[llvm] [AArch64] Lower extending sitofp using tbl (PR #92528)

Momchil Velikov via llvm-commits llvm-commits at lists.llvm.org
Fri May 31 03:40:05 PDT 2024


https://github.com/momchil-velikov updated https://github.com/llvm/llvm-project/pull/92528

>From 16166436a8e9c2fa528f9bd0a47d8251362a3c8c Mon Sep 17 00:00:00 2001
From: Momchil Velikov <momchil.velikov at arm.com>
Date: Fri, 17 May 2024 11:50:31 +0100
Subject: [PATCH 1/3] [AArch64] Refactor creation of a shuffle mask for TBL
 (NFC)

---
 .../Target/AArch64/AArch64ISelLowering.cpp    | 88 +++++++++++--------
 1 file changed, 50 insertions(+), 38 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index ac6f1e07c4184..7f64d09c00f1a 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -15756,48 +15756,51 @@ bool AArch64TargetLowering::shouldSinkOperands(
   return false;
 }
 
-static bool createTblShuffleForZExt(ZExtInst *ZExt, FixedVectorType *DstTy,
-                                    bool IsLittleEndian) {
-  Value *Op = ZExt->getOperand(0);
-  auto *SrcTy = cast<FixedVectorType>(Op->getType());
-  auto SrcWidth = cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
-  auto DstWidth = cast<IntegerType>(DstTy->getElementType())->getBitWidth();
+static bool createTblShuffleMask(unsigned SrcWidth, unsigned DstWidth,
+                                 unsigned NumElts, bool IsLittleEndian,
+                                 SmallVectorImpl<int> &Mask) {
   if (DstWidth % 8 != 0 || DstWidth <= 16 || DstWidth >= 64)
     return false;
 
-  assert(DstWidth % SrcWidth == 0 &&
-         "TBL lowering is not supported for a ZExt instruction with this "
-         "source & destination element type.");
-  unsigned ZExtFactor = DstWidth / SrcWidth;
+  if (DstWidth % SrcWidth != 0)
+    return false;
+
+  unsigned Factor = DstWidth / SrcWidth;
+  unsigned MaskLen = NumElts * Factor;
+
+  Mask.clear();
+  Mask.resize(MaskLen, NumElts);
+
+  unsigned SrcIndex = 0;
+  for (unsigned I = 0; I < MaskLen; I += Factor)
+    Mask[I] = SrcIndex++;
+
+  if (!IsLittleEndian)
+    std::rotate(Mask.rbegin(), Mask.rbegin() + Factor - 1, Mask.rend());
+
+  return true;
+}
+
+static Value *createTblShuffleForZExt(IRBuilderBase &Builder, Value *Op,
+                                      FixedVectorType *ZExtTy,
+                                      FixedVectorType *DstTy,
+                                      bool IsLittleEndian) {
+  auto *SrcTy = cast<FixedVectorType>(Op->getType());
   unsigned NumElts = SrcTy->getNumElements();
-  IRBuilder<> Builder(ZExt);
+  auto SrcWidth = cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
+  auto DstWidth = cast<IntegerType>(DstTy->getElementType())->getBitWidth();
+
   SmallVector<int> Mask;
-  // Create a mask that selects <0,...,Op[i]> for each lane of the destination
-  // vector to replace the original ZExt. This can later be lowered to a set of
-  // tbl instructions.
-  for (unsigned i = 0; i < NumElts * ZExtFactor; i++) {
-    if (IsLittleEndian) {
-      if (i % ZExtFactor == 0)
-        Mask.push_back(i / ZExtFactor);
-      else
-        Mask.push_back(NumElts);
-    } else {
-      if ((i + 1) % ZExtFactor == 0)
-        Mask.push_back((i - ZExtFactor + 1) / ZExtFactor);
-      else
-        Mask.push_back(NumElts);
-    }
-  }
+  if (!createTblShuffleMask(SrcWidth, DstWidth, NumElts, IsLittleEndian, Mask))
+    return nullptr;
 
   auto *FirstEltZero = Builder.CreateInsertElement(
       PoisonValue::get(SrcTy), Builder.getInt8(0), uint64_t(0));
   Value *Result = Builder.CreateShuffleVector(Op, FirstEltZero, Mask);
   Result = Builder.CreateBitCast(Result, DstTy);
-  if (DstTy != ZExt->getType())
-    Result = Builder.CreateZExt(Result, ZExt->getType());
-  ZExt->replaceAllUsesWith(Result);
-  ZExt->eraseFromParent();
-  return true;
+  if (DstTy != ZExtTy)
+    Result = Builder.CreateZExt(Result, ZExtTy);
+  return Result;
 }
 
 static void createTblForTrunc(TruncInst *TI, bool IsLittleEndian) {
@@ -15962,21 +15965,30 @@ bool AArch64TargetLowering::optimizeExtendOrTruncateConversion(
 
