[llvm] [WebAssembly] extadd_pairwise for PartialReduce (PR #157669)
Sam Parker via llvm-commits
llvm-commits at lists.llvm.org
Tue Sep 9 06:17:10 PDT 2025
https://github.com/sparker-arm updated https://github.com/llvm/llvm-project/pull/157669
>From c2aba6ff03eaa263fc2dbc327bef98ea35640575 Mon Sep 17 00:00:00 2001
From: Sam Parker <sam.parker at arm.com>
Date: Tue, 9 Sep 2025 13:49:38 +0100
Subject: [PATCH 1/2] [WebAssembly] extadd_pairwise for PartialReduce
Avoid using extends, and adding the high and low half and use
extadd_pairwise instead.
---
llvm/lib/Target/WebAssembly/WebAssemblyISD.def | 1 +
.../Target/WebAssembly/WebAssemblyISelLowering.cpp | 9 ++-------
llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td | 11 +++++++++--
.../CodeGen/WebAssembly/int-mac-reduction-loops.ll | 8 ++------
4 files changed, 14 insertions(+), 15 deletions(-)
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyISD.def b/llvm/lib/Target/WebAssembly/WebAssemblyISD.def
index 378ef2c8f250e..1eae3586d16b8 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyISD.def
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyISD.def
@@ -27,6 +27,7 @@ HANDLE_NODETYPE(WrapperREL)
HANDLE_NODETYPE(BR_IF)
HANDLE_NODETYPE(BR_TABLE)
HANDLE_NODETYPE(DOT)
+HANDLE_NODETYPE(EXT_ADD_PAIRWISE_U)
HANDLE_NODETYPE(SHUFFLE)
HANDLE_NODETYPE(SWIZZLE)
HANDLE_NODETYPE(VEC_SHL)
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
index 5a45134692865..7827d79697a94 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
@@ -2183,13 +2183,8 @@ SDValue performLowerPartialReduction(SDNode *N, SelectionDAG &DAG) {
SDValue MulLow = DAG.getNode(ISD::MUL, DL, MVT::v8i16, LowLHS, LowRHS);
SDValue MulHigh = DAG.getNode(ISD::MUL, DL, MVT::v8i16, HighLHS, HighRHS);
- SDValue LowLow = DAG.getNode(LowOpc, DL, MVT::v4i32, MulLow);
- SDValue LowHigh = DAG.getNode(LowOpc, DL, MVT::v4i32, MulHigh);
- SDValue HighLow = DAG.getNode(HighOpc, DL, MVT::v4i32, MulLow);
- SDValue HighHigh = DAG.getNode(HighOpc, DL, MVT::v4i32, MulHigh);
-
- SDValue AddLow = DAG.getNode(ISD::ADD, DL, MVT::v4i32, LowLow, HighLow);
- SDValue AddHigh = DAG.getNode(ISD::ADD, DL, MVT::v4i32, LowHigh, HighHigh);
+ SDValue AddLow = DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL, MVT::v4i32, MulLow);
+ SDValue AddHigh = DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL, MVT::v4i32, MulHigh);
SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::v4i32, AddLow, AddHigh);
return DAG.getNode(ISD::ADD, DL, MVT::v4i32, N->getOperand(1), Add);
}
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index f06f8d5174e3e..3c26b453c4482 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -1453,15 +1453,22 @@ if !ne(t1, t2) then
def : Pat<(t1.vt (bitconvert (t2.vt V128:$v))), (t1.vt V128:$v)>;
// Extended pairwise addition
+def extadd_pairwise_u : SDNode<"WebAssemblyISD::EXT_ADD_PAIRWISE_U", extend_t>;
+
defm "" : SIMDConvert<I16x8, I8x16, int_wasm_extadd_pairwise_signed,
