Skip to content

[RISCV] Attempt to widen SEW before generic shuffle lowering #122311

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 3 commits into from
Jan 10, 2025
Merged

[RISCV] Attempt to widen SEW before generic shuffle lowering #122311

merged 3 commits into from
Jan 10, 2025

Conversation

@preames
Copy link
Collaborator

@preames preames commented Jan 9, 2025

This takes inspiration from AArch64 which does the same thing to assist with zip/trn/etc.. Doing this recursion unconditionally when the mask allows is slightly questionable, but seems to work out okay in practice. If reviewers would rather see a more narrow heuristic for widening, let me know - and please suggest a heuristic which makes sense to you.

My actual motivation for this involves matching the proposed zip/unzip/zipeven/zipodd instructions being discussed on sig-vector, but this also applies to other shuffle masks as well.

As a bit of context, it's helpful to realize that we have existing logic in both DAGCombine and InstCombine which mutates the element width of in an analogous manner. However, that code has two restriction which prevent it from handling the motivating cases here. First, it only triggers if there is a bitcast involving a different element type. Second, the matcher used considers a partially undef wide element to be a non-match. I considered trying to relax those assumptions, but the information loss for undef in mid-level opt seemed more likely to open a can of worms than I wanted.

@preames
[RISCV] Attempt to widen SEW before generic shuffle lowering
024353a 
This takes inspiration from AArch64 which does the same thing to assist
with zip/trn/etc..  Doing this recursion unconditionally when the mask
allows is slightly questionable, but seems to work out okay in practice.
If reviewers would rather see a more narrow heuristic for widening, let
me know - and please suggest a heuristic which makes sense to you.

My actual motivation for this involves matching the proposed
zip/unzip/zipeven/zipodd instructions being discussed on sig-vector,
but this also applies to other shuffle masks as well.

As a bit of context, it's helpful to realize that we have existing
logic in both DAGCombine and InstCombine which mutates the element
width of in an analogous manner.  However, that code has two
restriction which prevent it from handling the motivating cases here.
First, it only triggers if there is a bitcast involving a different
element type.  Second, the matcher used considers a partially undef
wide element to be a non-match.  I considered trying to relax those
assumptions, but the information loss for undef in mid-level opt
seemed more likely to open a can of worms than I wanted.
@preames preames requested review from lukel97 and topperc January 9, 2025 16:41
@llvmbot llvmbot added backend:AArch64 backend:RISC-V llvm:analysis Includes value tracking, cost tables and constant folding labels Jan 9, 2025
@llvmbot
Copy link
Member

llvmbot commented Jan 9, 2025 •
edited
Loading

@llvm/pr-subscribers-backend-risc-v

@llvm/pr-subscribers-backend-aarch64

Author: Philip Reames (preames)

Changes

This takes inspiration from AArch64 which does the same thing to assist with zip/trn/etc.. Doing this recursion unconditionally when the mask allows is slightly questionable, but seems to work out okay in practice. If reviewers would rather see a more narrow heuristic for widening, let me know - and please suggest a heuristic which makes sense to you.

My actual motivation for this involves matching the proposed zip/unzip/zipeven/zipodd instructions being discussed on sig-vector, but this also applies to other shuffle masks as well.

As a bit of context, it's helpful to realize that we have existing logic in both DAGCombine and InstCombine which mutates the element width of in an analogous manner. However, that code has two restriction which prevent it from handling the motivating cases here. First, it only triggers if there is a bitcast involving a different element type. Second, the matcher used considers a partially undef wide element to be a non-match. I considered trying to relax those assumptions, but the information loss for undef in mid-level opt seemed more likely to open a can of worms than I wanted.

Patch is 48.16 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/122311.diff

8 Files Affected:

  • (modified) llvm/include/llvm/Analysis/VectorUtils.h (+6)
  • (modified) llvm/lib/Analysis/VectorUtils.cpp (+36)
  • (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+1-39)
  • (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+47)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll (+17-29)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll (+237-279)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-reverse.ll (+32-119)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-rotate.ll (+14-10)
diff --git a/llvm/include/llvm/Analysis/VectorUtils.h b/llvm/include/llvm/Analysis/VectorUtils.h
index b72efac0a4887f..d99a225874abda 100644
--- a/llvm/include/llvm/Analysis/VectorUtils.h
+++ b/llvm/include/llvm/Analysis/VectorUtils.h
@@ -235,6 +235,12 @@ void narrowShuffleMaskElts(int Scale, ArrayRef<int> Mask,
 bool widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
                           SmallVectorImpl<int> &ScaledMask);
 
+/// A variant of the previous method which is specialized for Scale=2, and
+/// treats -1 as undef and allows widening when a wider element is partially
+/// undef in the narrow form of the mask.  This transformation discards
+/// information about which bytes in the original shuffle were undef.
+bool widenShuffleMaskElts(ArrayRef<int> M, SmallVectorImpl<int> &NewMask);
+
 /// Attempt to narrow/widen the \p Mask shuffle mask to the \p NumDstElts target
 /// width. Internally this will call narrowShuffleMaskElts/widenShuffleMaskElts.
 /// This will assert unless NumDstElts is a multiple of Mask.size (or vice-versa).
diff --git a/llvm/lib/Analysis/VectorUtils.cpp b/llvm/lib/Analysis/VectorUtils.cpp
index 6c2502ce21cca5..9e7b907ace848e 100644
--- a/llvm/lib/Analysis/VectorUtils.cpp
+++ b/llvm/lib/Analysis/VectorUtils.cpp
@@ -479,6 +479,42 @@ bool llvm::widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
   return true;
 }
 
+bool llvm::widenShuffleMaskElts(ArrayRef<int> M,
+                                SmallVectorImpl<int> &NewMask) {
+  unsigned NumElts = M.size();;
+  if (NumElts % 2 != 0)
+    return false;
+
+  NewMask.clear();
+  for (unsigned i = 0; i < NumElts; i += 2) {
+    int M0 = M[i];
+    int M1 = M[i + 1];
+
+    // If both elements are undef, new mask is undef too.
+    if (M0 == -1 && M1 == -1) {
+      NewMask.push_back(-1);
+      continue;
+    }
+
+    if (M0 == -1 && M1 != -1 && (M1 % 2) == 1) {
+      NewMask.push_back(M1 / 2);
+      continue;
+    }
+
+    if (M0 != -1 && (M0 % 2) == 0 && ((M0 + 1) == M1 || M1 == -1)) {
+      NewMask.push_back(M0 / 2);
+      continue;
+    }
+
+    NewMask.clear();
+    return false;
+  }
+
+  assert(NewMask.size() == NumElts / 2 && "Incorrect size for mask!");
+  return true;
+}
+
+
 bool llvm::scaleShuffleMaskElts(unsigned NumDstElts, ArrayRef<int> Mask,
                                 SmallVectorImpl<int> &ScaledMask) {
   unsigned NumSrcElts = Mask.size();
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 3ad2905ce52076..797b150d70f83a 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -13721,44 +13721,6 @@ static SDValue constructDup(SDValue V, int Lane, SDLoc dl, EVT VT,
   return DAG.getNode(Opcode, dl, VT, V, DAG.getConstant(Lane, dl, MVT::i64));
 }
 
-// Return true if we can get a new shuffle mask by checking the parameter mask
-// array to test whether every two adjacent mask values are continuous and
-// starting from an even number.
-static bool isWideTypeMask(ArrayRef<int> M, EVT VT,
-                           SmallVectorImpl<int> &NewMask) {
-  unsigned NumElts = VT.getVectorNumElements();
-  if (NumElts % 2 != 0)
-    return false;
-
-  NewMask.clear();
-  for (unsigned i = 0; i < NumElts; i += 2) {
-    int M0 = M[i];
-    int M1 = M[i + 1];
-
-    // If both elements are undef, new mask is undef too.
-    if (M0 == -1 && M1 == -1) {
-      NewMask.push_back(-1);
-      continue;
-    }
-
-    if (M0 == -1 && M1 != -1 && (M1 % 2) == 1) {
-      NewMask.push_back(M1 / 2);
-      continue;
-    }
-
-    if (M0 != -1 && (M0 % 2) == 0 && ((M0 + 1) == M1 || M1 == -1)) {
-      NewMask.push_back(M0 / 2);
-      continue;
-    }
-
-    NewMask.clear();
-    return false;
-  }
-
-  assert(NewMask.size() == NumElts / 2 && "Incorrect size for mask!");
-  return true;
-}
-
 // Try to widen element type to get a new mask value for a better permutation
 // sequence, so that we can use NEON shuffle instructions, such as zip1/2,
 // UZP1/2, TRN1/2, REV, INS, etc.
@@ -13785,7 +13747,7 @@ static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
     return SDValue();
 
   SmallVector<int, 8> NewMask;
-  if (isWideTypeMask(Mask, VT, NewMask)) {
+  if (widenShuffleMaskElts(Mask, NewMask)) {
     MVT NewEltVT = VT.isFloatingPoint()
                        ? MVT::getFloatingPointVT(ElementSize * 2)
                        : MVT::getIntegerVT(ElementSize * 2);
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 2eeca45ac414bd..aff09925dfe478 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -5261,6 +5261,42 @@ static SDValue lowerDisjointIndicesShuffle(ShuffleVectorSDNode *SVN,
   return DAG.getVectorShuffle(VT, DL, Select, DAG.getUNDEF(VT), NewMask);
 }
 
+/// Try to widen element type to get a new mask value for a better permutation
+/// sequence.  This doesn't try to inspect the widened mask for profitability;
+/// we speculate the widened form is equal or better.  This has the effect of
+/// reducing mask constant sizes - allowing cheaper materialization sequences
+/// - and index sequence sizes - reducing register pressure and materialization
+/// cost, at the cost of (possibly) an extra VTYPE toggle.
+static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  EVT VT = Op.getValueType();
+  EVT ScalarVT = VT.getVectorElementType();
+  unsigned ElementSize = ScalarVT.getFixedSizeInBits();
+  SDValue V0 = Op.getOperand(0);
+  SDValue V1 = Op.getOperand(1);
+  ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
+
+  // Avoid wasted work leading to isTypeLegal check failing below
+  if (ElementSize > 32)
+    return SDValue();
+
+  SmallVector<int, 8> NewMask;
+  if (widenShuffleMaskElts(Mask, NewMask)) {
+    MVT NewEltVT = VT.isFloatingPoint()
+                       ? MVT::getFloatingPointVT(ElementSize * 2)
+                       : MVT::getIntegerVT(ElementSize * 2);
+    MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
+    if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
+      V0 = DAG.getBitcast(NewVT, V0);
+      V1 = DAG.getBitcast(NewVT, V1);
+      return DAG.getBitcast(VT,
+                            DAG.getVectorShuffle(NewVT, DL, V0, V1, NewMask));
+    }
+  }
+
+  return SDValue();
+}
+
 static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
                                    const RISCVSubtarget &Subtarget) {
   SDValue V1 = Op.getOperand(0);
@@ -5506,6 +5542,11 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
     if (SDValue V = lowerVECTOR_SHUFFLEAsRotate(SVN, DAG, Subtarget))
       return V;
 
+    // Before hitting generic lowering fallbacks, try to widen the mask
+    // to a wider SEW.
+    if (SDValue V = tryWidenMaskForShuffle(Op, DAG))
+      return V;
+
     // Can we generate a vcompress instead of a vrgather?  These scale better
     // at high LMUL, at the cost of not being able to fold a following select
     // into them.  The mask constants are also smaller than the index vector
@@ -5615,6 +5656,12 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
     if (SDValue V = lowerDisjointIndicesShuffle(SVN, DAG, Subtarget))
       return V;
 
+
+  // Before hitting generic lowering fallbacks, try to widen the mask
+  // to a wider SEW.
+  if (SDValue V = tryWidenMaskForShuffle(Op, DAG))
+    return V;
+
   // Try to pick a profitable operand order.
   bool SwapOps = DAG.isSplatValue(V2) && !DAG.isSplatValue(V1);
   SwapOps = SwapOps ^ ShuffleVectorInst::isIdentityMask(ShuffleMaskRHS, NumElts);
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
index 0bd8466669dc80..00138744a7bda9 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
@@ -603,10 +603,8 @@ define <8 x i8> @concat_4xi8_start_undef(<8 x i8> %v, <8 x i8> %w) {
 define <8 x i8> @concat_4xi8_start_undef_at_start(<8 x i8> %v, <8 x i8> %w) {
 ; CHECK-LABEL: concat_4xi8_start_undef_at_start:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a0, -32
-; CHECK-NEXT:    vsetivli zero, 8, e8, mf2, ta, mu
-; CHECK-NEXT:    vmv.s.x v0, a0
-; CHECK-NEXT:    vslideup.vi v8, v9, 4, v0.t
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vslideup.vi v8, v9, 2
 ; CHECK-NEXT:    ret
   %res = shufflevector <8 x i8> %v, <8 x i8> %w, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 9, i32 10, i32 11>
   ret <8 x i8> %res
@@ -704,8 +702,8 @@ define <8 x i32> @shuffle_v8i32_2(<8 x i32> %x, <8 x i32> %y) {
 ; CHECK-LABEL: shuffle_v8i32_2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetivli zero, 1, e8, mf8, ta, ma
-; CHECK-NEXT:    vmv.v.i v0, -13
-; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vmv.v.i v0, 13
+; CHECK-NEXT:    vsetivli zero, 4, e64, m2, ta, ma
 ; CHECK-NEXT:    vmerge.vvm v8, v10, v8, v0
 ; CHECK-NEXT:    ret
   %s = shufflevector <8 x i32> %x, <8 x i32> %y, <8 x i32> <i32 0, i32 1, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7>
@@ -756,9 +754,9 @@ define <8 x i16> @shuffle_compress_singlesrc_e16(<8 x i16> %v) {
 define <8 x i32> @shuffle_compress_singlesrc_e32(<8 x i32> %v) {
 ; CHECK-LABEL: shuffle_compress_singlesrc_e32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a0, 115
-; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; CHECK-NEXT:    vmv.s.x v12, a0
+; CHECK-NEXT:    vsetivli zero, 1, e8, mf8, ta, ma
+; CHECK-NEXT:    vmv.v.i v12, 13
+; CHECK-NEXT:    vsetivli zero, 4, e64, m2, ta, ma
 ; CHECK-NEXT:    vcompress.vm v10, v8, v12
 ; CHECK-NEXT:    vmv.v.v v8, v10
 ; CHECK-NEXT:    ret
@@ -832,26 +830,16 @@ define <8 x i32> @shuffle_spread2_singlesrc_e32_index2(<8 x i32> %v) {
 }
 
 define <8 x i32> @shuffle_spread3_singlesrc_e32(<8 x i32> %v) {
-; RV32-LABEL: shuffle_spread3_singlesrc_e32:
-; RV32:       # %bb.0:
-; RV32-NEXT:    lui a0, %hi(.LCPI57_0)
-; RV32-NEXT:    addi a0, a0, %lo(.LCPI57_0)
-; RV32-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; RV32-NEXT:    vle16.v v12, (a0)
-; RV32-NEXT:    vrgatherei16.vv v10, v8, v12
-; RV32-NEXT:    vmv.v.v v8, v10
-; RV32-NEXT:    ret
-;
-; RV64-LABEL: shuffle_spread3_singlesrc_e32:
-; RV64:       # %bb.0:
-; RV64-NEXT:    lui a0, 32769
-; RV64-NEXT:    slli a0, a0, 21
-; RV64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
-; RV64-NEXT:    vmv.v.x v12, a0
-; RV64-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; RV64-NEXT:    vrgatherei16.vv v10, v8, v12
-; RV64-NEXT:    vmv.v.v v8, v10
-; RV64-NEXT:    ret
+; CHECK-LABEL: shuffle_spread3_singlesrc_e32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vmv.v.i v10, 0
+; CHECK-NEXT:    li a0, 1
+; CHECK-NEXT:    vslide1down.vx v12, v10, a0
+; CHECK-NEXT:    vsetvli zero, zero, e64, m2, ta, ma
+; CHECK-NEXT:    vrgatherei16.vv v10, v8, v12
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
   %out = shufflevector <8 x i32> %v, <8 x i32> poison, <8 x i32> <i32 0, i32 undef, i32 undef, i32 1, i32 undef, i32 undef, i32 2, i32 undef>
   ret <8 x i32> %out
 }
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
index 67d649902b022a..0c7d7925edf39c 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
@@ -183,498 +183,456 @@ define {<8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>} @load_
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    csrr a2, vlenb
-; RV32-NEXT:    li a3, 96
+; RV32-NEXT:    li a3, 92
 ; RV32-NEXT:    mul a2, a2, a3
 ; RV32-NEXT:    sub sp, sp, a2
-; RV32-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0xe0, 0x00, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 96 * vlenb
-; RV32-NEXT:    addi a3, a1, 128
-; RV32-NEXT:    addi a4, a1, 256
+; RV32-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0xdc, 0x00, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 92 * vlenb
+; RV32-NEXT:    addi a3, a1, 256
+; RV32-NEXT:    addi a4, a1, 128
 ; RV32-NEXT:    li a2, 32
-; RV32-NEXT:    li a5, 768
-; RV32-NEXT:    lui a6, 12291
-; RV32-NEXT:    lui a7, %hi(.LCPI8_1)
-; RV32-NEXT:    addi a7, a7, %lo(.LCPI8_1)
+; RV32-NEXT:    lui a5, 12291
+; RV32-NEXT:    lui a6, %hi(.LCPI8_0)
+; RV32-NEXT:    addi a6, a6, %lo(.LCPI8_0)
+; RV32-NEXT:    li a7, 768
+; RV32-NEXT:    lui t0, 49164
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
+; RV32-NEXT:    vle32.v v16, (a1)
+; RV32-NEXT:    csrr a1, vlenb
+; RV32-NEXT:    li t1, 76
+; RV32-NEXT:    mul a1, a1, t1
+; RV32-NEXT:    add a1, sp, a1
+; RV32-NEXT:    addi a1, a1, 16
+; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    vle32.v v8, (a4)
-; RV32-NEXT:    csrr a4, vlenb
-; RV32-NEXT:    li t0, 88
-; RV32-NEXT:    mul a4, a4, t0
-; RV32-NEXT:    add a4, sp, a4
-; RV32-NEXT:    addi a4, a4, 16
-; RV32-NEXT:    vs8r.v v8, (a4) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv.s.x v0, a5
-; RV32-NEXT:    vle32.v v24, (a1)
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a4, 72
+; RV32-NEXT:    li a4, 68
 ; RV32-NEXT:    mul a1, a1, a4
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v24, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    addi a5, a5, 3
+; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
+; RV32-NEXT:    vle16.v v6, (a6)
+; RV32-NEXT:    vmv.s.x v0, a5
+; RV32-NEXT:    lui a1, %hi(.LCPI8_1)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI8_1)
+; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
+; RV32-NEXT:    vmerge.vvm v16, v8, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v24, v16, v6
+; RV32-NEXT:    csrr a4, vlenb
+; RV32-NEXT:    li a5, 52
+; RV32-NEXT:    mul a4, a4, a5
+; RV32-NEXT:    add a4, sp, a4
+; RV32-NEXT:    addi a4, a4, 16
+; RV32-NEXT:    vs8r.v v24, (a4) # Unknown-size Folded Spill
+; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
 ; RV32-NEXT:    vle32.v v16, (a3)
+; RV32-NEXT:    addi t0, t0, 12
+; RV32-NEXT:    vmv.s.x v0, a7
+; RV32-NEXT:    vmv.s.x v7, t0
+; RV32-NEXT:    vsetivli zero, 16, e16, m2, ta, ma
+; RV32-NEXT:    vle16.v v4, (a1)
+; RV32-NEXT:    vsetivli zero, 16, e32, m8, ta, ma
+; RV32-NEXT:    vslidedown.vi v24, v16, 16
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
+; RV32-NEXT:    li a3, 60
+; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    addi a6, a6, 3
-; RV32-NEXT:    vle16.v v4, (a7)
-; RV32-NEXT:    vmv.s.x v3, a6
-; RV32-NEXT:    vsetivli zero, 16, e32, m8, ta, ma
-; RV32-NEXT:    vslidedown.vi v16, v8, 16
+; RV32-NEXT:    vs8r.v v24, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80
+; RV32-NEXT:    li a3, 84
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
-; RV32-NEXT:    vmerge.vvm v8, v16, v8, v0
+; RV32-NEXT:    vmerge.vvm v20, v24, v16, v0
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 52
+; RV32-NEXT:    li a3, 40
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs4r.v v8, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv1r.v v0, v3
+; RV32-NEXT:    vs4r.v v20, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vmv1r.v v0, v7
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
+; RV32-NEXT:    li a3, 76
+; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vl8r.v v16, (a1) # Unknown-size Folded Reload
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
-; RV32-NEXT:    vmerge.vvm v16, v8, v24, v0
-; RV32-NEXT:    vrgatherei16.vv v8, v16, v4
+; RV32-NEXT:    vmerge.vvm v24, v8, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v8, v24, v4
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 56
+; RV32-NEXT:    li a3, 44
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    li a1, 3
-; RV32-NEXT:    lui a3, 49164
-; RV32-NEXT:    lui a4, %hi(.LCPI8_3)
-; RV32-NEXT:    addi a4, a4, %lo(.LCPI8_3)
+; RV32-NEXT:    lui a3, 196656
+; RV32-NEXT:    lui a4, %hi(.LCPI8_2)
+; RV32-NEXT:    addi a4, a4, %lo(.LCPI8_2)
 ; RV32-NEXT:    slli a1, a1, 10
-; RV32-NEXT:    addi a3, a3, 12
+; RV32-NEXT:    addi a3, a3, 48
 ; RV32-NEXT:    vmv.s.x v0, a1
-; RV32-NEXT:    vle16.v v16, (a4)
-; RV32-NEXT:    vmv.s.x v20, a3
+; RV32-NEXT:    vle16.v v14, (a4)
+; RV32-NEXT:    vmv.s.x v12, a3
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 88
+; RV32-NEXT:    li a3, 84
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vmv4r.v v8, v24
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80
+; RV32-NEXT:    li a3, 60
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
-; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
-; RV32-NEXT:    vmerge.vvm v8, v8, v24, v0
+; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
+; RV32-NEXT:    vmerge.vvm v8, v24, v8, v0
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 48
+; RV32-NEXT:    li a3, 24
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs4r.v v8, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv1r.v v0, v20
-; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
-; RV32-NEXT:    add a1, sp, a1
-; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vmv1r.v v0, v12
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 72
+; RV32-NEXT:    li a3, 68
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
-; RV32-NEXT:    vmerge.vvm v24, v8, v24, v0
-; RV32-NEXT:    vrgatherei16.vv v8, v24, v16
+; RV32-NEXT:    vmerge.vvm v24, v24, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v16, v24, v14
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 40
-; RV32-NEXT:    mul a1, a1, a3
+; RV32-NEXT:    slli a1, a1, 5
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    lui a1, 3
-; RV32-NEXT:    lui a3, 196656
+; RV32-NEXT:    lui a3, 786624
 ; RV32-NEXT:    lui a4, 12
-; RV32-NEXT:    lui a5, 786624
+; RV32-NEXT:    lui a5, 768
 ; RV32-NEXT:    li a6, 48
-; RV32-NEXT:    lui a7, 768
+; RV32-NEXT:    lui a7, 3073
 ; RV32-NEXT:    li t0, 192
 ; RV32-NEXT:    addi a1, a1, 3
-; RV32-NEXT:    addi a3, a3, 48
+; RV32-NEXT:    addi a3, a3, 192
 ; RV32-NEXT:    addi a4, a4, 12
-; RV32-NEXT:    addi a5, a5, 192
-; RV32-NEXT:    addi a7, a7, 768
+; RV32-NEXT:    addi a5, a5, 768
+; RV32-NEXT:    addi a7, a7, -1024
 ; RV32-NEXT:    vmv.s.x v1, a6
-; RV32-NEXT:    vmv.s.x v8, t0
-; RV32-NEXT:    addi a6, sp, 16
-; RV32-NEXT:    vs1r.v v8, (a6) # Unknown-size Folded Spill
+; RV32-NEXT:    vmv.s.x v12, t0
 ; RV32-NEXT:    vmv.s.x v0, a1
-; RV32-NEXT:    vmv.s.x v14, a3
-; RV32-NEXT:    vmv.s.x v7, a4
-; RV32-NEXT:    vmv.s.x v3, a5
-; RV32-NEXT:    vmv.s.x v2, a7
+; RV32-NEXT:    vmv.s.x v3, a3
+; RV32-NEXT:    vmv.s.x v2, a4
+; RV32-NEXT:    vmv.s.x v13, a5
+; RV32-NEXT:    vmv.s.x v14, a7
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80...
[truncated]

@llvmbot
Copy link
Member

llvmbot commented Jan 9, 2025

@llvm/pr-subscribers-llvm-analysis

Author: Philip Reames (preames)

Changes

This takes inspiration from AArch64 which does the same thing to assist with zip/trn/etc.. Doing this recursion unconditionally when the mask allows is slightly questionable, but seems to work out okay in practice. If reviewers would rather see a more narrow heuristic for widening, let me know - and please suggest a heuristic which makes sense to you.

My actual motivation for this involves matching the proposed zip/unzip/zipeven/zipodd instructions being discussed on sig-vector, but this also applies to other shuffle masks as well.

As a bit of context, it's helpful to realize that we have existing logic in both DAGCombine and InstCombine which mutates the element width of in an analogous manner. However, that code has two restriction which prevent it from handling the motivating cases here. First, it only triggers if there is a bitcast involving a different element type. Second, the matcher used considers a partially undef wide element to be a non-match. I considered trying to relax those assumptions, but the information loss for undef in mid-level opt seemed more likely to open a can of worms than I wanted.

Patch is 48.16 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/122311.diff

8 Files Affected:

  • (modified) llvm/include/llvm/Analysis/VectorUtils.h (+6)
  • (modified) llvm/lib/Analysis/VectorUtils.cpp (+36)
  • (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+1-39)
  • (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+47)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll (+17-29)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll (+237-279)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-reverse.ll (+32-119)
  • (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-rotate.ll (+14-10)
diff --git a/llvm/include/llvm/Analysis/VectorUtils.h b/llvm/include/llvm/Analysis/VectorUtils.h
index b72efac0a4887f..d99a225874abda 100644
--- a/llvm/include/llvm/Analysis/VectorUtils.h
+++ b/llvm/include/llvm/Analysis/VectorUtils.h
@@ -235,6 +235,12 @@ void narrowShuffleMaskElts(int Scale, ArrayRef<int> Mask,
 bool widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
                           SmallVectorImpl<int> &ScaledMask);
 
+/// A variant of the previous method which is specialized for Scale=2, and
+/// treats -1 as undef and allows widening when a wider element is partially
+/// undef in the narrow form of the mask.  This transformation discards
+/// information about which bytes in the original shuffle were undef.
+bool widenShuffleMaskElts(ArrayRef<int> M, SmallVectorImpl<int> &NewMask);
+
 /// Attempt to narrow/widen the \p Mask shuffle mask to the \p NumDstElts target
 /// width. Internally this will call narrowShuffleMaskElts/widenShuffleMaskElts.
 /// This will assert unless NumDstElts is a multiple of Mask.size (or vice-versa).
diff --git a/llvm/lib/Analysis/VectorUtils.cpp b/llvm/lib/Analysis/VectorUtils.cpp
index 6c2502ce21cca5..9e7b907ace848e 100644
--- a/llvm/lib/Analysis/VectorUtils.cpp
+++ b/llvm/lib/Analysis/VectorUtils.cpp
@@ -479,6 +479,42 @@ bool llvm::widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
   return true;
 }
 
+bool llvm::widenShuffleMaskElts(ArrayRef<int> M,
+                                SmallVectorImpl<int> &NewMask) {
+  unsigned NumElts = M.size();;
+  if (NumElts % 2 != 0)
+    return false;
+
+  NewMask.clear();
+  for (unsigned i = 0; i < NumElts; i += 2) {
+    int M0 = M[i];
+    int M1 = M[i + 1];
+
+    // If both elements are undef, new mask is undef too.
+    if (M0 == -1 && M1 == -1) {
+      NewMask.push_back(-1);
+      continue;
+    }
+
+    if (M0 == -1 && M1 != -1 && (M1 % 2) == 1) {
+      NewMask.push_back(M1 / 2);
+      continue;
+    }
+
+    if (M0 != -1 && (M0 % 2) == 0 && ((M0 + 1) == M1 || M1 == -1)) {
+      NewMask.push_back(M0 / 2);
+      continue;
+    }
+
+    NewMask.clear();
+    return false;
+  }
+
+  assert(NewMask.size() == NumElts / 2 && "Incorrect size for mask!");
+  return true;
+}
+
+
 bool llvm::scaleShuffleMaskElts(unsigned NumDstElts, ArrayRef<int> Mask,
                                 SmallVectorImpl<int> &ScaledMask) {
   unsigned NumSrcElts = Mask.size();
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 3ad2905ce52076..797b150d70f83a 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -13721,44 +13721,6 @@ static SDValue constructDup(SDValue V, int Lane, SDLoc dl, EVT VT,
   return DAG.getNode(Opcode, dl, VT, V, DAG.getConstant(Lane, dl, MVT::i64));
 }
 
-// Return true if we can get a new shuffle mask by checking the parameter mask
-// array to test whether every two adjacent mask values are continuous and
-// starting from an even number.
-static bool isWideTypeMask(ArrayRef<int> M, EVT VT,
-                           SmallVectorImpl<int> &NewMask) {
-  unsigned NumElts = VT.getVectorNumElements();
-  if (NumElts % 2 != 0)
-    return false;
-
-  NewMask.clear();
-  for (unsigned i = 0; i < NumElts; i += 2) {
-    int M0 = M[i];
-    int M1 = M[i + 1];
-
-    // If both elements are undef, new mask is undef too.
-    if (M0 == -1 && M1 == -1) {
-      NewMask.push_back(-1);
-      continue;
-    }
-
-    if (M0 == -1 && M1 != -1 && (M1 % 2) == 1) {
-      NewMask.push_back(M1 / 2);
-      continue;
-    }
-
-    if (M0 != -1 && (M0 % 2) == 0 && ((M0 + 1) == M1 || M1 == -1)) {
-      NewMask.push_back(M0 / 2);
-      continue;
-    }
-
-    NewMask.clear();
-    return false;
-  }
-
-  assert(NewMask.size() == NumElts / 2 && "Incorrect size for mask!");
-  return true;
-}
-
 // Try to widen element type to get a new mask value for a better permutation
 // sequence, so that we can use NEON shuffle instructions, such as zip1/2,
 // UZP1/2, TRN1/2, REV, INS, etc.
@@ -13785,7 +13747,7 @@ static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
     return SDValue();
 
   SmallVector<int, 8> NewMask;
-  if (isWideTypeMask(Mask, VT, NewMask)) {
+  if (widenShuffleMaskElts(Mask, NewMask)) {
     MVT NewEltVT = VT.isFloatingPoint()
                        ? MVT::getFloatingPointVT(ElementSize * 2)
                        : MVT::getIntegerVT(ElementSize * 2);
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 2eeca45ac414bd..aff09925dfe478 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -5261,6 +5261,42 @@ static SDValue lowerDisjointIndicesShuffle(ShuffleVectorSDNode *SVN,
   return DAG.getVectorShuffle(VT, DL, Select, DAG.getUNDEF(VT), NewMask);
 }
 
+/// Try to widen element type to get a new mask value for a better permutation
+/// sequence.  This doesn't try to inspect the widened mask for profitability;
+/// we speculate the widened form is equal or better.  This has the effect of
+/// reducing mask constant sizes - allowing cheaper materialization sequences
+/// - and index sequence sizes - reducing register pressure and materialization
+/// cost, at the cost of (possibly) an extra VTYPE toggle.
+static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  EVT VT = Op.getValueType();
+  EVT ScalarVT = VT.getVectorElementType();
+  unsigned ElementSize = ScalarVT.getFixedSizeInBits();
+  SDValue V0 = Op.getOperand(0);
+  SDValue V1 = Op.getOperand(1);
+  ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();
+
+  // Avoid wasted work leading to isTypeLegal check failing below
+  if (ElementSize > 32)
+    return SDValue();
+
+  SmallVector<int, 8> NewMask;
+  if (widenShuffleMaskElts(Mask, NewMask)) {
+    MVT NewEltVT = VT.isFloatingPoint()
+                       ? MVT::getFloatingPointVT(ElementSize * 2)
+                       : MVT::getIntegerVT(ElementSize * 2);
+    MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
+    if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
+      V0 = DAG.getBitcast(NewVT, V0);
+      V1 = DAG.getBitcast(NewVT, V1);
+      return DAG.getBitcast(VT,
+                            DAG.getVectorShuffle(NewVT, DL, V0, V1, NewMask));
+    }
+  }
+
+  return SDValue();
+}
+
 static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
                                    const RISCVSubtarget &Subtarget) {
   SDValue V1 = Op.getOperand(0);
@@ -5506,6 +5542,11 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
     if (SDValue V = lowerVECTOR_SHUFFLEAsRotate(SVN, DAG, Subtarget))
       return V;
 
