[RISCV] Use ri.vunzip2{a,b} for e64 fixed length deinterleave(2) shuffles #137217
+99
−96
Conversation
This file contains hidden or 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
…fles
If we have xrivosvizip, we can use the vunzip2{a,b} instructions for these
cases *provided* that we can prove the layout in the two registers matches
the fixed length semantics.
The majority of this patch is a straight-forward port of the existing
vnsrl logic which has the same requirement (though for slightly different
reasoning).
The one complicated bit is the addition of the scalable splitting logic
inside lowerVZIP to exploit the independent register operands, and allow
the use of lower LMUL. This bit is annoyingly complicated, and really
"should" be a DAG combine - except that the VL and mask reduction becomes
hard when it's not known to be a constant.
|
@llvm/pr-subscribers-backend-risc-v Author: Philip Reames (preames) ChangesIf we have xrivosvizip, we can use the vunzip2{a,b} instructions for these cases provided that we can prove the layout in the two registers matches the fixed length semantics. The majority of this patch is a straight-forward port of the existing vnsrl logic which has the same requirement (though for slightly different reasoning). The one complicated bit is the addition of the scalable splitting logic inside lowerVZIP to exploit the independent register operands, and allow the use of lower LMUL. This bit is annoyingly complicated, and really "should" be a DAG combine - except that the VL and mask reduction becomes hard when it's not known to be a constant. Full diff: https://github.com/llvm/llvm-project/pull/137217.diff 3 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index c440df5a3e638..a20d89cd8f8d1 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -5062,9 +5062,27 @@ static SDValue lowerVZIP(unsigned Opc, SDValue Op0, SDValue Op1,
Op1 = convertToScalableVector(ContainerVT, Op1, DAG, Subtarget);
}
- auto [Mask, VL] = getDefaultVLOps(IntVT, ContainerVT, DL, DAG, Subtarget);
- SDValue Passthru = DAG.getUNDEF(ContainerVT);
- SDValue Res = DAG.getNode(Opc, DL, ContainerVT, Op0, Op1, Passthru, Mask, VL);
+ MVT InnerVT = ContainerVT;
+ auto [Mask, VL] = getDefaultVLOps(IntVT, InnerVT, DL, DAG, Subtarget);
+ if (Op1.isUndef() && ContainerVT.bitsGT(getLMUL1VT(ContainerVT)) &&
+ (RISCVISD::RI_VUNZIP2A_VL == Opc || RISCVISD::RI_VUNZIP2B_VL == Opc)) {
+ InnerVT = ContainerVT.getHalfNumVectorElementsVT();
+ VL = DAG.getConstant(VT.getVectorNumElements() / 2, DL,
+ Subtarget.getXLenVT());
+ Mask = getAllOnesMask(InnerVT, VL, DL, DAG);
+ unsigned HighIdx = InnerVT.getVectorElementCount().getKnownMinValue();
+ Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InnerVT, Op0,
+ DAG.getVectorIdxConstant(HighIdx, DL));
+ Op0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InnerVT, Op0,
+ DAG.getVectorIdxConstant(0, DL));
+ }
+
+ SDValue Passthru = DAG.getUNDEF(InnerVT);
+ SDValue Res = DAG.getNode(Opc, DL, InnerVT, Op0, Op1, Passthru, Mask, VL);
+ if (InnerVT.bitsLT(ContainerVT))
+ Res = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT,
+ DAG.getUNDEF(ContainerVT), Res,
+ DAG.getVectorIdxConstant(0, DL));
if (IntVT.isFixedLengthVector())
Res = convertFromScalableVector(IntVT, Res, DAG, Subtarget);
Res = DAG.getBitcast(VT, Res);
@@ -5768,6 +5786,26 @@ static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
}
}
+ // If this is a deinterleave(2), try using vunzip{a,b}. This mostly catches
+ // e64 which can't match above.
+ unsigned Index = 0;
+ if (Subtarget.hasVendorXRivosVizip() &&
+ ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask, 2, Index) &&
+ 1 < count_if(Mask, [](int Idx) { return Idx != -1; })) {
+ unsigned Opc =
+ Index == 0 ? RISCVISD::RI_VUNZIP2A_VL : RISCVISD::RI_VUNZIP2B_VL;
+ if (V2.isUndef())
+ return lowerVZIP(Opc, V1, V2, DL, DAG, Subtarget);
+ if (SDValue Src = foldConcatVector(V1, V2)) {
+ EVT NewVT = VT.getDoubleNumVectorElementsVT();
+ SDValue ZeroIdx = DAG.getVectorIdxConstant(0, DL);
+ Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NewVT, Src, ZeroIdx);
+ SDValue Res =
+ lowerVZIP(Opc, Src, DAG.getUNDEF(NewVT), DL, DAG, Subtarget);
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, ZeroIdx);
+ }
+ }
+
if (SDValue V =
lowerVECTOR_SHUFFLEAsVSlideup(DL, VT, V1, V2, Mask, Subtarget, DAG))
return V;
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
index c65d7c36a2198..b692a80159288 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-deinterleave2.ll
@@ -347,9 +347,8 @@ define void @vnsrl_0_i64(ptr %in, ptr %out) {
; ZIP-NEXT: vsetivli zero, 4, e64, m1, ta, ma
; ZIP-NEXT: vle64.v v8, (a0)
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: vslidedown.vi v9, v8, 2
-; ZIP-NEXT: vslideup.vi v8, v9, 1
-; ZIP-NEXT: vse64.v v8, (a1)
+; ZIP-NEXT: ri.vunzip2a.vv v10, v8, v9
+; ZIP-NEXT: vse64.v v10, (a1)
; ZIP-NEXT: ret
entry:
%0 = load <4 x i64>, ptr %in, align 8
@@ -383,8 +382,7 @@ define void @vnsrl_64_i64(ptr %in, ptr %out) {
; ZIP-NEXT: vsetivli zero, 4, e64, m1, ta, ma
; ZIP-NEXT: vle64.v v8, (a0)
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: vslidedown.vi v9, v8, 2
-; ZIP-NEXT: ri.vzipodd.vv v10, v8, v9
+; ZIP-NEXT: ri.vunzip2b.vv v10, v8, v9
; ZIP-NEXT: vse64.v v10, (a1)
; ZIP-NEXT: ret
entry:
@@ -417,10 +415,9 @@ define void @vnsrl_0_double(ptr %in, ptr %out) {
; ZIP: # %bb.0: # %entry
; ZIP-NEXT: vsetivli zero, 4, e64, m1, ta, ma
; ZIP-NEXT: vle64.v v8, (a0)
+; ZIP-NEXT: ri.vunzip2a.vv v10, v8, v9
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: vslidedown.vi v9, v8, 2
-; ZIP-NEXT: vslideup.vi v8, v9, 1
-; ZIP-NEXT: vse64.v v8, (a1)
+; ZIP-NEXT: vse64.v v10, (a1)
; ZIP-NEXT: ret
entry:
%0 = load <4 x double>, ptr %in, align 8
@@ -453,9 +450,8 @@ define void @vnsrl_64_double(ptr %in, ptr %out) {
; ZIP: # %bb.0: # %entry
; ZIP-NEXT: vsetivli zero, 4, e64, m1, ta, ma
; ZIP-NEXT: vle64.v v8, (a0)
+; ZIP-NEXT: ri.vunzip2b.vv v10, v8, v9
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: vslidedown.vi v9, v8, 2
-; ZIP-NEXT: ri.vzipodd.vv v10, v8, v9
; ZIP-NEXT: vse64.v v10, (a1)
; ZIP-NEXT: ret
entry:
diff --git a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-fixed.ll b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-fixed.ll
index f6b5a35aa06d6..3263ad121a9db 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-fixed.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-fixed.ll
@@ -85,11 +85,11 @@ define {<2 x i64>, <2 x i64>} @vector_deinterleave_v2i64_v4i64(<4 x i64> %vec) {
;
; ZIP-LABEL: vector_deinterleave_v2i64_v4i64:
; ZIP: # %bb.0:
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v10, v8, 2
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: ri.vzipodd.vv v9, v8, v10
-; ZIP-NEXT: vslideup.vi v8, v10, 1
+; ZIP-NEXT: ri.vunzip2a.vv v10, v8, v9
+; ZIP-NEXT: ri.vunzip2b.vv v11, v8, v9
+; ZIP-NEXT: vmv.v.v v8, v10
+; ZIP-NEXT: vmv.v.v v9, v11
; ZIP-NEXT: ret