"extadd_pairwise_i8x16_s", 0x7c>;
-defm "" : SIMDConvert<I16x8, I8x16, int_wasm_extadd_pairwise_unsigned,
+defm "" : SIMDConvert<I16x8, I8x16, extadd_pairwise_u,
"extadd_pairwise_i8x16_u", 0x7d>;
defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_signed,
"extadd_pairwise_i16x8_s", 0x7e>;
-defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_unsigned,
+defm "" : SIMDConvert<I32x4, I16x8, extadd_pairwise_u,
"extadd_pairwise_i16x8_u", 0x7f>;
+def : Pat<(v4i32 (int_wasm_extadd_pairwise_unsigned (v8i16 V128:$in))),
+ (extadd_pairwise_u_I32x4 V128:$in)>;
+def : Pat<(v8i16 (int_wasm_extadd_pairwise_unsigned (v16i8 V128:$in))),
+ (extadd_pairwise_u_I16x8 V128:$in)>;
+
// f64x2 <-> f32x4 conversions
def demote_t : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;
def demote_zero : SDNode<"WebAssemblyISD::DEMOTE_ZERO", demote_t>;
diff --git a/llvm/test/CodeGen/WebAssembly/int-mac-reduction-loops.ll b/llvm/test/CodeGen/WebAssembly/int-mac-reduction-loops.ll
index 0184e22a3b40d..04a2268db1755 100644
--- a/llvm/test/CodeGen/WebAssembly/int-mac-reduction-loops.ll
+++ b/llvm/test/CodeGen/WebAssembly/int-mac-reduction-loops.ll
@@ -191,13 +191,9 @@ define hidden i32 @i32_mac_u8(ptr nocapture noundef readonly %a, ptr nocapture n
; MAX-BANDWIDTH: v128.load
; MAX-BANDWIDTH: v128.load
; MAX-BANDWIDTH: i16x8.extmul_low_i8x16_u
-; MAX-BANDWIDTH: i32x4.extend_low_i16x8_u
-; MAX-BANDWIDTH: i32x4.extend_high_i16x8_u
-; MAX-BANDWIDTH: i32x4.add
+; MAX-BANDWIDTH: i32x4.extadd_pairwise_i16x8_u
; MAX-BANDWIDTH: i16x8.extmul_high_i8x16_u
-; MAX-BANDWIDTH: i32x4.extend_low_i16x8_u
-; MAX-BANDWIDTH: i32x4.extend_high_i16x8_u
-; MAX-BANDWIDTH: i32x4.add
+; MAX-BANDWIDTH: i32x4.extadd_pairwise_i16x8_u
; MAX-BANDWIDTH: i32x4.add
; MAX-BANDWIDTH: i32x4.add
>From fbf9eccd4c033a8563c36b7747a5c894ec3614c3 Mon Sep 17 00:00:00 2001
From: Sam Parker <sam.parker at arm.com>
Date: Tue, 9 Sep 2025 14:16:51 +0100
Subject: [PATCH 2/2] run clang-format
---
llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
index 7827d79697a94..fe100dab427ef 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
@@ -2183,8 +2183,10 @@ SDValue performLowerPartialReduction(SDNode *N, SelectionDAG &DAG) {
SDValue MulLow = DAG.getNode(ISD::MUL, DL, MVT::v8i16, LowLHS, LowRHS);
SDValue MulHigh = DAG.getNode(ISD::MUL, DL, MVT::v8i16, HighLHS, HighRHS);
- SDValue AddLow = DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL, MVT::v4i32, MulLow);
- SDValue AddHigh = DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL, MVT::v4i32, MulHigh);
+ SDValue AddLow =
+ DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL, MVT::v4i32, MulLow);
+ SDValue AddHigh = DAG.getNode(WebAssemblyISD::EXT_ADD_PAIRWISE_U, DL,
+ MVT::v4i32, MulHigh);
SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::v4i32, AddLow, AddHigh);
return DAG.getNode(ISD::ADD, DL, MVT::v4i32, N->getOperand(1), Add);
}
More information about the llvm-commits
mailing list