[RISCV] Improve performCONCAT_VECTORCombine stride matching #68726
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-risc-v Author: Michael Maitland (michaelmaitland) ChangesI intend to commit the two commits in this patch independently: Full diff: https://github.com/llvm/llvm-project/pull/68726.diff 2 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 6be3fa71479be5c..955674bd13e1c79 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -27,6 +27,7 @@
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAGAddressAnalysis.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DiagnosticInfo.h"
@@ -13821,6 +13822,58 @@ static SDValue performCONCAT_VECTORSCombine(SDNode *N, SelectionDAG &DAG,
Align = std::min(Align, Ld->getAlign());
}
+ // If the load ptrs can be decomposed into a common (Base + Index) with a
+ // common constant stride, then return the constant stride. This matcher
+ // enables some additional optimization since BaseIndexOffset is capable of
+ // decomposing the load ptrs to (add (add Base, Index), Stride) instead of
+ // (add LastPtr, Stride) or (add NextPtr, Stride) that matchForwardStrided and
+ // matchReverseStrided use respectively.
+ auto matchConstantStride = [&DAG, &N](ArrayRef<SDUse> Loads) {
+ // Initialize match constraints based on the first load. Initialize
+ // ConstStride by taking the difference between the offset of the first two
+ // loads.
+ if (Loads.size() < 2)
+ return SDValue();
+ BaseIndexOffset BaseLdBIO =
+ BaseIndexOffset::match(cast<LoadSDNode>(Loads[0]), DAG);
+ BaseIndexOffset LastLdBIO =
+ BaseIndexOffset::match(cast<LoadSDNode>(Loads[1]), DAG);
+ bool AllValidOffset =
+ BaseLdBIO.hasValidOffset() && LastLdBIO.hasValidOffset();
+ if (!AllValidOffset)
+ return SDValue();
+ bool BaseIndexMatch = BaseLdBIO.equalBaseIndex(LastLdBIO, DAG);
+ if (!BaseIndexMatch)
+ return SDValue();
+ int64_t ConstStride = LastLdBIO.getOffset() - BaseLdBIO.getOffset();
+
+ // Check that constraints hold for all subsequent loads and the ConstStride
+ // is the same.
+ for (auto Idx : enumerate(Loads.drop_front(2))) {
+ auto *Ld = cast<LoadSDNode>(Idx.value());
+ BaseIndexOffset BIO = BaseIndexOffset::match(Ld, DAG);
+ AllValidOffset &= BIO.hasValidOffset();
+ if (!AllValidOffset)
+ return SDValue();
+ BaseIndexMatch |= BaseLdBIO.equalBaseIndex(BIO, DAG);
+ // Add 3 to index because the first two loads have been processed before
+ // the loop.
+ bool StrideMatches =
+ ConstStride == BIO.getOffset() - LastLdBIO.getOffset();
+ if (!BaseIndexMatch || !StrideMatches)
+ return SDValue();
+ LastLdBIO = BIO;
+ }
+
+ // The match is a success if all the constraints hold.
+ if (BaseIndexMatch && AllValidOffset)
+ return DAG.getConstant(
+ ConstStride, SDLoc(N),
+ cast<LoadSDNode>(N->getOperand(0))->getOffset().getValueType());
+
+ // The match failed.
+ return SDValue();
+ };
auto matchForwardStrided = [](ArrayRef<SDValue> Ptrs) {
SDValue Stride;
for (auto Idx : enumerate(Ptrs)) {
@@ -13862,13 +13915,21 @@ static SDValue performCONCAT_VECTORSCombine(SDNode *N, SelectionDAG &DAG,
SDValue Stride = matchForwardStrided(Ptrs);
if (!Stride) {
Stride = matchReverseStrided(Ptrs);
- Reversed = true;
- // TODO: At this point, we've successfully matched a generalized gather
- // load. Maybe we should emit that, and then move the specialized
- // matchers above and below into a DAG combine?
- if (!Stride)
- return SDValue();
+ if (Stride) {
+ Reversed = true;
+ Stride = DAG.getNegative(Stride, DL, Stride->getValueType(0));
+ } else {
+ Stride = matchConstantStride(N->ops());
+ if (Stride) {
+ Reversed = cast<ConstantSDNode>(Stride)->getSExtValue() < 0;
+ } else {
+ return SDValue();
+ }
+ }
}
+ // TODO: At this point, we've successfully matched a generalized gather
+ // load. Maybe we should emit that, and then move the specialized
+ // matchers above and below into a DAG combine?
