[llvm] [RISCV] Pack build_vectors into largest available element type (PR #97351)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 1 14:35:52 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-risc-v
Author: Philip Reames (preames)
<details>
<summary>Changes</summary>
Our worst case build_vector lowering is a serial chain of vslide1down.vx operations which creates a serial dependency chain through a relatively high latency operation. We can instead pack together elements into ELEN sized chunks, and move them from integer to scalar in a single operation.
This reduces the length of the serial chain on the vector side, and costs at most three scalar instructions per element. This is a win for all cores when the sum of the latencies of the scalar instructions is less than the vslide1down.vx being replaced, and is particularly profitable for out-of-order cores which can overlap the scalar computation.
This patch is restricted to configurations with zba and zbb. Without both, the zero extend might require two instructions which would bring the total scalar instructions per element to 4. zba and zba are both present in the rva22u64 baseline which is looking to be quite common for hardware in practice; we could extend this to systems without bitmanip with a bit of extra effort.
---
Patch is 27.76 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/97351.diff
2 Files Affected:
- (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+67)
- (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll (+419-144)
``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 5e94fbec5a04a..3a431e8aa8c0e 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -3896,6 +3896,66 @@ static SDValue lowerBuildVectorOfConstants(SDValue Op, SelectionDAG &DAG,
return SDValue();
}
+/// Double the element size of the build vector to reduce the number
+/// of vslide1down in the build vector chain. In the worst case, this
+/// trades three scalar operations for 1 vector operation. Scalar
+/// operations are generally lower latency, and for out-of-order cores
+/// we also benefit from additional parallelism.
+static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
+ const RISCVSubtarget &Subtarget) {
+ SDLoc DL(Op);
+ MVT VT = Op.getSimpleValueType();
+ assert(VT.isFixedLengthVector() && "Unexpected vector!");
+ MVT ElemVT = VT.getVectorElementType();
+ if (!ElemVT.isInteger())
+ return SDValue();
+
+ // TODO: Relax these architectural restrictions, possibly with costing
+ // of the actual instructions required.
+ if (!Subtarget.hasStdExtZbb() || !Subtarget.hasStdExtZba())
+ return SDValue();
+
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned ElemSizeInBits = ElemVT.getSizeInBits();
+ if (ElemSizeInBits >= Subtarget.getELen() || NumElts % 2 != 0)
+ return SDValue();
+
+ // Produce [B,A] packed into a type twice as wide. Note that all
+ // scalars are XLenVT, possibly masked (see below).
+ MVT XLenVT = Subtarget.getXLenVT();
+ auto pack = [&](SDValue A, SDValue B) {
+ // Bias the scheduling of the inserted operations to near the
+ // definition of the element - this tends to reduce register
+ // pressure overall.
+ SDLoc ElemDL(B);
+ SDValue ShtAmt = DAG.getConstant(ElemSizeInBits, ElemDL, XLenVT);
+ return DAG.getNode(ISD::OR, ElemDL, XLenVT, A,
+ DAG.getNode(ISD::SHL, ElemDL, XLenVT, B, ShtAmt));
+ };
+
+ SDValue Mask = DAG.getConstant(
+ APInt::getLowBitsSet(XLenVT.getSizeInBits(), ElemSizeInBits), DL, XLenVT);
+ SmallVector<SDValue> NewOperands;
+ NewOperands.reserve(NumElts / 2);
+ for (unsigned i = 0; i < VT.getVectorNumElements(); i += 2) {
+ SDValue A = Op.getOperand(i);
+ SDValue B = Op.getOperand(i + 1);
+ if (ElemVT != XLenVT) {
+ // Bias the scheduling of the inserted operations to near the
+ // definition of the element - this tends to reduce register
+ // pressure overall.