%retval = call {<2 x i64>, <2 x i64>} @llvm.vector.deinterleave2.v4i64(<4 x i64> %vec)
ret {<2 x i64>, <2 x i64>} %retval
@@ -129,62 +129,51 @@ define {<4 x i64>, <4 x i64>} @vector_deinterleave_v4i64_v8i64(<8 x i64> %vec) {
; ZIP-LABEL: vector_deinterleave_v4i64_v8i64:
; ZIP: # %bb.0:
; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v12, v8, 1
-; ZIP-NEXT: vsetivli zero, 1, e8, mf8, ta, ma
-; ZIP-NEXT: vmv.v.i v0, 2
-; ZIP-NEXT: vmv.v.i v14, 12
-; ZIP-NEXT: vsetivli zero, 4, e64, m4, ta, ma
-; ZIP-NEXT: vslidedown.vi v16, v8, 4
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v10, v8, 2
-; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, mu
-; ZIP-NEXT: vslidedown.vi v12, v8, 2, v0.t
-; ZIP-NEXT: ri.vzip2a.vv v18, v8, v10
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v8, v16, 2
-; ZIP-NEXT: vmv1r.v v0, v14
-; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, mu
-; ZIP-NEXT: ri.vzip2a.vv v12, v16, v8, v0.t
-; ZIP-NEXT: vsetivli zero, 1, e8, mf8, ta, ma
-; ZIP-NEXT: vmv.v.i v0, 8
-; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, mu
-; ZIP-NEXT: vslideup.vi v8, v16, 2
-; ZIP-NEXT: vslideup.vi v8, v16, 1, v0.t
-; ZIP-NEXT: vmv1r.v v0, v14
-; ZIP-NEXT: vmerge.vvm v8, v18, v8, v0
-; ZIP-NEXT: vmv2r.v v10, v12
+; ZIP-NEXT: ri.vunzip2a.vv v12, v8, v10
+; ZIP-NEXT: ri.vunzip2b.vv v14, v8, v10
+; ZIP-NEXT: vmv.v.v v8, v12
+; ZIP-NEXT: vmv.v.v v10, v14
; ZIP-NEXT: ret
%retval = call {<4 x i64>, <4 x i64>} @llvm.vector.deinterleave2.v8i64(<8 x i64> %vec)
ret {<4 x i64>, <4 x i64>} %retval
}
define {<8 x i64>, <8 x i64>} @vector_deinterleave_v8i64_v16i64(<16 x i64> %vec) {
-; CHECK-LABEL: vector_deinterleave_v8i64_v16i64:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a0, 85
-; CHECK-NEXT: vsetivli zero, 8, e16, m1, ta, ma
-; CHECK-NEXT: vmv.v.i v0, -16
-; CHECK-NEXT: vid.v v16
-; CHECK-NEXT: vsetivli zero, 8, e64, m8, ta, ma
-; CHECK-NEXT: vslidedown.vi v24, v8, 8
-; CHECK-NEXT: vmv.s.x v12, a0
-; CHECK-NEXT: li a0, 170
-; CHECK-NEXT: vsetivli zero, 8, e16, m1, ta, ma
-; CHECK-NEXT: vadd.vv v20, v16, v16
-; CHECK-NEXT: vmv.s.x v21, a0
-; CHECK-NEXT: vsetvli zero, zero, e64, m4, ta, ma
-; CHECK-NEXT: vcompress.vm v16, v8, v12
-; CHECK-NEXT: vsetvli zero, zero, e16, m1, ta, ma
-; CHECK-NEXT: vadd.vi v22, v20, -8
-; CHECK-NEXT: vsetvli zero, zero, e64, m4, ta, ma
-; CHECK-NEXT: vcompress.vm v12, v8, v21
-; CHECK-NEXT: vsetvli zero, zero, e16, m1, ta, ma
-; CHECK-NEXT: vadd.vi v8, v20, -7
-; CHECK-NEXT: vsetvli zero, zero, e64, m4, ta, mu
-; CHECK-NEXT: vrgatherei16.vv v16, v24, v22, v0.t
-; CHECK-NEXT: vrgatherei16.vv v12, v24, v8, v0.t
-; CHECK-NEXT: vmv.v.v v8, v16
-; CHECK-NEXT: ret
+; V-LABEL: vector_deinterleave_v8i64_v16i64:
+; V: # %bb.0:
+; V-NEXT: li a0, 85
+; V-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; V-NEXT: vmv.v.i v0, -16
+; V-NEXT: vid.v v16
+; V-NEXT: vsetivli zero, 8, e64, m8, ta, ma
+; V-NEXT: vslidedown.vi v24, v8, 8
+; V-NEXT: vmv.s.x v12, a0
+; V-NEXT: li a0, 170
+; V-NEXT: vsetivli zero, 8, e16, m1, ta, ma
+; V-NEXT: vadd.vv v20, v16, v16
+; V-NEXT: vmv.s.x v21, a0
+; V-NEXT: vsetvli zero, zero, e64, m4, ta, ma
+; V-NEXT: vcompress.vm v16, v8, v12
+; V-NEXT: vsetvli zero, zero, e16, m1, ta, ma
+; V-NEXT: vadd.vi v22, v20, -8
+; V-NEXT: vsetvli zero, zero, e64, m4, ta, ma
+; V-NEXT: vcompress.vm v12, v8, v21
+; V-NEXT: vsetvli zero, zero, e16, m1, ta, ma
+; V-NEXT: vadd.vi v8, v20, -7
+; V-NEXT: vsetvli zero, zero, e64, m4, ta, mu
+; V-NEXT: vrgatherei16.vv v16, v24, v22, v0.t
+; V-NEXT: vrgatherei16.vv v12, v24, v8, v0.t
+; V-NEXT: vmv.v.v v8, v16
+; V-NEXT: ret
+;
+; ZIP-LABEL: vector_deinterleave_v8i64_v16i64:
+; ZIP: # %bb.0:
+; ZIP-NEXT: vsetivli zero, 8, e64, m4, ta, ma
+; ZIP-NEXT: ri.vunzip2a.vv v16, v8, v12
+; ZIP-NEXT: ri.vunzip2b.vv v20, v8, v12
+; ZIP-NEXT: vmv.v.v v8, v16
+; ZIP-NEXT: vmv.v.v v12, v20
+; ZIP-NEXT: ret
%retval = call {<8 x i64>, <8 x i64>} @llvm.vector.deinterleave2.v16i64(<16 x i64> %vec)
ret {<8 x i64>, <8 x i64>} %retval
}
@@ -622,11 +611,11 @@ define {<2 x double>, <2 x double>} @vector_deinterleave_v2f64_v4f64(<4 x double
;
; ZIP-LABEL: vector_deinterleave_v2f64_v4f64:
; ZIP: # %bb.0:
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v10, v8, 2
; ZIP-NEXT: vsetivli zero, 2, e64, m1, ta, ma
-; ZIP-NEXT: ri.vzipodd.vv v9, v8, v10
-; ZIP-NEXT: vslideup.vi v8, v10, 1
+; ZIP-NEXT: ri.vunzip2a.vv v10, v8, v9
+; ZIP-NEXT: ri.vunzip2b.vv v12, v8, v9
+; ZIP-NEXT: vmv.v.v v8, v10
+; ZIP-NEXT: vmv.v.v v9, v12
; ZIP-NEXT: ret
%retval = call {<2 x double>, <2 x double>} @llvm.vector.deinterleave2.v4f64(<4 x double> %vec)
ret {<2 x double>, <2 x double>} %retval
@@ -665,31 +654,11 @@ define {<4 x double>, <4 x double>} @vector_deinterleave_v4f64_v8f64(<8 x double
;
; ZIP-LABEL: vector_deinterleave_v4f64_v8f64:
; ZIP: # %bb.0:
-; ZIP-NEXT: vsetivli zero, 1, e8, mf8, ta, ma
-; ZIP-NEXT: vmv.v.i v0, 8
-; ZIP-NEXT: vsetivli zero, 4, e64, m4, ta, ma
-; ZIP-NEXT: vslidedown.vi v16, v8, 4
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v12, v8, 2
-; ZIP-NEXT: vsetivli zero, 1, e8, mf8, ta, ma
-; ZIP-NEXT: vmv.v.i v10, 12
-; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, mu
-; ZIP-NEXT: vslideup.vi v14, v16, 2
-; ZIP-NEXT: vslideup.vi v14, v16, 1, v0.t
-; ZIP-NEXT: ri.vzip2a.vv v18, v8, v12
-; ZIP-NEXT: vmv1r.v v0, v10
-; ZIP-NEXT: vmerge.vvm v12, v18, v14, v0
-; ZIP-NEXT: vsetivli zero, 2, e64, m2, ta, ma
-; ZIP-NEXT: vslidedown.vi v14, v16, 2
-; ZIP-NEXT: vsetivli zero, 1, e8, mf8, ta, ma
-; ZIP-NEXT: vmv.v.i v0, 2
-; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, mu
-; ZIP-NEXT: ri.vzip2a.vv v18, v16, v14
-; ZIP-NEXT: vslidedown.vi v14, v8, 1
-; ZIP-NEXT: vslidedown.vi v14, v8, 2, v0.t
-; ZIP-NEXT: vmv1r.v v0, v10
-; ZIP-NEXT: vmerge.vvm v10, v14, v18, v0
-; ZIP-NEXT: vmv2r.v v8, v12
+; ZIP-NEXT: vsetivli zero, 4, e64, m2, ta, ma
+; ZIP-NEXT: ri.vunzip2a.vv v12, v8, v10
+; ZIP-NEXT: ri.vunzip2b.vv v16, v8, v10
+; ZIP-NEXT: vmv.v.v v8, v12
+; ZIP-NEXT: vmv.v.v v10, v16
; ZIP-NEXT: ret
%retval = call {<4 x double>, <4 x double>} @llvm.vector.deinterleave2.v8f64(<8 x double> %vec)
ret {<4 x double>, <4 x double>} %retval
|
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.
If we have xrivosvizip, we can use the vunzip2{a,b} instructions for these cases provided that we can prove the layout in the two registers matches the fixed length semantics.
The majority of this patch is a straight-forward port of the existing vnsrl logic which has the same requirement (though for slightly different reasoning).
The one complicated bit is the addition of the scalable splitting logic inside lowerVZIP to exploit the independent register operands, and allow the use of lower LMUL. This bit is annoyingly complicated, and really "should" be a DAG combine - except that the VL and mask reduction becomes hard when it's not known to be a constant.