// Get the widened scalar type, e.g. v4i8 -> i64
unsigned WideScalarBitWidth =
@@ -13888,8 +13949,6 @@ static SDValue performCONCAT_VECTORSCombine(SDNode *N, SelectionDAG &DAG,
SDValue IntID =
DAG.getTargetConstant(Intrinsic::riscv_masked_strided_load, DL,
Subtarget.getXLenVT());
- if (Reversed)
- Stride = DAG.getNegative(Stride, DL, Stride->getValueType(0));
SDValue AllOneMask =
DAG.getSplat(WideVecVT.changeVectorElementType(MVT::i1), DL,
DAG.getConstant(1, DL, MVT::i1));
diff --git a/llvm/test/CodeGen/RISCV/rvv/concat-vectors-constant-stride.ll b/llvm/test/CodeGen/RISCV/rvv/concat-vectors-constant-stride.ll
new file mode 100644
index 000000000000000..611270ab98ebdaf
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/rvv/concat-vectors-constant-stride.ll
@@ -0,0 +1,231 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+v,+unaligned-vector-mem -target-abi=ilp32 \
+; RUN: -verify-machineinstrs < %s | FileCheck %s --check-prefixes=CHECK,RV32
+; RUN: llc -mtriple=riscv64 -mattr=+v,+unaligned-vector-mem -target-abi=lp64 \
+; RUN: -verify-machineinstrs < %s | FileCheck %s --check-prefixes=CHECK,RV64
+
+define void @constant_forward_stride(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_forward_stride:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, 16
+; CHECK-NEXT: addi a3, a0, 32
+; CHECK-NEXT: addi a4, a0, 48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vle8.v v9, (a2)
+; CHECK-NEXT: vle8.v v10, (a3)
+; CHECK-NEXT: vle8.v v11, (a4)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 16
+ %2 = getelementptr inbounds i8, ptr %s, i64 32
+ %3 = getelementptr inbounds i8, ptr %s, i64 48
+ %4 = load <2 x i8>, ptr %s, align 1
+ %5 = load <2 x i8>, ptr %1, align 1
+ %6 = load <2 x i8>, ptr %2, align 1
+ %7 = load <2 x i8>, ptr %3, align 1
+ %8 = shufflevector <2 x i8> %4, <2 x i8> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %9 = shufflevector <2 x i8> %6, <2 x i8> %7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %10 = shufflevector <4 x i8> %8, <4 x i8> %9, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %10, ptr %d, align 1
+ ret void
+}
+
+define void @constant_forward_stride2(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_forward_stride2:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, -16
+; CHECK-NEXT: addi a3, a0, -32
+; CHECK-NEXT: addi a4, a0, -48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a4)
+; CHECK-NEXT: vle8.v v9, (a3)
+; CHECK-NEXT: vle8.v v10, (a2)
+; CHECK-NEXT: vle8.v v11, (a0)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 -16
+ %2 = getelementptr inbounds i8, ptr %s, i64 -32
+ %3 = getelementptr inbounds i8, ptr %s, i64 -48
+ %4 = load <2 x i8>, ptr %3, align 1
+ %5 = load <2 x i8>, ptr %2, align 1
+ %6 = load <2 x i8>, ptr %1, align 1
+ %7 = load <2 x i8>, ptr %s, align 1
+ %8 = shufflevector <2 x i8> %4, <2 x i8> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %9 = shufflevector <2 x i8> %6, <2 x i8> %7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %10 = shufflevector <4 x i8> %8, <4 x i8> %9, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %10, ptr %d, align 1
+ ret void
+}
+
+define void @constant_forward_stride3(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_forward_stride3:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, 16
+; CHECK-NEXT: addi a3, a0, 32
+; CHECK-NEXT: addi a4, a0, 48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vle8.v v9, (a2)
+; CHECK-NEXT: vle8.v v10, (a3)
+; CHECK-NEXT: vle8.v v11, (a4)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 16
+ %2 = getelementptr inbounds i8, ptr %s, i64 32
+ %3 = getelementptr inbounds i8, ptr %s, i64 48
+ %4 = getelementptr inbounds i8, ptr %1, i64 0
+ %5 = getelementptr inbounds i8, ptr %2, i64 0
+ %6 = getelementptr inbounds i8, ptr %3, i64 0
+ %7 = load <2 x i8>, ptr %s, align 1
+ %8 = load <2 x i8>, ptr %4, align 1
+ %9 = load <2 x i8>, ptr %5, align 1
+ %10 = load <2 x i8>, ptr %6, align 1
+ %11 = shufflevector <2 x i8> %7, <2 x i8> %8, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %12 = shufflevector <2 x i8> %9, <2 x i8> %10, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %13 = shufflevector <4 x i8> %11, <4 x i8> %12, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %13, ptr %d, align 1
+ ret void
+}
+
+define void @constant_back_stride(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_back_stride:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, -16
