[llvm] [RISCV] Eagerly optimize scalar packing for buildvector lowering (PR #156062)

Philip Reames via llvm-commits llvm-commits at lists.llvm.org
Sat Aug 30 18:59:29 PDT 2025


https://github.com/preames updated https://github.com/llvm/llvm-project/pull/156062

>From 482102c02639acd6d46b6a9c4fa6fb2ed901e468 Mon Sep 17 00:00:00 2001
From: Philip Reames <preames at rivosinc.com>
Date: Fri, 29 Aug 2025 10:02:14 -0700
Subject: [PATCH 1/4] [RISCV] Eagerly optimize scalar packing for buildvector
 lowering

Instead of relying on DAG to cleanup the redundant ANDs, go ahead
actively try to prove them redundant at construction.

I noticed this because SimplifyDemandedBits was dropping the
disjoint on the OR, but there's at least one other codegen diff
in the tests as well.
---
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp               | 7 +++++--
 llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll | 2 +-
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 4c39bcf8494a4..80083f843f9af 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4257,6 +4257,7 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
   MVT XLenVT = Subtarget.getXLenVT();
   SDValue Mask = DAG.getConstant(
       APInt::getLowBitsSet(XLenVT.getSizeInBits(), ElemSizeInBits), DL, XLenVT);
+  unsigned ZeroPrefix = XLenVT.getSizeInBits() - ElemSizeInBits;
   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
@@ -4270,8 +4271,10 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
                              ElemDL, XLenVT, A, B),
           0);
 
-    A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
-    B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
+    if (DAG.computeKnownBits(A).countMinLeadingZeros() < ZeroPrefix)
+      A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
+    if (DAG.computeKnownBits(B).countMinLeadingZeros() < ZeroPrefix)
+      B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
     SDValue ShtAmt = DAG.getConstant(ElemSizeInBits, ElemDL, XLenVT);
     return DAG.getNode(ISD::OR, ElemDL, XLenVT, A,
                        DAG.getNode(ISD::SHL, ElemDL, XLenVT, B, ShtAmt),
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 d9bb007a10f71..8aaffd41cbd16 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
@@ -2629,7 +2629,7 @@ define <16 x i8> @buildvec_v16i8_undef_edges(ptr %p) {
 ; RVA22U64-NEXT:    or a0, a0, a4
 ; RVA22U64-NEXT:    slli a6, a6, 24
 ; RVA22U64-NEXT:    or a1, a1, a2
-; RVA22U64-NEXT:    add.uw a1, a6, a1
+; RVA22U64-NEXT:    or a1, a6, a1
 ; RVA22U64-NEXT:    or a0, a0, a3
 ; RVA22U64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RVA22U64-NEXT:    vmv.v.x v8, a1

>From 1fef56edae697f51dd3f03d207d2f1df34c272f4 Mon Sep 17 00:00:00 2001
From: Philip Reames <preames at rivosinc.com>
Date: Fri, 29 Aug 2025 10:15:35 -0700
Subject: [PATCH 2/4] Missing comment

---
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 80083f843f9af..efd0d779680d4 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4271,6 +4271,8 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
                              ElemDL, XLenVT, A, B),
           0);
 
+    // Manually optimize away the ANDs if we can, DAGCombiner will
+    // sometimes end of perturbing codegen if we don't.
     if (DAG.computeKnownBits(A).countMinLeadingZeros() < ZeroPrefix)
       A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
     if (DAG.computeKnownBits(B).countMinLeadingZeros() < ZeroPrefix)

>From 00d1907b55cf9ede90bae25a33a3a9902be304f3 Mon Sep 17 00:00:00 2001
From: Philip Reames <preames at rivosinc.com>
Date: Sat, 30 Aug 2025 18:51:10 -0700
Subject: [PATCH 3/4] Update llvm/lib/Target/RISCV/RISCVISelLowering.cpp

Co-authored-by: Craig Topper <craig.topper at sifive.com>
---
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index efd0d779680d4..34fce5da829e9 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4272,7 +4272,7 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
           0);
 
     // Manually optimize away the ANDs if we can, DAGCombiner will
-    // sometimes end of perturbing codegen if we don't.
+    // sometimes end up perturbing codegen if we don't.
     if (DAG.computeKnownBits(A).countMinLeadingZeros() < ZeroPrefix)
       A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
     if (DAG.computeKnownBits(B).countMinLeadingZeros() < ZeroPrefix)

>From bd98ff06fda7632313818cb744cede44c73d5b8b Mon Sep 17 00:00:00 2001
From: Philip Reames <preames at rivosinc.com>
Date: Sat, 30 Aug 2025 18:53:59 -0700
Subject: [PATCH 4/4] Address review comment

---
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 34fce5da829e9..d44d8bc289c8d 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4257,7 +4257,6 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
   MVT XLenVT = Subtarget.getXLenVT();
   SDValue Mask = DAG.getConstant(
       APInt::getLowBitsSet(XLenVT.getSizeInBits(), ElemSizeInBits), DL, XLenVT);
-  unsigned ZeroPrefix = XLenVT.getSizeInBits() - ElemSizeInBits;
   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
@@ -4273,9 +4272,9 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
 
     // Manually optimize away the ANDs if we can, DAGCombiner will
     // sometimes end up perturbing codegen if we don't.
-    if (DAG.computeKnownBits(A).countMinLeadingZeros() < ZeroPrefix)
+    if (DAG.computeKnownBits(A).countMaxActiveBits() > ElemSizeInBits)
       A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
-    if (DAG.computeKnownBits(B).countMinLeadingZeros() < ZeroPrefix)
+    if (DAG.computeKnownBits(B).countMaxActiveBits() > ElemSizeInBits)
       B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
     SDValue ShtAmt = DAG.getConstant(ElemSizeInBits, ElemDL, XLenVT);
     return DAG.getNode(ISD::OR, ElemDL, XLenVT, A,



More information about the llvm-commits mailing list