       DstTy = TruncDstType;
     }
-
-    return createTblShuffleForZExt(ZExt, DstTy, Subtarget->isLittleEndian());
+    IRBuilder<> Builder(ZExt);
+    Value *Result = createTblShuffleForZExt(
+        Builder, ZExt->getOperand(0), cast<FixedVectorType>(ZExt->getType()),
+        DstTy, Subtarget->isLittleEndian());
+    if (!Result)
+      return false;
+    ZExt->replaceAllUsesWith(Result);
+    ZExt->eraseFromParent();
+    return true;
   }
 
   auto *UIToFP = dyn_cast<UIToFPInst>(I);
   if (UIToFP && SrcTy->getElementType()->isIntegerTy(8) &&
       DstTy->getElementType()->isFloatTy()) {
     IRBuilder<> Builder(I);
-    auto *ZExt = cast<ZExtInst>(
-        Builder.CreateZExt(I->getOperand(0), VectorType::getInteger(DstTy)));
+    Value *ZExt = createTblShuffleForZExt(
+        Builder, I->getOperand(0), FixedVectorType::getInteger(DstTy),
+        FixedVectorType::getInteger(DstTy), Subtarget->isLittleEndian());
+    if (!ZExt)
+      return false;
     auto *UI = Builder.CreateUIToFP(ZExt, DstTy);
     I->replaceAllUsesWith(UI);
     I->eraseFromParent();
-    return createTblShuffleForZExt(ZExt, cast<FixedVectorType>(ZExt->getType()),
-                                   Subtarget->isLittleEndian());
+    return true;
   }
 
   // Convert 'fptoui <(8|16) x float> to <(8|16) x i8>' to a wide fptoui

>From af2dab51fb52bcdce46c0774cadbc9b69bd7f8b1 Mon Sep 17 00:00:00 2001
From: Momchil Velikov <momchil.velikov at arm.com>
Date: Tue, 28 May 2024 14:35:33 +0100
Subject: [PATCH 2/3] [fixup] Simplify code

Change-Id: I0e51f00113d0692d5d3751c89eac8167511d1d9e
---
 llvm/lib/Target/AArch64/AArch64ISelLowering.cpp | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 7f64d09c00f1a..04fbaf30a4b0d 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -15772,12 +15772,9 @@ static bool createTblShuffleMask(unsigned SrcWidth, unsigned DstWidth,
   Mask.resize(MaskLen, NumElts);
 
   unsigned SrcIndex = 0;
-  for (unsigned I = 0; I < MaskLen; I += Factor)
+  for (unsigned I = IsLittleEndian ? 0 : Factor - 1; I < MaskLen; I += Factor)
     Mask[I] = SrcIndex++;
 
-  if (!IsLittleEndian)
-    std::rotate(Mask.rbegin(), Mask.rbegin() + Factor - 1, Mask.rend());
-
   return true;
 }
 

>From e8e2ab0156ac14f1781cdc40b0adc5b6ce3cc40c Mon Sep 17 00:00:00 2001
From: Momchil Velikov <momchil.velikov at arm.com>
Date: Tue, 21 May 2024 15:57:16 +0100
Subject: [PATCH 3/3] [AArch64] Lower extending sitofp using tbl

In a similar manner as in https://reviews.llvm.org/D133494
use `TBL` to place bytes in the *upper* part of `i32` elements
and then convert to float using fixed-point `scvtf`, i.e.

    scvtf Vd.4s, Vn.4s, #24

Change-Id: Ib9df3e4243612cbee8560907b24b14e76b61f265
---
 llvm/lib/CodeGen/CodeGenPrepare.cpp           |   3 +-
 .../Target/AArch64/AArch64ISelLowering.cpp    |  35 ++++
 llvm/test/CodeGen/AArch64/sitofp-to-tbl.ll    | 196 ++++++++++++++++++
 3 files changed, 233 insertions(+), 1 deletion(-)
 create mode 100644 llvm/test/CodeGen/AArch64/sitofp-to-tbl.ll

diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
index 339a1f1f2f002..4c52dbfa23903 100644
--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -8333,7 +8333,8 @@ bool CodeGenPrepare::optimizeInst(Instruction *I, ModifyDT &ModifiedDT) {
     if (OptimizeNoopCopyExpression(CI, *TLI, *DL))
       return true;
 