+ A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
+ B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
+ }
+ NewOperands.push_back(pack(A, B));
+ }
+ assert(NumElts == NewOperands.size() * 2);
+ MVT WideVT = MVT::getIntegerVT(ElemSizeInBits * 2);
+ MVT WideVecVT = MVT::getVectorVT(WideVT, NumElts / 2);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getBuildVector(WideVecVT, DL, NewOperands));
+}
+
static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
const RISCVSubtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
@@ -3981,6 +4041,13 @@ static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
+ // If we're about to resort to vslide1down (or stack usage), pack our
+ // elements into the widest scalar type we can. This will force a VL/VTYPE
+ // toggle, but reduces the critical path, the number of vslide1down ops
+ // required, and possibly enables scalar folds of the values.
+ if (SDValue Res = lowerBuildVectorViaPacking(Op, DAG, Subtarget))
+ return Res;
+
// For m1 vectors, if we have non-undef values in both halves of our vector,
// split the vector into low and high halves, build them separately, then
// use a vselect to combine them. For long vectors, this cuts the critical
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
index 6cd69bac46e3c..94f9f480d0ba0 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
@@ -1267,43 +1267,53 @@ define <16 x i8> @buildvec_v16i8_loads_contigous(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_loads_contigous:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: addi a6, a0, 8
-; RVA22U64-NEXT: lbu t6, 1(a0)
+; RVA22U64-NEXT: lbu a1, 1(a0)
+; RVA22U64-NEXT: lbu a2, 0(a0)
; RVA22U64-NEXT: lbu a3, 2(a0)
; RVA22U64-NEXT: lbu a4, 3(a0)
-; RVA22U64-NEXT: lbu a5, 4(a0)
-; RVA22U64-NEXT: lbu t5, 5(a0)
-; RVA22U64-NEXT: lbu a7, 6(a0)
-; RVA22U64-NEXT: lbu t0, 7(a0)
-; RVA22U64-NEXT: lbu t1, 9(a0)
-; RVA22U64-NEXT: lbu t2, 10(a0)
-; RVA22U64-NEXT: lbu t3, 11(a0)
-; RVA22U64-NEXT: lbu t4, 12(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a0), zero
-; RVA22U64-NEXT: lbu a1, 13(a0)
-; RVA22U64-NEXT: lbu a2, 14(a0)
+; RVA22U64-NEXT: slli a1, a1, 8
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: slli a3, a3, 16
+; RVA22U64-NEXT: slli a4, a4, 24
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: lbu a2, 4(a0)
+; RVA22U64-NEXT: or a1, a1, a3
+; RVA22U64-NEXT: lbu a3, 5(a0)
+; RVA22U64-NEXT: lbu a4, 6(a0)
+; RVA22U64-NEXT: slli a2, a2, 32
+; RVA22U64-NEXT: lbu a5, 7(a0)
+; RVA22U64-NEXT: slli a3, a3, 40
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 48
+; RVA22U64-NEXT: slli a5, a5, 56
+; RVA22U64-NEXT: or a4, a4, a5
+; RVA22U64-NEXT: or a2, a2, a4
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: lbu a2, 9(a0)
+; RVA22U64-NEXT: lbu a3, 8(a0)
+; RVA22U64-NEXT: lbu a4, 10(a0)
+; RVA22U64-NEXT: lbu a5, 11(a0)
+; RVA22U64-NEXT: slli a2, a2, 8
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 16
+; RVA22U64-NEXT: slli a5, a5, 24
+; RVA22U64-NEXT: or a4, a4, a5
+; RVA22U64-NEXT: lbu a3, 12(a0)
+; RVA22U64-NEXT: or a2, a2, a4
+; RVA22U64-NEXT: lbu a4, 13(a0)
+; RVA22U64-NEXT: lbu a5, 14(a0)