+; CHECK-NEXT: addi a3, a0, -32
+; CHECK-NEXT: addi a4, a0, -48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vle8.v v9, (a2)
+; CHECK-NEXT: vle8.v v10, (a3)
+; CHECK-NEXT: vle8.v v11, (a4)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 -16
+ %2 = getelementptr inbounds i8, ptr %s, i64 -32
+ %3 = getelementptr inbounds i8, ptr %s, i64 -48
+ %4 = load <2 x i8>, ptr %s, align 1
+ %5 = load <2 x i8>, ptr %1, align 1
+ %6 = load <2 x i8>, ptr %2, align 1
+ %7 = load <2 x i8>, ptr %3, align 1
+ %8 = shufflevector <2 x i8> %4, <2 x i8> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %9 = shufflevector <2 x i8> %6, <2 x i8> %7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %10 = shufflevector <4 x i8> %8, <4 x i8> %9, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %10, ptr %d, align 1
+ ret void
+}
+
+define void @constant_back_stride2(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_back_stride2:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, 16
+; CHECK-NEXT: addi a3, a0, 32
+; CHECK-NEXT: addi a4, a0, 48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a4)
+; CHECK-NEXT: vle8.v v9, (a3)
+; CHECK-NEXT: vle8.v v10, (a2)
+; CHECK-NEXT: vle8.v v11, (a0)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 16
+ %2 = getelementptr inbounds i8, ptr %s, i64 32
+ %3 = getelementptr inbounds i8, ptr %s, i64 48
+ %4 = load <2 x i8>, ptr %3, align 1
+ %5 = load <2 x i8>, ptr %2, align 1
+ %6 = load <2 x i8>, ptr %1, align 1
+ %7 = load <2 x i8>, ptr %s, align 1
+ %8 = shufflevector <2 x i8> %4, <2 x i8> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %9 = shufflevector <2 x i8> %6, <2 x i8> %7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %10 = shufflevector <4 x i8> %8, <4 x i8> %9, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %10, ptr %d, align 1
+ ret void
+}
+
+define void @constant_back_stride3(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_back_stride3:
+; CHECK: # %bb.0:
+; CHECK-NEXT: addi a2, a0, -16
+; CHECK-NEXT: addi a3, a0, -32
+; CHECK-NEXT: addi a4, a0, -48
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vle8.v v9, (a2)
+; CHECK-NEXT: vle8.v v10, (a3)
+; CHECK-NEXT: vle8.v v11, (a4)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v9, 2
+; CHECK-NEXT: vsetivli zero, 6, e8, mf2, tu, ma
+; CHECK-NEXT: vslideup.vi v8, v10, 4
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vslideup.vi v8, v11, 6
+; CHECK-NEXT: vse8.v v8, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 -16
+ %2 = getelementptr inbounds i8, ptr %s, i64 -32
+ %3 = getelementptr inbounds i8, ptr %s, i64 -48
+ %4 = getelementptr inbounds i8, ptr %1, i64 0
+ %5 = getelementptr inbounds i8, ptr %2, i64 0
+ %6 = getelementptr inbounds i8, ptr %3, i64 0
+ %7 = load <2 x i8>, ptr %s, align 1
+ %8 = load <2 x i8>, ptr %4, align 1
+ %9 = load <2 x i8>, ptr %5, align 1
+ %10 = load <2 x i8>, ptr %6, align 1
+ %11 = shufflevector <2 x i8> %7, <2 x i8> %8, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %12 = shufflevector <2 x i8> %9, <2 x i8> %10, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ %13 = shufflevector <4 x i8> %11, <4 x i8> %12, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ store <8 x i8> %13, ptr %d, align 1
+ ret void
+}
+
+define void @constant_zero_stride(ptr %s, ptr %d) {
+; CHECK-LABEL: constant_zero_stride:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; CHECK-NEXT: vle8.v v8, (a0)
+; CHECK-NEXT: vsetivli zero, 4, e8, mf4, ta, ma
+; CHECK-NEXT: vmv1r.v v9, v8
+; CHECK-NEXT: vslideup.vi v9, v8, 2
+; CHECK-NEXT: vse8.v v9, (a1)
+; CHECK-NEXT: ret
+ %1 = getelementptr inbounds i8, ptr %s, i64 0
+ %2 = load <2 x i8>, ptr %s, align 1
+ %3 = load <2 x i8>, ptr %1, align 1
+ %4 = shufflevector <2 x i8> %2, <2 x i8> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+ store <4 x i8> %4, ptr %d, align 1
+ ret void
+}
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; RV32: {{.*}}
+; RV64: {{.*}}
|
michaelmaitland
changed the title
michaelmaitland
changed the title
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
2 times, most recently
from
ce03922
to
9e87a89
Compare
topperc
reviewed
Oct 10, 2023
preames
reviewed
Oct 10, 2023
| // loads. | ||
| if (Loads.size() < 2) | ||
| return SDValue(); | ||
| BaseIndexOffset BaseLdBIO = |
There was a problem hiding this comment.