-    if ((isa<UIToFPInst>(I) || isa<FPToUIInst>(I) || isa<TruncInst>(I)) &&
+    if ((isa<UIToFPInst>(I) || isa<SIToFPInst>(I) || isa<FPToUIInst>(I) ||
+         isa<TruncInst>(I)) &&
         TLI->optimizeExtendOrTruncateConversion(
             I, LI->getLoopFor(I->getParent()), *TTI))
       return true;
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 04fbaf30a4b0d..72ad160528501 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -15800,6 +15800,24 @@ static Value *createTblShuffleForZExt(IRBuilderBase &Builder, Value *Op,
   return Result;
 }
 
+static Value *createTblShuffleForSExt(IRBuilderBase &Builder, Value *Op,
+                                      FixedVectorType *DstTy,
+                                      bool IsLittleEndian) {
+  auto *SrcTy = cast<FixedVectorType>(Op->getType());
+  auto SrcWidth = cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
+  auto DstWidth = cast<IntegerType>(DstTy->getElementType())->getBitWidth();
+
+  SmallVector<int> Mask;
+  if (!createTblShuffleMask(SrcWidth, DstWidth, SrcTy->getNumElements(),
+                            !IsLittleEndian, Mask))
+    return nullptr;
+
+  auto *FirstEltZero = Builder.CreateInsertElement(
+      PoisonValue::get(SrcTy), Builder.getInt8(0), uint64_t(0));
+
+  return Builder.CreateShuffleVector(Op, FirstEltZero, Mask);
+}
+
 static void createTblForTrunc(TruncInst *TI, bool IsLittleEndian) {
   IRBuilder<> Builder(TI);
   SmallVector<Value *> Parts;
@@ -15988,6 +16006,23 @@ bool AArch64TargetLowering::optimizeExtendOrTruncateConversion(
     return true;
   }
 
+  auto *SIToFP = dyn_cast<SIToFPInst>(I);
+  if (SIToFP && SrcTy->getElementType()->isIntegerTy(8) &&
+      DstTy->getElementType()->isFloatTy()) {
+    IRBuilder<> Builder(I);
+    auto *Shuffle = createTblShuffleForSExt(Builder, I->getOperand(0),
+                                            FixedVectorType::getInteger(DstTy),
+                                            Subtarget->isLittleEndian());
+    if (!Shuffle)
+      return false;
+    auto *Cast = Builder.CreateBitCast(Shuffle, VectorType::getInteger(DstTy));
+    auto *AShr = Builder.CreateAShr(Cast, 24, "", true);
+    auto *SI = Builder.CreateSIToFP(AShr, DstTy);
+    I->replaceAllUsesWith(SI);
+    I->eraseFromParent();
+    return true;
+  }
+
   // Convert 'fptoui <(8|16) x float> to <(8|16) x i8>' to a wide fptoui
   // followed by a truncate lowered to using tbl.4.
   auto *FPToUI = dyn_cast<FPToUIInst>(I);
diff --git a/llvm/test/CodeGen/AArch64/sitofp-to-tbl.ll b/llvm/test/CodeGen/AArch64/sitofp-to-tbl.ll
new file mode 100644
index 0000000000000..b7e8a3a3defe4
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sitofp-to-tbl.ll
@@ -0,0 +1,196 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -verify-machineinstrs < %s | FileCheck %s
+
+target triple = "aarch64-linux"
+
+; CHECK-LABEL: .LCPI0_0:
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    4
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    5
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    6
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    7
+; CHECK-NEXT:  .LCPI0_1:
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    0
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    1
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    2
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+; CHECK-NEXT:  .byte    255
+
+define void @sitofp_v8i8_to_v8f32(ptr %src, ptr %dst) {
+; CHECK-LABEL: sitofp_v8i8_to_v8f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    adrp x8, .LCPI0_0
+; CHECK-NEXT:    adrp x9, .LCPI0_1
+; CHECK-NEXT:    ldr q0, [x8, :lo12:.LCPI0_0]
+; CHECK-NEXT:    ldr q1, [x9, :lo12:.LCPI0_1]
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  .LBB0_1: // %loop
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr d2, [x0, x8, lsl #3]
+; CHECK-NEXT:    add x9, x1, x8, lsl #5
+; CHECK-NEXT:    add x8, x8, #1
+; CHECK-NEXT:    cmp x8, #1000
+; CHECK-NEXT:    tbl v3.16b, { v2.16b }, v0.16b
+; CHECK-NEXT:    tbl v2.16b, { v2.16b }, v1.16b
+; CHECK-NEXT:    scvtf v3.4s, v3.4s, #24
+; CHECK-NEXT:    scvtf v2.4s, v2.4s, #24
+; CHECK-NEXT:    stp q2, q3, [x9]
+; CHECK-NEXT:    b.eq .LBB0_1
+; CHECK-NEXT:  // %bb.2: // %exit
+; CHECK-NEXT:    ret
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep.src = getelementptr inbounds <8 x i8>, ptr %src, i64 %iv
+  %l = load <8 x i8>, ptr %gep.src
+  %conv = sitofp <8 x i8> %l to <8 x float>
+  %gep.dst = getelementptr inbounds <8 x float>, ptr %dst, i64 %iv
+  store <8 x float> %conv, ptr %gep.dst
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, 1000
+  br i1 %ec, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+; CHECK-LABEL: .LCPI1_0:
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    12
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    13
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    14
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    15
+; CHECK-NEXT: .LCPI1_1:
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    8
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    9
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    10
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    11
+; CHECK-NEXT: .LCPI1_2:
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    4
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    5
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    6
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    7
+; CHECK-NEXT: .LCPI1_3:
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    0
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    1
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    2
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    255
+; CHECK-NEXT: .byte    3
+
+define void @sitofp_v16i8_to_v16f32(ptr %src, ptr %dst) {
+; CHECK-LABEL: sitofp_v16i8_to_v16f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    adrp x8, .LCPI1_0
+; CHECK-NEXT:    adrp x9, .LCPI1_1
+; CHECK-NEXT:    adrp x10, .LCPI1_2
+; CHECK-NEXT:    ldr q0, [x8, :lo12:.LCPI1_0]
+; CHECK-NEXT:    adrp x8, .LCPI1_3
+; CHECK-NEXT:    ldr q1, [x9, :lo12:.LCPI1_1]
+; CHECK-NEXT:    ldr q2, [x10, :lo12:.LCPI1_2]
+; CHECK-NEXT:    ldr q3, [x8, :lo12:.LCPI1_3]
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  .LBB1_1: // %loop
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr q4, [x0, x8, lsl #4]
+; CHECK-NEXT:    add x9, x1, x8, lsl #6
+; CHECK-NEXT:    add x8, x8, #1
+; CHECK-NEXT:    cmp x8, #1000
+; CHECK-NEXT:    tbl v5.16b, { v4.16b }, v0.16b
+; CHECK-NEXT:    tbl v6.16b, { v4.16b }, v1.16b
+; CHECK-NEXT:    tbl v7.16b, { v4.16b }, v2.16b
+; CHECK-NEXT:    tbl v4.16b, { v4.16b }, v3.16b
+; CHECK-NEXT:    scvtf v5.4s, v5.4s, #24
+; CHECK-NEXT:    scvtf v6.4s, v6.4s, #24
+; CHECK-NEXT:    scvtf v7.4s, v7.4s, #24
+; CHECK-NEXT:    scvtf v4.4s, v4.4s, #24
+; CHECK-NEXT:    stp q6, q5, [x9, #32]
+; CHECK-NEXT:    stp q4, q7, [x9]
+; CHECK-NEXT:    b.eq .LBB1_1
+; CHECK-NEXT:  // %bb.2: // %exit
+; CHECK-NEXT:    ret
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep.src = getelementptr inbounds <16 x i8>, ptr %src, i64 %iv
+  %l = load <16 x i8>, ptr %gep.src
+  %conv = sitofp <16 x i8> %l to <16 x float>
+  %gep.dst = getelementptr inbounds <16 x float>, ptr %dst, i64 %iv
+  store <16 x float> %conv, ptr %gep.dst
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, 1000
+  br i1 %ec, label %loop, label %exit
+
+exit:
+  ret void
+}



More information about the llvm-commits mailing list