+; RVA22U64-NEXT: slli a3, a3, 32
; RVA22U64-NEXT: lbu a0, 15(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t6
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: vlse8.v v9, (a6), zero
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t5
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a7
-; RVA22U64-NEXT: vslide1down.vx v10, v8, t0
-; RVA22U64-NEXT: vslide1down.vx v8, v9, t1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t2
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
-; RVA22U64-NEXT: li a1, 255
-; RVA22U64-NEXT: vsetvli zero, zero, e16, m2, ta, ma
-; RVA22U64-NEXT: vmv.s.x v0, a1
-; RVA22U64-NEXT: vsetvli zero, zero, e8, m1, ta, mu
+; RVA22U64-NEXT: slli a4, a4, 40
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: slli a5, a5, 48
+; RVA22U64-NEXT: slli a0, a0, 56
+; RVA22U64-NEXT: or a0, a0, a5
+; RVA22U64-NEXT: or a0, a0, a3
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.x v8, a1
; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
-; RVA22U64-NEXT: vslidedown.vi v8, v10, 8, v0.t
; RVA22U64-NEXT: ret
;
; RV64ZVE32-LABEL: buildvec_v16i8_loads_contigous:
@@ -1484,43 +1494,53 @@ define <16 x i8> @buildvec_v16i8_loads_gather(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_loads_gather:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: addi a6, a0, 82
-; RVA22U64-NEXT: lbu t6, 1(a0)
+; RVA22U64-NEXT: lbu a1, 1(a0)
+; RVA22U64-NEXT: lbu a2, 0(a0)
; RVA22U64-NEXT: lbu a3, 22(a0)
; RVA22U64-NEXT: lbu a4, 31(a0)
-; RVA22U64-NEXT: lbu a5, 44(a0)
-; RVA22U64-NEXT: lbu t5, 55(a0)
-; RVA22U64-NEXT: lbu a7, 623(a0)
-; RVA22U64-NEXT: lbu t0, 75(a0)
-; RVA22U64-NEXT: lbu t1, 93(a0)
-; RVA22U64-NEXT: lbu t2, 105(a0)
-; RVA22U64-NEXT: lbu t3, 161(a0)
-; RVA22U64-NEXT: lbu t4, 124(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a0), zero
-; RVA22U64-NEXT: lbu a1, 163(a0)
-; RVA22U64-NEXT: lbu a2, 144(a0)
+; RVA22U64-NEXT: slli a1, a1, 8
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: slli a3, a3, 16
+; RVA22U64-NEXT: slli a4, a4, 24
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: lbu a2, 44(a0)
+; RVA22U64-NEXT: or a1, a1, a3
+; RVA22U64-NEXT: lbu a3, 55(a0)
+; RVA22U64-NEXT: lbu a4, 623(a0)
+; RVA22U64-NEXT: slli a2, a2, 32
+; RVA22U64-NEXT: lbu a5, 75(a0)
+; RVA22U64-NEXT: slli a3, a3, 40
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 48
+; RVA22U64-NEXT: slli a5, a5, 56
+; RVA22U64-NEXT: or a4, a4, a5
+; RVA22U64-NEXT: or a2, a2, a4
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: lbu a2, 93(a0)
+; RVA22U64-NEXT: lbu a3, 82(a0)
+; RVA22U64-NEXT: lbu a4, 105(a0)
+; RVA22U64-NEXT: lbu a5, 161(a0)
+; RVA22U64-NEXT: slli a2, a2, 8
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 16
+; RVA22U64-NEXT: slli a5, a5, 24
+; RVA22U64-NEXT: or a4, a4, a5
+; RVA22U64-NEXT: lbu a3, 124(a0)
+; RVA22U64-NEXT: or a2, a2, a4
+; RVA22U64-NEXT: lbu a4, 163(a0)
+; RVA22U64-NEXT: lbu a5, 144(a0)
+; RVA22U64-NEXT: slli a3, a3, 32
; RVA22U64-NEXT: lbu a0, 154(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t6
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: vlse8.v v9, (a6), zero
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t5
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a7
-; RVA22U64-NEXT: vslide1down.vx v10, v8, t0
-; RVA22U64-NEXT: vslide1down.vx v8, v9, t1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t2
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, t4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
-; RVA22U64-NEXT: li a1, 255
-; RVA22U64-NEXT: vsetvli zero, zero, e16, m2, ta, ma
-; RVA22U64-NEXT: vmv.s.x v0, a1
-; RVA22U64-NEXT: vsetvli zero, zero, e8, m1, ta, mu
+; RVA22U64-NEXT: slli a4, a4, 40
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: slli a5, a5, 48
+; RVA22U64-NEXT: slli a0, a0, 56
+; RVA22U64-NEXT: or a0, a0, a5
+; RVA22U64-NEXT: or a0, a0, a3
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.x v8, a1
; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
-; RVA22U64-NEXT: vslidedown.vi v8, v10, 8, v0.t
; RVA22U64-NEXT: ret
;
; RV64ZVE32-LABEL: buildvec_v16i8_loads_gather:
@@ -1660,22 +1680,30 @@ define <16 x i8> @buildvec_v16i8_undef_low_half(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_undef_low_half:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: addi a1, a0, 82
-; RVA22U64-NEXT: lbu a6, 93(a0)
+; RVA22U64-NEXT: lbu a1, 93(a0)
+; RVA22U64-NEXT: lbu a2, 82(a0)
; RVA22U64-NEXT: lbu a3, 105(a0)
; RVA22U64-NEXT: lbu a4, 161(a0)
-; RVA22U64-NEXT: lbu a5, 124(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a1), zero
-; RVA22U64-NEXT: lbu a1, 163(a0)
-; RVA22U64-NEXT: lbu a2, 144(a0)
+; RVA22U64-NEXT: slli a1, a1, 8
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: slli a3, a3, 16
+; RVA22U64-NEXT: slli a4, a4, 24
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: lbu a2, 124(a0)
+; RVA22U64-NEXT: or a1, a1, a3
+; RVA22U64-NEXT: lbu a3, 163(a0)
+; RVA22U64-NEXT: lbu a4, 144(a0)
+; RVA22U64-NEXT: slli a2, a2, 32
; RVA22U64-NEXT: lbu a0, 154(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a6
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
+; RVA22U64-NEXT: slli a3, a3, 40
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 48
+; RVA22U64-NEXT: slli a0, a0, 56
+; RVA22U64-NEXT: or a0, a0, a4
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: or a0, a0, a1
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.i v8, 0
; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
; RVA22U64-NEXT: ret
;
@@ -1773,23 +1801,31 @@ define <16 x i8> @buildvec_v16i8_undef_high_half(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_undef_high_half:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: lbu a6, 1(a0)
-; RVA22U64-NEXT: lbu a2, 22(a0)
-; RVA22U64-NEXT: lbu a3, 31(a0)
-; RVA22U64-NEXT: lbu a4, 44(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a0), zero
-; RVA22U64-NEXT: lbu a5, 55(a0)
-; RVA22U64-NEXT: lbu a1, 623(a0)
+; RVA22U64-NEXT: lbu a1, 1(a0)
+; RVA22U64-NEXT: lbu a2, 0(a0)
+; RVA22U64-NEXT: lbu a3, 22(a0)
+; RVA22U64-NEXT: lbu a4, 31(a0)
+; RVA22U64-NEXT: slli a1, a1, 8
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: slli a3, a3, 16
+; RVA22U64-NEXT: slli a4, a4, 24
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: lbu a2, 44(a0)
+; RVA22U64-NEXT: or a1, a1, a3
+; RVA22U64-NEXT: lbu a3, 55(a0)
+; RVA22U64-NEXT: lbu a4, 623(a0)
+; RVA22U64-NEXT: slli a2, a2, 32
; RVA22U64-NEXT: lbu a0, 75(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a6
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 8
+; RVA22U64-NEXT: slli a3, a3, 40
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 48
+; RVA22U64-NEXT: slli a0, a0, 56
+; RVA22U64-NEXT: or a0, a0, a4
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: or a0, a0, a1
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.x v8, a0
+; RVA22U64-NEXT: vslide1down.vx v8, v8, zero
; RVA22U64-NEXT: ret
;
; RV64ZVE32-LABEL: buildvec_v16i8_undef_high_half:
@@ -1901,31 +1937,33 @@ define <16 x i8> @buildvec_v16i8_undef_edges(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_undef_edges:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: addi a1, a0, 31
-; RVA22U64-NEXT: addi a6, a0, 82
-; RVA22U64-NEXT: lbu a3, 44(a0)
-; RVA22U64-NEXT: lbu a4, 55(a0)
-; RVA22U64-NEXT: lbu a5, 623(a0)
-; RVA22U64-NEXT: lbu a7, 75(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a1), zero
-; RVA22U64-NEXT: lbu a1, 93(a0)
-; RVA22U64-NEXT: lbu a2, 105(a0)
+; RVA22U64-NEXT: lbu a1, 44(a0)
+; RVA22U64-NEXT: lbu a2, 55(a0)
+; RVA22U64-NEXT: lbu a3, 31(a0)
+; RVA22U64-NEXT: lbu a4, 623(a0)
+; RVA22U64-NEXT: slli a1, a1, 32
+; RVA22U64-NEXT: slli a2, a2, 40
+; RVA22U64-NEXT: lbu a5, 75(a0)
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: slli a3, a3, 24
+; RVA22U64-NEXT: slli a4, a4, 48
+; RVA22U64-NEXT: slli a5, a5, 56
+; RVA22U64-NEXT: or a4, a4, a5
+; RVA22U64-NEXT: or a1, a1, a4
+; RVA22U64-NEXT: add.uw a1, a3, a1
+; RVA22U64-NEXT: lbu a2, 93(a0)
+; RVA22U64-NEXT: lbu a3, 82(a0)
+; RVA22U64-NEXT: lbu a4, 105(a0)
; RVA22U64-NEXT: lbu a0, 161(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vlse8.v v9, (a6), zero
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: vslide1down.vx v10, v8, a7
-; RVA22U64-NEXT: vslide1down.vx v8, v9, a1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
+; RVA22U64-NEXT: slli a2, a2, 8
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: slli a4, a4, 16
+; RVA22U64-NEXT: slli a0, a0, 24
+; RVA22U64-NEXT: or a0, a0, a4
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.x v8, a1
; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
-; RVA22U64-NEXT: li a0, 255
-; RVA22U64-NEXT: vsetvli zero, zero, e16, m2, ta, ma
-; RVA22U64-NEXT: vmv.s.x v0, a0
-; RVA22U64-NEXT: vsetvli zero, zero, e8, m1, ta, mu
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 4
-; RVA22U64-NEXT: vslidedown.vi v8, v10, 8, v0.t
; RVA22U64-NEXT: ret
;
; RV64ZVE32-LABEL: buildvec_v16i8_undef_edges:
@@ -2057,35 +2095,35 @@ define <16 x i8> @buildvec_v16i8_loads_undef_scattered(ptr %p) {
;
; RVA22U64-LABEL: buildvec_v16i8_loads_undef_scattered:
; RVA22U64: # %bb.0:
-; RVA22U64-NEXT: addi a6, a0, 82
-; RVA22U64-NEXT: lbu a2, 1(a0)
+; RVA22U64-NEXT: lbu a1, 1(a0)
+; RVA22U64-NEXT: lbu a2, 0(a0)
+; RVA22U64-NEXT: slli a1, a1, 8
; RVA22U64-NEXT: lbu a3, 44(a0)
; RVA22U64-NEXT: lbu a4, 55(a0)
-; RVA22U64-NEXT: lbu t0, 75(a0)
-; RVA22U64-NEXT: lbu a7, 93(a0)
-; RVA22U64-NEXT: vsetivli zero, 16, e8, m1, ta, ma
-; RVA22U64-NEXT: vlse8.v v8, (a0), zero
-; RVA22U64-NEXT: lbu a1, 124(a0)
-; RVA22U64-NEXT: lbu a5, 144(a0)
-; RVA22U64-NEXT: lbu a0, 154(a0)
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a2
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 2
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a3
-; RVA22U64-NEXT: vlse8.v v9, (a6), zero
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a4
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 1
-; RVA22U64-NEXT: vslide1down.vx v10, v8, t0
-; RVA22U64-NEXT: vslide1down.vx v8, v9, a7
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 2
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a1
-; RVA22U64-NEXT: vslidedown.vi v8, v8, 1
-; RVA22U64-NEXT: vslide1down.vx v8, v8, a5
-; RVA22U64-NEXT: li a1, 255
-; RVA22U64-NEXT: vsetvli zero, zero, e16, m2, ta, ma
-; RVA22U64-NEXT: vmv.s.x v0, a1
-; RVA22U64-NEXT: vsetvli zero, zero, e8, m1, ta, mu
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: lbu a2, 75(a0)
+; RVA22U64-NEXT: slli a3, a3, 32
+; RVA22U64-NEXT: slli a4, a4, 40
+; RVA22U64-NEXT: or a3, a3, a4
+; RVA22U64-NEXT: slli a2, a2, 56
+; RVA22U64-NEXT: lbu a4, 93(a0)
+; RVA22U64-NEXT: or a2, a2, a3
+; RVA22U64-NEXT: or a1, a1, a2
+; RVA22U64-NEXT: lbu a2, 82(a0)
+; RVA22U64-NEXT: slli a4, a4, 8
+; RVA22U64-NEXT: lbu a3, 144(a0)
+; RVA22U64-NEXT: lbu a5, 154(a0)
+; RVA22U64-NEXT: or a2, a2, a4
+; RVA22U64-NEXT: lbu a0, 124(a0)
+; RVA22U64-NEXT: slli a3, a3, 48
+; RVA22U64-NEXT: slli a5, a5, 56
+; RVA22U64-NEXT: or a3, a3, a5
+; RVA22U64-NEXT: slli a0, a0, 32
+; RVA22U64-NEXT: or a0, a0, a3
+; RVA22U64-NEXT: or a0, a0, a2
+; RVA22U64-NEXT: vsetivli zero, 2, e64, m1, ta, ma
+; RVA22U64-NEXT: vmv.v.x v8, a1
; RVA22U64-NEXT: vslide1down.vx v8, v8, a0
-; RVA22U64-NEXT: vslidedown.vi v8, v10, 8, v0.t
; RVA22U64-NEXT: ret
;
; RV64ZVE32-LABEL: buildvec_v16i8_loads_undef_scattered:
@@ -2171,3 +2209,240 @@ define <16 x i8> @buildvec_v16i8_loads_undef_scattered(ptr %p) {
%v16 = insertelement <16 x i8> %v15, i8 %ld16, i32 15
ret <16 x i8> %v16
}
+
+define <8 x i8> @buildvec_v8i8_pack(ptr %p, i8 %e1, i8 %e2, i8 %e3, i8 %e4, i8 %e5, i8 %e6, i8 %e7, i8 %e8) {
+; RV32-LABEL: buildvec_v8i8_pack:
+; RV32: # %bb.0:
+; RV32-NEXT: lbu a0, 0(sp)
+; RV32-NEXT: vsetivli zero, 8, e8, mf2, ta, mu
+; RV32-NEXT: vmv.v.x v8, a1
+; RV32-NEXT: vslide1down.vx v8, v8, a2
+; RV32-NEXT: vslide1down.vx v8, v8, a3
+; RV32-NEXT: vslide1down.vx v9, v8, a4
+; RV32-NEXT: vmv.v.x v8, a5
+; RV32-NEXT: vslide1down.vx v8, v8, a6
+; RV32-NEXT: vslide1down.vx v8, v8, a7
+; RV32-NEXT: vmv.v.i v0, 15
+; RV32-NEXT: vslide1down.vx v8, v8, a0
+; RV32-NEXT: vslidedown.vi v8, v9, 4, v0.t
+; RV32-NEXT: ret
+;
+; RV64V-ONLY-LABEL: buildvec_v8i8_pack:
+; RV64V-ONLY: # %bb.0:
+; RV64V-ONLY-NEXT: lbu a0, 0(sp)
+; RV64V-ONLY-NEXT: vsetivli zero, 8, e8, mf2, ta, mu
+; RV64V-ONLY-NEXT: vmv.v.x v8, a1
+; RV64V-ONLY-NEXT: vslide1down.vx v8, v8, a2
+; RV64V-ONLY-NEXT: vslide1down.vx v8, v8, a3
+; RV64V-ONLY-NEXT: vsli...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/97351
More information about the llvm-commits
mailing list