Hm, an idea...
Looking at this code, I realized we could factor out a helper routine of the form optional<int64_t> getPtrDiff(LD1, LD2). Then we could loop over all of the indices, and check that the offset is known and the same.
After that, it occurred to me that both of the previous two matchers can also be written as rules inside a getPtrDiff helper. (The return value has to become SDValue instead, but otherwise it works.)
Maybe we should switch over in a separate commit, and then implement this matcher in that style?
There was a problem hiding this comment.
I realized we could factor out a helper routine of the form optional<int64_t> getPtrDiff(LD1, LD2). Then we could loop over all of the indices, and check that the offset is known and the same.
This seems to be very close to what BaseIndexOffset match does since if the (Base+Index) matches then the ptr diff is Offset. We could definitely add this helper method to simplify our checks in matchConstantStride.
After that, it occurred to me that both of the previous two matchers can also be written as rules inside a getPtrDiff helper. (The return value has to become SDValue instead, but otherwise it works.)
What happens in the case where there is no constant stride in matchFoward and matchReverse (i.e. stride can be a common add SDValue that is not a SDConstant).
There was a problem hiding this comment.
I think I understand, getPtrDiff would return the canonicalized representation of the non-constant value. We wouldn't be able to use BaseIndexOffset here. Do you have any ideas on what we could use?
There was a problem hiding this comment.
I pulled out getPtrDiff myself when trying wrap my head around this, hope this doesn't step on anyones toes: #69068
We should be able to extend it to check BaseIndexOffsets and return DAG.getConstant like we do here if they match
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
from
7b1867e
to
add2d4d
Compare
lukel97
added a commit
to lukel97/llvm-project
that referenced
this pull request
Oct 14, 2023
Instead of doing a forward pass for positive strides and a reverse pass for negative strides, we can just do one pass by negating the offset if the pointers do happen to be in reverse order. We can extend getPtrDiff later in llvm#68726 to handle more constant offset sequences.
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
from
add2d4d
to
100702c
Compare
lukel97
reviewed
Oct 16, 2023
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
from
7eb3b59
to
f5266db
Compare
lukel97
added a commit
that referenced
this pull request
Oct 16, 2023
Instead of doing a forward pass for positive strides and a reverse pass for negative strides, we can just do one pass by negating the offset if the pointers do happen to be in reverse order. We can extend getPtrDiff later in #68726 to handle more constant offset sequences.
lukel97
reviewed
Oct 16, 2023
| BaseIndexOffset BIO1 = BaseIndexOffset::match(Ld1, DAG); | ||
| BaseIndexOffset BIO2 = BaseIndexOffset::match(Ld2, DAG); | ||
| if (BIO1.equalBaseIndex(BIO2, DAG)) | ||
| return {{DAG.getConstant(BIO2.getOffset() - BIO1.getOffset(), DL, |
There was a problem hiding this comment.
Just realising we should probably defer creating the constant here too. One thought would to use a std::variant<uint64_t, SDValue> to represent either a constant or variable stride. But then we would need to extract the constants out of the SDValues below so we can compare the two etc. There's probably a simpler way of doing this
There was a problem hiding this comment.
You are right. Let me take a look.
There was a problem hiding this comment.
I think what you propose makes sense. I thought about trying to do it using std::function to defer creation of constant or negative, but that doesn't allow us to compare getPtrDiff() != BaseDiff. We need to do a comparison by value on three cases:
- SDValue, true
- SDValue, false
- int64_t, false
And the pair<variant, bool> structure permits us to do this check.
This patch commits tests that can be optimized by improving performCONCAT_VECTORCombine to do a better job at decomposing the base pointer and recognizing a constant offset.
If the load ptrs can be decomposed into a common (Base + Index) with a common constant stride, then return the constant stride.
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
from
0bcf88c
to
23687a3
Compare
michaelmaitland
force-pushed
the
comcat-vectors-combine
branch
from
46f6725
to
60d5c35
Compare
lukel97
approved these changes
Oct 16, 2023
There was a problem hiding this comment.
LGTM. Just leaving a note here that I think we'll fail to detect sequences of pointers that are a mixture of (base + index) and (add LastPtr, constStride), because the latter will be an SDValue and not a int64_t. But it sounds like an unlikely pattern to show up and I doubt its worth the complexity
My guess is that if BaseIndexOffset couldn't resolve to a match, then we would have no way to compare a constant to a non-constant anyway. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
I intend to commit the two commits in this patch independently: