Skip to content

[LoongArch] Split 256-bit build_vector to avoid using LASX element insertion #154918

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 6 commits into from
Sep 22, 2025
Merged

[LoongArch] Split 256-bit build_vector to avoid using LASX element insertion #154918

merged 6 commits into from
Sep 22, 2025

Conversation

@zhaoqi5
Copy link
Contributor

@zhaoqi5 zhaoqi5 commented Aug 22, 2025 •
edited
Loading

Note: Only worse for v8i32/v8f32/v4i64/v4f64 types when the high
part only has one non-undef element. Skip spliting to avoid this.

@llvmbot
Copy link
Member

llvmbot commented Aug 22, 2025

@llvm/pr-subscribers-backend-loongarch

Author: ZhaoQi (zhaoqi5)

Changes

Note: Only worse for v8i32/v8f32/v4i64/v4f64 types when the high
part only has one non-undef element.

Patch is 45.89 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/154918.diff

7 Files Affected:

  • (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+33-10)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/bitreverse.ll (+38-34)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/build-vector.ll (+164-304)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/fpowi.ll (+69-70)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/ir-instruction/fix-xvshuf.ll (+7-6)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insertelement.ll (+4-6)
  • (modified) llvm/test/CodeGen/LoongArch/lasx/scalar-to-vector.ll (+2-2)
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 0499e2c3f1d4c..53be8e5804229 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -2578,24 +2578,47 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
       return DAG.getBitcast(ResTy, SplatVec);
     }
 
-    // Use INSERT_VECTOR_ELT operations rather than expand to stores.
-    // The resulting code is the same length as the expansion, but it doesn't
-    // use memory operations.
-    assert(ResTy.isVector());
+    // Use INSERT_VECTOR_ELT operations rather than expand to stores, because
+    // using memory operations is much lower.
+    EVT VecTy = ResTy;
+    unsigned VecNumElts = NumElts;
+
+    // Split the 256-bits vector and fill them separately, concat the two parts
+    // to get the result vector.
+    if (Is256Vec) {
+      VecTy = ResTy.getHalfNumVectorElementsVT(*DAG.getContext());
+      VecNumElts = NumElts / 2;
+    }
 
+    SDValue Vector = DAG.getUNDEF(VecTy);
     SDValue Op0 = Node->getOperand(0);
-    SDValue Vector = DAG.getUNDEF(ResTy);
-
     if (!Op0.isUndef())
-      Vector = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ResTy, Op0);
-    for (unsigned i = 1; i < NumElts; ++i) {
+      Vector = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecTy, Op0);
+    for (unsigned i = 1; i < VecNumElts; ++i) {
       SDValue Opi = Node->getOperand(i);
       if (Opi.isUndef())
         continue;
-      Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector, Opi,
+      Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecTy, Vector, Opi,
                            DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
     }
-    return Vector;
+
+    if (Is128Vec)
+      return Vector;
+
+    SDValue VectorHi = DAG.getUNDEF(VecTy);
+    SDValue OpHi0 = Node->getOperand(VecNumElts);
+    if (!OpHi0.isUndef())
+      VectorHi = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecTy, OpHi0);
+    for (unsigned i = VecNumElts + 1; i < NumElts; ++i) {
+      SDValue Opi = Node->getOperand(i);
+      if (Opi.isUndef())
+        continue;
+      VectorHi = DAG.getNode(
+          ISD::INSERT_VECTOR_ELT, DL, VecTy, VectorHi, Opi,
+          DAG.getConstant(i - VecNumElts, DL, Subtarget.getGRLenVT()));
+    }
+
+    return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResTy, Vector, VectorHi);
   }
 
   return SDValue();
diff --git a/llvm/test/CodeGen/LoongArch/lasx/bitreverse.ll b/llvm/test/CodeGen/LoongArch/lasx/bitreverse.ll
index 11f1bce55fad6..87ee4ad025395 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/bitreverse.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/bitreverse.ll
@@ -7,18 +7,19 @@ declare <32 x i8> @llvm.bitreverse.v32i8(<32 x i8>)
 define <32 x i8> @test_bitreverse_v32i8(<32 x i8> %a) nounwind {
 ; CHECK-LABEL: test_bitreverse_v32i8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
-; CHECK-NEXT:    bitrev.8b $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 0
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
-; CHECK-NEXT:    bitrev.8b $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 1
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 2
 ; CHECK-NEXT:    bitrev.8b $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 2
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 0
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
 ; CHECK-NEXT:    bitrev.8b $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 3
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 1
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
+; CHECK-NEXT:    bitrev.8b $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 0
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
+; CHECK-NEXT:    bitrev.8b $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 1
+; CHECK-NEXT:    xvpermi.q $xr1, $xr2, 2
 ; CHECK-NEXT:    xvori.b $xr0, $xr1, 0
 ; CHECK-NEXT:    ret
   %b = call <32 x i8> @llvm.bitreverse.v32i8(<32 x i8> %a)
@@ -30,19 +31,20 @@ declare <16 x i16> @llvm.bitreverse.v16i16(<16 x i16>)
 define <16 x i16> @test_bitreverse_v16i16(<16 x i16> %a) nounwind {
 ; CHECK-LABEL: test_bitreverse_v16i16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 0
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 1
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 2
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 2
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 0
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 3
-; CHECK-NEXT:    xvshuf4i.h $xr0, $xr1, 27
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 1
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 0
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 1
+; CHECK-NEXT:    xvpermi.q $xr2, $xr1, 2
+; CHECK-NEXT:    xvshuf4i.h $xr0, $xr2, 27
 ; CHECK-NEXT:    ret
   %b = call <16 x i16> @llvm.bitreverse.v16i16(<16 x i16> %a)
   ret <16 x i16> %b
@@ -53,19 +55,20 @@ declare <8 x i32> @llvm.bitreverse.v8i32(<8 x i32>)
 define <8 x i32> @test_bitreverse_v8i32(<8 x i32> %a) nounwind {
 ; CHECK-LABEL: test_bitreverse_v8i32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 0
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 1
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 2
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 2
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 0
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 3
-; CHECK-NEXT:    xvshuf4i.w $xr0, $xr1, 177
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 1
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 0
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 1
+; CHECK-NEXT:    xvpermi.q $xr2, $xr1, 2
+; CHECK-NEXT:    xvshuf4i.w $xr0, $xr2, 177
 ; CHECK-NEXT:    ret
   %b = call <8 x i32> @llvm.bitreverse.v8i32(<8 x i32> %a)
   ret <8 x i32> %b
@@ -76,18 +79,19 @@ declare <4 x i64> @llvm.bitreverse.v4i64(<4 x i64>)
 define <4 x i64> @test_bitreverse_v4i64(<4 x i64> %a) nounwind {
 ; CHECK-LABEL: test_bitreverse_v4i64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 0
-; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
-; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 1
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 2
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 2
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 0
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
 ; CHECK-NEXT:    bitrev.d $a0, $a0
-; CHECK-NEXT:    xvinsgr2vr.d $xr1, $a0, 3
+; CHECK-NEXT:    vinsgr2vr.d $vr2, $a0, 1
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 0
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 0
+; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 1
+; CHECK-NEXT:    bitrev.d $a0, $a0
+; CHECK-NEXT:    vinsgr2vr.d $vr1, $a0, 1
+; CHECK-NEXT:    xvpermi.q $xr1, $xr2, 2
 ; CHECK-NEXT:    xvori.b $xr0, $xr1, 0
 ; CHECK-NEXT:    ret
   %b = call <4 x i64> @llvm.bitreverse.v4i64(<4 x i64> %a)
diff --git a/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll b/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