+    // Before hitting generic lowering fallbacks, try to widen the mask
+    // to a wider SEW.
+    if (SDValue V = tryWidenMaskForShuffle(Op, DAG))
+      return V;
+
     // Can we generate a vcompress instead of a vrgather?  These scale better
     // at high LMUL, at the cost of not being able to fold a following select
     // into them.  The mask constants are also smaller than the index vector
@@ -5615,6 +5656,12 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
     if (SDValue V = lowerDisjointIndicesShuffle(SVN, DAG, Subtarget))
       return V;
 
+
+  // Before hitting generic lowering fallbacks, try to widen the mask
+  // to a wider SEW.
+  if (SDValue V = tryWidenMaskForShuffle(Op, DAG))
+    return V;
+
   // Try to pick a profitable operand order.
   bool SwapOps = DAG.isSplatValue(V2) && !DAG.isSplatValue(V1);
   SwapOps = SwapOps ^ ShuffleVectorInst::isIdentityMask(ShuffleMaskRHS, NumElts);
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
index 0bd8466669dc80..00138744a7bda9 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll
@@ -603,10 +603,8 @@ define <8 x i8> @concat_4xi8_start_undef(<8 x i8> %v, <8 x i8> %w) {
 define <8 x i8> @concat_4xi8_start_undef_at_start(<8 x i8> %v, <8 x i8> %w) {
 ; CHECK-LABEL: concat_4xi8_start_undef_at_start:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a0, -32
-; CHECK-NEXT:    vsetivli zero, 8, e8, mf2, ta, mu
-; CHECK-NEXT:    vmv.s.x v0, a0
-; CHECK-NEXT:    vslideup.vi v8, v9, 4, v0.t
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vslideup.vi v8, v9, 2
 ; CHECK-NEXT:    ret
   %res = shufflevector <8 x i8> %v, <8 x i8> %w, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 9, i32 10, i32 11>
   ret <8 x i8> %res
@@ -704,8 +702,8 @@ define <8 x i32> @shuffle_v8i32_2(<8 x i32> %x, <8 x i32> %y) {
 ; CHECK-LABEL: shuffle_v8i32_2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetivli zero, 1, e8, mf8, ta, ma
-; CHECK-NEXT:    vmv.v.i v0, -13
-; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vmv.v.i v0, 13
+; CHECK-NEXT:    vsetivli zero, 4, e64, m2, ta, ma
 ; CHECK-NEXT:    vmerge.vvm v8, v10, v8, v0
 ; CHECK-NEXT:    ret
   %s = shufflevector <8 x i32> %x, <8 x i32> %y, <8 x i32> <i32 0, i32 1, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7>
@@ -756,9 +754,9 @@ define <8 x i16> @shuffle_compress_singlesrc_e16(<8 x i16> %v) {
 define <8 x i32> @shuffle_compress_singlesrc_e32(<8 x i32> %v) {
 ; CHECK-LABEL: shuffle_compress_singlesrc_e32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    li a0, 115
-; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; CHECK-NEXT:    vmv.s.x v12, a0
+; CHECK-NEXT:    vsetivli zero, 1, e8, mf8, ta, ma
+; CHECK-NEXT:    vmv.v.i v12, 13
+; CHECK-NEXT:    vsetivli zero, 4, e64, m2, ta, ma
 ; CHECK-NEXT:    vcompress.vm v10, v8, v12
 ; CHECK-NEXT:    vmv.v.v v8, v10
 ; CHECK-NEXT:    ret
@@ -832,26 +830,16 @@ define <8 x i32> @shuffle_spread2_singlesrc_e32_index2(<8 x i32> %v) {
 }
 
 define <8 x i32> @shuffle_spread3_singlesrc_e32(<8 x i32> %v) {
-; RV32-LABEL: shuffle_spread3_singlesrc_e32:
-; RV32:       # %bb.0:
-; RV32-NEXT:    lui a0, %hi(.LCPI57_0)
-; RV32-NEXT:    addi a0, a0, %lo(.LCPI57_0)
-; RV32-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; RV32-NEXT:    vle16.v v12, (a0)
-; RV32-NEXT:    vrgatherei16.vv v10, v8, v12
-; RV32-NEXT:    vmv.v.v v8, v10
-; RV32-NEXT:    ret
-;
-; RV64-LABEL: shuffle_spread3_singlesrc_e32:
-; RV64:       # %bb.0:
-; RV64-NEXT:    lui a0, 32769
-; RV64-NEXT:    slli a0, a0, 21
-; RV64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
-; RV64-NEXT:    vmv.v.x v12, a0
-; RV64-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; RV64-NEXT:    vrgatherei16.vv v10, v8, v12
-; RV64-NEXT:    vmv.v.v v8, v10
-; RV64-NEXT:    ret
+; CHECK-LABEL: shuffle_spread3_singlesrc_e32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vmv.v.i v10, 0
+; CHECK-NEXT:    li a0, 1
+; CHECK-NEXT:    vslide1down.vx v12, v10, a0
+; CHECK-NEXT:    vsetvli zero, zero, e64, m2, ta, ma
+; CHECK-NEXT:    vrgatherei16.vv v10, v8, v12
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
   %out = shufflevector <8 x i32> %v, <8 x i32> poison, <8 x i32> <i32 0, i32 undef, i32 undef, i32 1, i32 undef, i32 undef, i32 2, i32 undef>
   ret <8 x i32> %out
 }
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
index 67d649902b022a..0c7d7925edf39c 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll
@@ -183,498 +183,456 @@ define {<8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>} @load_
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    csrr a2, vlenb
-; RV32-NEXT:    li a3, 96
+; RV32-NEXT:    li a3, 92
 ; RV32-NEXT:    mul a2, a2, a3
 ; RV32-NEXT:    sub sp, sp, a2
-; RV32-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0xe0, 0x00, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 96 * vlenb
-; RV32-NEXT:    addi a3, a1, 128
-; RV32-NEXT:    addi a4, a1, 256
+; RV32-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0xdc, 0x00, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 92 * vlenb
+; RV32-NEXT:    addi a3, a1, 256
+; RV32-NEXT:    addi a4, a1, 128
 ; RV32-NEXT:    li a2, 32
-; RV32-NEXT:    li a5, 768
-; RV32-NEXT:    lui a6, 12291
-; RV32-NEXT:    lui a7, %hi(.LCPI8_1)
-; RV32-NEXT:    addi a7, a7, %lo(.LCPI8_1)
+; RV32-NEXT:    lui a5, 12291
+; RV32-NEXT:    lui a6, %hi(.LCPI8_0)
+; RV32-NEXT:    addi a6, a6, %lo(.LCPI8_0)
+; RV32-NEXT:    li a7, 768
+; RV32-NEXT:    lui t0, 49164
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
+; RV32-NEXT:    vle32.v v16, (a1)
+; RV32-NEXT:    csrr a1, vlenb
+; RV32-NEXT:    li t1, 76
+; RV32-NEXT:    mul a1, a1, t1
+; RV32-NEXT:    add a1, sp, a1
+; RV32-NEXT:    addi a1, a1, 16
+; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    vle32.v v8, (a4)
-; RV32-NEXT:    csrr a4, vlenb
-; RV32-NEXT:    li t0, 88
-; RV32-NEXT:    mul a4, a4, t0
-; RV32-NEXT:    add a4, sp, a4
-; RV32-NEXT:    addi a4, a4, 16
-; RV32-NEXT:    vs8r.v v8, (a4) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv.s.x v0, a5
-; RV32-NEXT:    vle32.v v24, (a1)
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a4, 72
+; RV32-NEXT:    li a4, 68
 ; RV32-NEXT:    mul a1, a1, a4
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v24, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    addi a5, a5, 3
+; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
+; RV32-NEXT:    vle16.v v6, (a6)
+; RV32-NEXT:    vmv.s.x v0, a5
+; RV32-NEXT:    lui a1, %hi(.LCPI8_1)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI8_1)
+; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
+; RV32-NEXT:    vmerge.vvm v16, v8, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v24, v16, v6
+; RV32-NEXT:    csrr a4, vlenb
+; RV32-NEXT:    li a5, 52
+; RV32-NEXT:    mul a4, a4, a5
+; RV32-NEXT:    add a4, sp, a4
+; RV32-NEXT:    addi a4, a4, 16
+; RV32-NEXT:    vs8r.v v24, (a4) # Unknown-size Folded Spill
+; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
 ; RV32-NEXT:    vle32.v v16, (a3)
+; RV32-NEXT:    addi t0, t0, 12
+; RV32-NEXT:    vmv.s.x v0, a7
+; RV32-NEXT:    vmv.s.x v7, t0
+; RV32-NEXT:    vsetivli zero, 16, e16, m2, ta, ma
+; RV32-NEXT:    vle16.v v4, (a1)
+; RV32-NEXT:    vsetivli zero, 16, e32, m8, ta, ma
+; RV32-NEXT:    vslidedown.vi v24, v16, 16
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
+; RV32-NEXT:    li a3, 60
+; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    addi a6, a6, 3
-; RV32-NEXT:    vle16.v v4, (a7)
-; RV32-NEXT:    vmv.s.x v3, a6
-; RV32-NEXT:    vsetivli zero, 16, e32, m8, ta, ma
-; RV32-NEXT:    vslidedown.vi v16, v8, 16
+; RV32-NEXT:    vs8r.v v24, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80
+; RV32-NEXT:    li a3, 84
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
-; RV32-NEXT:    vmerge.vvm v8, v16, v8, v0
+; RV32-NEXT:    vmerge.vvm v20, v24, v16, v0
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 52
+; RV32-NEXT:    li a3, 40
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs4r.v v8, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv1r.v v0, v3
+; RV32-NEXT:    vs4r.v v20, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vmv1r.v v0, v7
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
+; RV32-NEXT:    li a3, 76
+; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vl8r.v v16, (a1) # Unknown-size Folded Reload
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
-; RV32-NEXT:    vmerge.vvm v16, v8, v24, v0
-; RV32-NEXT:    vrgatherei16.vv v8, v16, v4
+; RV32-NEXT:    vmerge.vvm v24, v8, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v8, v24, v4
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 56
+; RV32-NEXT:    li a3, 44
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    li a1, 3
-; RV32-NEXT:    lui a3, 49164
-; RV32-NEXT:    lui a4, %hi(.LCPI8_3)
-; RV32-NEXT:    addi a4, a4, %lo(.LCPI8_3)
+; RV32-NEXT:    lui a3, 196656
+; RV32-NEXT:    lui a4, %hi(.LCPI8_2)
+; RV32-NEXT:    addi a4, a4, %lo(.LCPI8_2)
 ; RV32-NEXT:    slli a1, a1, 10
-; RV32-NEXT:    addi a3, a3, 12
+; RV32-NEXT:    addi a3, a3, 48
 ; RV32-NEXT:    vmv.s.x v0, a1
-; RV32-NEXT:    vle16.v v16, (a4)
-; RV32-NEXT:    vmv.s.x v20, a3
+; RV32-NEXT:    vle16.v v14, (a4)
+; RV32-NEXT:    vmv.s.x v12, a3
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 88
+; RV32-NEXT:    li a3, 84
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vmv4r.v v8, v24
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80
+; RV32-NEXT:    li a3, 60
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
-; RV32-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
-; RV32-NEXT:    vmerge.vvm v8, v8, v24, v0
+; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
+; RV32-NEXT:    vmerge.vvm v8, v24, v8, v0
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 48
+; RV32-NEXT:    li a3, 24
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vs4r.v v8, (a1) # Unknown-size Folded Spill
-; RV32-NEXT:    vmv1r.v v0, v20
-; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    slli a1, a1, 6
-; RV32-NEXT:    add a1, sp, a1
-; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vl8r.v v8, (a1) # Unknown-size Folded Reload
+; RV32-NEXT:    vmv1r.v v0, v12
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 72
+; RV32-NEXT:    li a3, 68
 ; RV32-NEXT:    mul a1, a1, a3
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
 ; RV32-NEXT:    vl8r.v v24, (a1) # Unknown-size Folded Reload
 ; RV32-NEXT:    vsetvli zero, a2, e32, m8, ta, ma
-; RV32-NEXT:    vmerge.vvm v24, v8, v24, v0
-; RV32-NEXT:    vrgatherei16.vv v8, v24, v16
+; RV32-NEXT:    vmerge.vvm v24, v24, v16, v0
+; RV32-NEXT:    vsetivli zero, 16, e64, m8, ta, ma
+; RV32-NEXT:    vrgatherei16.vv v16, v24, v14
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 40
-; RV32-NEXT:    mul a1, a1, a3
+; RV32-NEXT:    slli a1, a1, 5
 ; RV32-NEXT:    add a1, sp, a1
 ; RV32-NEXT:    addi a1, a1, 16
-; RV32-NEXT:    vs8r.v v8, (a1) # Unknown-size Folded Spill
+; RV32-NEXT:    vs8r.v v16, (a1) # Unknown-size Folded Spill
 ; RV32-NEXT:    lui a1, 3
-; RV32-NEXT:    lui a3, 196656
+; RV32-NEXT:    lui a3, 786624
 ; RV32-NEXT:    lui a4, 12
-; RV32-NEXT:    lui a5, 786624
+; RV32-NEXT:    lui a5, 768
 ; RV32-NEXT:    li a6, 48
-; RV32-NEXT:    lui a7, 768
+; RV32-NEXT:    lui a7, 3073
 ; RV32-NEXT:    li t0, 192
 ; RV32-NEXT:    addi a1, a1, 3
-; RV32-NEXT:    addi a3, a3, 48
+; RV32-NEXT:    addi a3, a3, 192
 ; RV32-NEXT:    addi a4, a4, 12
-; RV32-NEXT:    addi a5, a5, 192
-; RV32-NEXT:    addi a7, a7, 768
+; RV32-NEXT:    addi a5, a5, 768
+; RV32-NEXT:    addi a7, a7, -1024
 ; RV32-NEXT:    vmv.s.x v1, a6
-; RV32-NEXT:    vmv.s.x v8, t0
-; RV32-NEXT:    addi a6, sp, 16
-; RV32-NEXT:    vs1r.v v8, (a6) # Unknown-size Folded Spill
+; RV32-NEXT:    vmv.s.x v12, t0
 ; RV32-NEXT:    vmv.s.x v0, a1
-; RV32-NEXT:    vmv.s.x v14, a3
-; RV32-NEXT:    vmv.s.x v7, a4
-; RV32-NEXT:    vmv.s.x v3, a5
-; RV32-NEXT:    vmv.s.x v2, a7
+; RV32-NEXT:    vmv.s.x v3, a3
+; RV32-NEXT:    vmv.s.x v2, a4
+; RV32-NEXT:    vmv.s.x v13, a5
+; RV32-NEXT:    vmv.s.x v14, a7
 ; RV32-NEXT:    csrr a1, vlenb
-; RV32-NEXT:    li a3, 80...
[truncated]

@github-actions
Copy link

github-actions bot commented Jan 9, 2025 •
edited
Loading

✅ With the latest revision this PR passed the C/C++ code formatter.

@github-actions
Copy link

github-actions bot commented Jan 9, 2025 •
edited
Loading

⚠️ undef deprecator found issues in your code. ⚠️

You can test this locally with the following command: 
git diff -U0 --pickaxe-regex -S '([^a-zA-Z0-9#_-]undef[^a-zA-Z0-9_-]|UndefValue::get)' cdbba15c6cd53291358bf95a9a9057042fcbf163 7f83c935439ea1b4fb85dc5b9776899f8b39c7cd llvm/include/llvm/Analysis/VectorUtils.h llvm/lib/Analysis/VectorUtils.cpp llvm/lib/Target/AArch64/AArch64ISelLowering.cpp llvm/lib/Target/RISCV/RISCVISelLowering.cpp llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-shuffles.ll llvm/test/CodeGen/RISCV/rvv/fixed-vectors-interleaved-access.ll llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-reverse.ll llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-rotate.ll

The following files introduce new uses of undef:

  • llvm/include/llvm/Analysis/VectorUtils.h
  • llvm/lib/Analysis/VectorUtils.cpp

Undef is now deprecated and should only be used in the rare cases where no replacement is possible. For example, a load of uninitialized memory yields undef. You should use poison values for placeholders instead.

In tests, avoid using undef and having tests that trigger undefined behavior. If you need an operand with some unimportant value, you can add a new argument to the function and use that instead.

For example, this is considered a bad practice:

define void @fn() {
  ...
  br i1 undef, ...
}

Please use the following instead:

define void @fn(i1 %cond) {
  ...
  br i1 %cond, ...
}

Please refer to the Undefined Behavior Manual for more information.

topperc
topperc reviewed Jan 10, 2025
View reviewed changes
llvm/lib/Target/RISCV/RISCVISelLowering.cpp Outdated
/// cost, at the cost of (possibly) an extra VTYPE toggle.
static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
SDLoc DL(Op);
EVT VT = Op.getValueType();
Copy link
Collaborator

@topperc topperc Jan 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Use MVT. It results in less compiled code for getVectorElementType() and .getFixedSizeInBits().

lukel97
lukel97 approved these changes Jan 10, 2025
View reviewed changes
Copy link
Contributor

@lukel97 lukel97 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM + what Craig said. Test diff seems good so I don't think we need a heuristic. It looks like it has the benefit of shrinking the vrgather indices.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp Outdated
Comment on lines 5284 to 5295
if (widenShuffleMaskElts(Mask, NewMask)) {
MVT NewEltVT = VT.isFloatingPoint()
? MVT::getFloatingPointVT(ElementSize * 2)
: MVT::getIntegerVT(ElementSize * 2);
MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
V0 = DAG.getBitcast(NewVT, V0);
V1 = DAG.getBitcast(NewVT, V1);
return DAG.getBitcast(VT,
DAG.getVectorShuffle(NewVT, DL, V0, V1, NewMask));
}
}
Copy link
Contributor

@lukel97 lukel97 Jan 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit, structure these checks as early returns?

@preames preames merged commit 24bb180 into llvm:main Jan 10, 2025
7 of 8 checks passed
@preames preames deleted the pr-riscv-shuffle-widen-recursive branch January 10, 2025 15:12
BaiXilin pushed a commit to BaiXilin/llvm-project that referenced this pull request Jan 12, 2025
@preames @BaiXilin
[RISCV] Attempt to widen SEW before generic shuffle lowering (llvm#12…
2c85fbd 
…2311)

This takes inspiration from AArch64 which does the same thing to assist
with zip/trn/etc.. Doing this recursion unconditionally when the mask
allows is slightly questionable, but seems to work out okay in practice.

As a bit of context, it's helpful to realize that we have existing logic
in both DAGCombine and InstCombine which mutates the element width of in
an analogous manner. However, that code has two restriction which
prevent it from handling the motivating cases here. First, it only
triggers if there is a bitcast involving a different element type.
Second, the matcher used considers a partially undef wide element to be
a non-match. I considered trying to relax those assumptions, but the
information loss for undef in mid-level opt seemed more likely to open a
can of worms than I wanted.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@topperc topperc topperc left review comments

@lukel97 lukel97 lukel97 approved these changes

Assignees

No one assigned

Labels

backend:AArch64 backend:RISC-V llvm:analysis Includes value tracking, cost tables and constant folding

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@preames @llvmbot @lukel97 @topperc