-; AVX-NEXT:    orl %eax, %edi
-; AVX-NEXT:    movl %edi, %eax
-; AVX-NEXT:    andl $858993459, %eax # imm = 0x33333333
-; AVX-NEXT:    shrl $2, %edi
-; AVX-NEXT:    andl $858993459, %edi # imm = 0x33333333
-; AVX-NEXT:    leal (%rdi,%rax,4), %eax
-; AVX-NEXT:    movl %eax, %ecx
-; AVX-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
-; AVX-NEXT:    shrl %eax
-; AVX-NEXT:    andl $1431655765, %eax # imm = 0x55555555
-; AVX-NEXT:    leal (%rax,%rcx,2), %eax
-; AVX-NEXT:    retq
+; AVX1-LABEL: test_bitreverse_i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    # kill: def $edi killed $edi def $rdi
+; AVX1-NEXT:    bswapl %edi
+; AVX1-NEXT:    movl %edi, %eax
+; AVX1-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
+; AVX1-NEXT:    shll $4, %eax
+; AVX1-NEXT:    shrl $4, %edi
+; AVX1-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
+; AVX1-NEXT:    orl %eax, %edi
+; AVX1-NEXT:    movl %edi, %eax
+; AVX1-NEXT:    andl $858993459, %eax # imm = 0x33333333
+; AVX1-NEXT:    shrl $2, %edi
+; AVX1-NEXT:    andl $858993459, %edi # imm = 0x33333333
+; AVX1-NEXT:    leal (%rdi,%rax,4), %eax
+; AVX1-NEXT:    movl %eax, %ecx
+; AVX1-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
+; AVX1-NEXT:    shrl %eax
+; AVX1-NEXT:    andl $1431655765, %eax # imm = 0x55555555
+; AVX1-NEXT:    leal (%rax,%rcx,2), %eax
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_bitreverse_i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpmovmskb %ymm0, %eax
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; AVX512F-LABEL: test_bitreverse_i32:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vmovd %edi, %xmm0
+; AVX512F-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT:    vpmovmskb %ymm0, %eax
+; AVX512F-NEXT:    vzeroupper
+; AVX512F-NEXT:    retq
+;
+; AVX512BWNOVL-LABEL: test_bitreverse_i32:
+; AVX512BWNOVL:       # %bb.0:
+; AVX512BWNOVL-NEXT:    vmovd %edi, %xmm0
+; AVX512BWNOVL-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX512BWNOVL-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX512BWNOVL-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX512BWNOVL-NEXT:    vpand %ymm1, %ymm0, %ymm0
+; AVX512BWNOVL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
+; AVX512BWNOVL-NEXT:    vpmovmskb %ymm0, %eax
+; AVX512BWNOVL-NEXT:    vzeroupper
+; AVX512BWNOVL-NEXT:    retq
+;
+; AVX512BWVL-LABEL: test_bitreverse_i32:
+; AVX512BWVL:       # %bb.0:
+; AVX512BWVL-NEXT:    vpbroadcastd %edi, %ymm0
+; AVX512BWVL-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX512BWVL-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX512BWVL-NEXT:    vptestmb %ymm1, %ymm0, %k0
+; AVX512BWVL-NEXT:    kmovd %k0, %eax
+; AVX512BWVL-NEXT:    vzeroupper
+; AVX512BWVL-NEXT:    retq
 ;
 ; XOP-LABEL: test_bitreverse_i32:
 ; XOP:       # %bb.0:
@@ -251,29 +390,75 @@ define i64 @test_bitreverse_i64(i64 %a) nounwind {
 ; SSE-NEXT:    leaq (%rax,%rdx,2), %rax
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: test_bitreverse_i64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    bswapq %rdi
-; AVX-NEXT:    movq %rdi, %rax
-; AVX-NEXT:    shrq $4, %rax
-; AVX-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
-; AVX-NEXT:    andq %rcx, %rax
-; AVX-NEXT:    andq %rcx, %rdi
-; AVX-NEXT:    shlq $4, %rdi
-; AVX-NEXT:    orq %rax, %rdi
-; AVX-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
-; AVX-NEXT:    movq %rdi, %rcx
-; AVX-NEXT:    andq %rax, %rcx
-; AVX-NEXT:    shrq $2, %rdi
-; AVX-NEXT:    andq %rax, %rdi
-; AVX-NEXT:    leaq (%rdi,%rcx,4), %rax
-; AVX-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
-; AVX-NEXT:    movq %rax, %rdx
-; AVX-NEXT:    andq %rcx, %rdx
-; AVX-NEXT:    shrq %rax
-; AVX-NEXT:    andq %rcx, %rax
-; AVX-NEXT:    leaq (%rax,%rdx,2), %rax
-; AVX-NEXT:    retq
+; AVX1-LABEL: test_bitreverse_i64:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    bswapq %rdi
+; AVX1-NEXT:    movq %rdi, %rax
+; AVX1-NEXT:    shrq $4, %rax
+; AVX1-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
+; AVX1-NEXT:    andq %rcx, %rax
+; AVX1-NEXT:    andq %rcx, %rdi
+; AVX1-NEXT:    shlq $4, %rdi
+; AVX1-NEXT:    orq %rax, %rdi
+; AVX1-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
+; AVX1-NEXT:    movq %rdi, %rcx
+; AVX1-NEXT:    andq %rax, %rcx
+; AVX1-NEXT:    shrq $2, %rdi
+; AVX1-NEXT:    andq %rax, %rdi
+; AVX1-NEXT:    leaq (%rdi,%rcx,4), %rax
+; AVX1-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
+; AVX1-NEXT:    movq %rax, %rdx
+; AVX1-NEXT:    andq %rcx, %rdx
+; AVX1-NEXT:    shrq %rax
+; AVX1-NEXT:    andq %rcx, %rax
+; AVX1-NEXT:    leaq (%rax,%rdx,2), %rax
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_bitreverse_i64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vmovq %rdi, %xmm0
+; AVX2-NEXT:    vpbroadcastq %xmm0, %ymm0
+; AVX2-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[7,7,7,7,7,7,7,7,6,6,6,6,6,6,6,6,21,21,21,21,21,21,21,21,20,20,20,20,20,20,20,20]
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX2-NEXT:    vpand %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    vpmovmskb %ymm1, %ecx
+; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vpmovmskb %ymm0, %eax
+; AVX2-NEXT:    shlq $32, %rax
+; AVX2-NEXT:    orq %rcx, %rax
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; AVX512F-LABEL: test_bitreverse_i64:
+; AVX512F:       # %bb.0:
+; AVX512F-NEXT:    vmovq %rdi, %xmm0
+; AVX512F-NEXT:    vpbroadcastq %xmm0, %ymm0
+; AVX512F-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[7,7,7,7,7,7,7,7,6,6,6,6,6,6,6,6,21,21,21,21,21,21,21,21,20,20,20,20,20,20,20,20]
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX512F-NEXT:    vpand %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpmovmskb %ymm1, %ecx
+; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,3,3,3,3,3,3,3,2,2,2,2,2,2,2,2,17,17,17,17,17,17,17,17,16,16,16,16,16,16,16,16]
+; AVX512F-NEXT:    vpand %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT:    vpmovmskb %ymm0, %eax
+; AVX512F-NEXT:    shlq $32, %rax
+; AVX512F-NEXT:    orq %rcx, %rax
+; AVX512F-NEXT:    vzeroupper
+; AVX512F-NEXT:    retq
+;
+; AVX512BW-LABEL: test_bitreverse_i64:
+; AVX512BW:       # %bb.0:
+; AVX512BW-NEXT:    vpbroadcastq %rdi, %zmm0
+; AVX512BW-NEXT:    vpshufb {{.*#+}} zmm0 = zmm0[7,7,7,7,7,7,7,7,6,6,6,6,6,6,6,6,21,21,21,21,21,21,21,21,20,20,20,20,20,20,20,20,35,35,35,35,35,35,35,35,34,34,34,34,34,34,34,34,49,49,49,49,49,49,49,49,48,48,48,48,48,48,48,48]
+; AVX512BW-NEXT:    vpbroadcastq {{.*#+}} zmm1 = [72624976668147840,72624976668147840,72624976668147840,72624976668147840,72624976668147840,72624976668147840,72624976668147840,72624976668147840]
+; AVX512BW-NEXT:    vptestmb %zmm1, %zmm0, %k0
+; AVX512BW-NEXT:    kmovq %k0, %rax
+; AVX512BW-NEXT:    vzeroupper
+; AVX512BW-NEXT:    retq
 ;
 ; XOP-LABEL: test_bitreverse_i64:
 ; XOP:       # %bb.0:



More information about the llvm-commits mailing list