index e759f5c98f1e7..83aaf3376cb29 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
@@ -224,144 +224,65 @@ entry:
 define void @buildvector_v32i8(ptr %dst, i8 %a0, i8 %a1, i8 %a2, i8 %a3, i8 %a4, i8 %a5, i8 %a6, i8 %a7, i8 %a8, i8 %a9, i8 %a10, i8 %a11, i8 %a12, i8 %a13, i8 %a14, i8 %a15, i8 %a16, i8 %a17, i8 %a18, i8 %a19, i8 %a20, i8 %a21, i8 %a22, i8 %a23, i8 %a24, i8 %a25, i8 %a26, i8 %a27, i8 %a28, i8 %a29, i8 %a30, i8 %a31) nounwind {
 ; CHECK-LABEL: buildvector_v32i8:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    addi.d $sp, $sp, -80
-; CHECK-NEXT:    fst.d $fs0, $sp, 72 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs1, $sp, 64 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs2, $sp, 56 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs3, $sp, 48 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs4, $sp, 40 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs5, $sp, 32 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs6, $sp, 24 # 8-byte Folded Spill
-; CHECK-NEXT:    fst.d $fs7, $sp, 16 # 8-byte Folded Spill
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 0
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a2
-; CHECK-NEXT:    xvreplgr2vr.b $xr2, $a3
-; CHECK-NEXT:    xvreplgr2vr.b $xr3, $a4
-; CHECK-NEXT:    ld.b $a1, $sp, 264
-; CHECK-NEXT:    xvreplgr2vr.b $xr4, $a5
-; CHECK-NEXT:    ld.b $a2, $sp, 80
-; CHECK-NEXT:    xvreplgr2vr.b $xr5, $a6
-; CHECK-NEXT:    ld.b $a3, $sp, 88
-; CHECK-NEXT:    xvreplgr2vr.b $xr6, $a7
-; CHECK-NEXT:    ld.b $a4, $sp, 96
-; CHECK-NEXT:    xvreplgr2vr.b $xr7, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 104
-; CHECK-NEXT:    xvreplgr2vr.b $xr8, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 112
-; CHECK-NEXT:    xvreplgr2vr.b $xr9, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 120
-; CHECK-NEXT:    xvreplgr2vr.b $xr10, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 128
-; CHECK-NEXT:    xvreplgr2vr.b $xr11, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 136
-; CHECK-NEXT:    xvreplgr2vr.b $xr12, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 144
-; CHECK-NEXT:    xvreplgr2vr.b $xr13, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 152
-; CHECK-NEXT:    xvreplgr2vr.b $xr14, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 160
-; CHECK-NEXT:    xvreplgr2vr.b $xr15, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 168
-; CHECK-NEXT:    xvreplgr2vr.b $xr16, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 176
-; CHECK-NEXT:    xvreplgr2vr.b $xr17, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 184
-; CHECK-NEXT:    xvreplgr2vr.b $xr18, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 192
-; CHECK-NEXT:    xvreplgr2vr.b $xr19, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 200
-; CHECK-NEXT:    xvreplgr2vr.b $xr20, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 208
-; CHECK-NEXT:    xvreplgr2vr.b $xr21, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 216
-; CHECK-NEXT:    xvreplgr2vr.b $xr22, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 224
-; CHECK-NEXT:    xvreplgr2vr.b $xr23, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 232
-; CHECK-NEXT:    xvreplgr2vr.b $xr24, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 240
-; CHECK-NEXT:    xvreplgr2vr.b $xr25, $a2
-; CHECK-NEXT:    ld.b $a2, $sp, 248
-; CHECK-NEXT:    xvreplgr2vr.b $xr26, $a3
-; CHECK-NEXT:    ld.b $a3, $sp, 256
-; CHECK-NEXT:    xvreplgr2vr.b $xr27, $a4
-; CHECK-NEXT:    ld.b $a4, $sp, 272
-; CHECK-NEXT:    xvreplgr2vr.b $xr28, $a2
-; CHECK-NEXT:    xvreplgr2vr.b $xr29, $a3
-; CHECK-NEXT:    xvreplgr2vr.b $xr30, $a1
-; CHECK-NEXT:    xvreplgr2vr.b $xr31, $a4
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 17
-; CHECK-NEXT:    xvpermi.q $xr2, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr2, 34
-; CHECK-NEXT:    xvpermi.q $xr3, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr3, 51
-; CHECK-NEXT:    xvpermi.q $xr4, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr4, 68
-; CHECK-NEXT:    xvpermi.q $xr5, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr5, 85
-; CHECK-NEXT:    xvpermi.q $xr6, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr6, 102
-; CHECK-NEXT:    xvpermi.q $xr7, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr7, 119
-; CHECK-NEXT:    xvpermi.q $xr8, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr8, 136
-; CHECK-NEXT:    xvpermi.q $xr9, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr9, 153
-; CHECK-NEXT:    xvpermi.q $xr10, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr10, 170
-; CHECK-NEXT:    xvpermi.q $xr11, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr11, 187
-; CHECK-NEXT:    xvpermi.q $xr12, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr12, 204
-; CHECK-NEXT:    xvpermi.q $xr13, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr13, 221
-; CHECK-NEXT:    xvpermi.q $xr14, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr14, 238
-; CHECK-NEXT:    xvpermi.q $xr15, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr15, 255
-; CHECK-NEXT:    xvpermi.q $xr16, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr16, 0
-; CHECK-NEXT:    xvpermi.q $xr17, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr17, 17
-; CHECK-NEXT:    xvpermi.q $xr18, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr18, 34
-; CHECK-NEXT:    xvpermi.q $xr19, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr19, 51
-; CHECK-NEXT:    xvpermi.q $xr20, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr20, 68
-; CHECK-NEXT:    xvpermi.q $xr21, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr21, 85
-; CHECK-NEXT:    xvpermi.q $xr22, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr22, 102
-; CHECK-NEXT:    xvpermi.q $xr23, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr23, 119
-; CHECK-NEXT:    xvpermi.q $xr24, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr24, 136
-; CHECK-NEXT:    xvpermi.q $xr25, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr25, 153
-; CHECK-NEXT:    xvpermi.q $xr26, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr26, 170
-; CHECK-NEXT:    xvpermi.q $xr27, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr27, 187
-; CHECK-NEXT:    xvpermi.q $xr28, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr28, 204
-; CHECK-NEXT:    xvpermi.q $xr29, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr29, 221
-; CHECK-NEXT:    xvpermi.q $xr30, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr30, 238
-; CHECK-NEXT:    xvpermi.q $xr31, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr31, 255
-; CHECK-NEXT:    xvst $xr0, $a0, 0
-; CHECK-NEXT:    fld.d $fs7, $sp, 16 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs6, $sp, 24 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs5, $sp, 32 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs4, $sp, 40 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs3, $sp, 48 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs2, $sp, 56 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs1, $sp, 64 # 8-byte Folded Reload
-; CHECK-NEXT:    fld.d $fs0, $sp, 72 # 8-byte Folded Reload
-; CHECK-NEXT:    addi.d $sp, $sp, 80
+; CHECK-NEXT:    ld.b $t0, $sp, 72
+; CHECK-NEXT:    ld.b $t1, $sp, 184
+; CHECK-NEXT:    ld.b $t2, $sp, 80
+; CHECK-NEXT:    ld.b $t3, $sp, 88
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t0, 0
+; CHECK-NEXT:    ld.b $t0, $sp, 96
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t2, 1
+; CHECK-NEXT:    ld.b $t2, $sp, 104
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t3, 2
+; CHECK-NEXT:    ld.b $t3, $sp, 112
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t0, 3
+; CHECK-NEXT:    ld.b $t0, $sp, 120
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t2, 4
+; CHECK-NEXT:    ld.b $t2, $sp, 128
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t3, 5
+; CHECK-NEXT:    ld.b $t3, $sp, 136
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t0, 6
+; CHECK-NEXT:    ld.b $t0, $sp, 144
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t2, 7
+; CHECK-NEXT:    ld.b $t2, $sp, 152
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t3, 8
+; CHECK-NEXT:    ld.b $t3, $sp, 160
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t0, 9
+; CHECK-NEXT:    ld.b $t0, $sp, 168
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t2, 10
+; CHECK-NEXT:    ld.b $t2, $sp, 176
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t3, 11
+; CHECK-NEXT:    ld.b $t3, $sp, 192
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t0, 12
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t2, 13
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t1, 14
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $t3, 15
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a1, 0
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a2, 1
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a3, 2
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a4, 3
+; CHECK-NEXT:    ld.b $a1, $sp, 56
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a5, 4
+; CHECK-NEXT:    ld.b $a2, $sp, 0
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a6, 5
+; CHECK-NEXT:    ld.b $a3, $sp, 8
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a7, 6
+; CHECK-NEXT:    ld.b $a4, $sp, 16
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a2, 7
+; CHECK-NEXT:    ld.b $a2, $sp, 24
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a3, 8
+; CHECK-NEXT:    ld.b $a3, $sp, 32
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a4, 9
+; CHECK-NEXT:    ld.b $a4, $sp, 40
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a2, 10
+; CHECK-NEXT:    ld.b $a2, $sp, 48
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a3, 11
+; CHECK-NEXT:    ld.b $a3, $sp, 64
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a4, 12
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a2, 13
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a1, 14
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $a3, 15
+; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 2
+; CHECK-NEXT:    xvst $xr1, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
   %ins0  = insertelement <32 x i8> undef,  i8 %a0,  i32 0
@@ -412,48 +333,21 @@ define void @buildvector_v32i8_partial(ptr %dst, i8 %a0, i8 %a1, i8 %a2, i8 %a5,
 ; CHECK-NEXT:    ld.b $t6, $sp, 8
 ; CHECK-NEXT:    ld.b $t7, $sp, 0
 ; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 0
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a2
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 17
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a3
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 34
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a4
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 85
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a5
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 119
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a6
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 136
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $a7
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 18
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 255
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t7
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 17
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t6
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 34
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t5
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 68
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t4
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 102
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t3
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 119
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t2
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 187
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t1
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 204
-; CHECK-NEXT:    xvreplgr2vr.b $xr1, $t0
-; CHECK-NEXT:    xvpermi.q $xr1, $xr0, 48
-; CHECK-NEXT:    xvextrins.b $xr0, $xr1, 255
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a2, 1
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a3, 2
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a4, 5
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a5, 7
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a6, 8
+; CHECK-NEXT:    vinsgr2vr.b $vr0, $a7, 15
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t7, 1
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t6, 2
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t5, 4
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t4, 6
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t3, 7
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t2, 11
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t1, 12
+; CHECK-NEXT:    vinsgr2vr.b $vr1, $t0, 15
+; CHECK-NEXT:    xvpermi.q $xr0, $xr1, 2
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -804,62 +698,33 @@ entry:
 define void @buildvector_v16i16(ptr %dst, i16 %a0, i16 %a1, i16 %a2, i16 %a3, i16 %a4, i16 %a5, i16 %a6, i16 %a7, i16 %a8, i16 %a9, i16 %a10, i16 %a11, i16 %a12, i16 %a13, i16 %a14, i16 %a15) nounwind {
 ; CHECK-LABEL: buildvector_v16i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    ld.h $t0, $sp, 64
-; CHECK-NEXT:    ld.h $t1, $sp, 56
-; CHECK-NEXT:    ld.h $t2, $sp, 48
-; CHECK-NEXT:    ld.h $t3, $sp, 40
-; CHECK-NEXT:    ld.h $t4, $sp, 32
-; CHECK-NEXT:    ld.h $t5, $sp, 24
+; CHECK-NEXT:    ld.h $t0, $sp, 0
+; CHECK-NEXT:    ld.h $t1, $sp, 64
+; CHECK-NEXT:    ld.h $t2, $sp, 56
+; CHECK-NEXT:    ld.h $t3, $sp, 48
+; CHECK-NEXT:    ld.h $t4, $sp, 40
+; CHECK-NEXT:    ld.h $t5, $sp, 8
 ;...
[truncated]

@zhaoqi5 zhaoqi5 changed the title [LoongArch] Spill 256-bit build_vector to avoid using LASX element insertion [LoongArch] Split 256-bit build_vector to avoid using LASX element insertion Aug 23, 2025
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/broadcast-repeated-buildvector branch from 6e52551 to ca08e88 Compare August 28, 2025 08:59
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/general-256-buildvector-split branch from cdad7e6 to 6ef9fd8 Compare August 28, 2025 09:05
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/broadcast-repeated-buildvector branch from ca08e88 to 0e3aa0d Compare September 2, 2025 02:21
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/general-256-buildvector-split branch from 6ef9fd8 to 8cd1a2a Compare September 2, 2025 03:26
zhaoqi5
zhaoqi5 commented Sep 2, 2025
View reviewed changes
llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
Comment on lines -1914 to +1982
APInt Imm(64, SplatIndex);
return DAG.getNode(LoongArchISD::VREPLVEI, DL, VT, V1,
DAG.getConstant(Imm, DL, Subtarget.getGRLenVT()));
DAG.getConstant(SplatIndex, DL, Subtarget.getGRLenVT()));
Copy link
Contributor Author

Choose a reason for hiding this comment

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

Tests for la32 passed.

@zhaoqi5 zhaoqi5 requested a review from heiher September 16, 2025 12:22
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/broadcast-repeated-buildvector branch from 0e3aa0d to f028d60 Compare September 17, 2025 10:12
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/general-256-buildvector-split branch from 8cd1a2a to 254a642 Compare September 17, 2025 10:19
@zhaoqi5
Copy link
Contributor Author

zhaoqi5 commented Sep 17, 2025

ping

Base automatically changed from users/zhaoqi5/broadcast-repeated-buildvector to main September 22, 2025 08:04
@zhaoqi5
[LoongArch] Spill 256-bit build_vector to avoid using LASX element in…
13fae6b 
…sertion

Note: Only worse for v8i32/v8f32/v4i64/v4f64 types when the high
part only has one non-undef element.
@zhaoqi5
update
993237b
@zhaoqi5
update
45a0adb
@zhaoqi5
fix typo
7bf017f
@zhaoqi5
la32 passed
580c2ad
@zhaoqi5
update tests
fa11be3
@zhaoqi5 zhaoqi5 force-pushed the users/zhaoqi5/general-256-buildvector-split branch from 254a642 to fa11be3 Compare September 22, 2025 08:10
@zhaoqi5 zhaoqi5 merged commit b280cf5 into main Sep 22, 2025
9 checks passed
@zhaoqi5 zhaoqi5 deleted the users/zhaoqi5/general-256-buildvector-split branch September 22, 2025 09:02
DavidSpickett added a commit that referenced this pull request Oct 3, 2025
@DavidSpickett
[llvm][LoongArch] Fix compiler warning produced by assert
b4f54bf 
Fixes #154918

tion ‘llvm::SDValue fillSubVectorFromBuildVector(llvm::BuildVectorSDNode*, llvm::SelectionDAG&, llvm::SDLoc, const llvm::LoongArchSubtarget&, llvm::EVT, unsigned int)’:
/home/davspi01/work/open_source/llvm-project/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp:2863:16: warning: comparison of unsigned expression in ‘>= 0’ is always true [-Wtype-limits]
 2863 |   assert(first >= 0 &&
      |          ~~~~~~^~~~

first is unsigned so this part of the expression is redundant.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@DavidSpickett DavidSpickett DavidSpickett left review comments

@SixWeining SixWeining SixWeining approved these changes

@tangaac tangaac Awaiting requested review from tangaac

@heiher heiher Awaiting requested review from heiher

Assignees

No one assigned

Labels

backend:loongarch

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

5 participants

@zhaoqi5 @llvmbot @DavidSpickett @SixWeining