[llvm] [AArch64] Avoid NEON dot product in streaming[-compatible] functions (PR #101677)
Sander de Smalen via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 2 06:54:49 PDT 2024
https://github.com/sdesmalen-arm created https://github.com/llvm/llvm-project/pull/101677
The NEON dot product is not valid in streaming mode.
A follow-up patch will improve codegen for these operations.
>From 79d074a3d5476cea53e98739cb5865752cbb4ab7 Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Fri, 2 Aug 2024 14:39:31 +0100
Subject: [PATCH 1/2] Precommit test
---
...-streaming-mode-fixed-length-reductions.ll | 95 +++++++++++++++++++
1 file changed, 95 insertions(+)
create mode 100644 llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
new file mode 100644
index 0000000000000..7d0918c7023cc
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
@@ -0,0 +1,95 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mattr=+sve < %s | FileCheck %s
+; RUN: llc -mattr=+dotprod,+sve < %s | FileCheck %s -check-prefix=DOT
+; RUN: llc -mattr=+dotprod,+sve -force-streaming-compatible < %s | FileCheck %s --check-prefix=STREAMING-SVE
+; RUN: llc -mattr=+dotprod,+sme -force-streaming < %s | FileCheck %s --check-prefix=STREAMING-SVE
+
+target triple = "aarch64-unknown-linux-gnu"
+
+define i32 @reduce_uaddv_v16i8(<32 x i8> %a) {
+; CHECK-LABEL: reduce_uaddv_v16i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: ushll2 v2.8h, v1.16b, #0
+; CHECK-NEXT: ushll2 v3.8h, v0.16b, #0
+; CHECK-NEXT: ushll v1.8h, v1.8b, #0
+; CHECK-NEXT: ushll v0.8h, v0.8b, #0
+; CHECK-NEXT: uaddl2 v4.4s, v3.8h, v2.8h
+; CHECK-NEXT: uaddl v2.4s, v3.4h, v2.4h
+; CHECK-NEXT: uaddl2 v5.4s, v0.8h, v1.8h
+; CHECK-NEXT: uaddl v0.4s, v0.4h, v1.4h
+; CHECK-NEXT: add v1.4s, v5.4s, v4.4s
+; CHECK-NEXT: add v0.4s, v0.4s, v2.4s
+; CHECK-NEXT: add v0.4s, v0.4s, v1.4s
+; CHECK-NEXT: addv s0, v0.4s
+; CHECK-NEXT: fmov w0, s0
+; CHECK-NEXT: ret
+;
+; DOT-LABEL: reduce_uaddv_v16i8:
+; DOT: // %bb.0:
+; DOT-NEXT: movi v2.16b, #1
+; DOT-NEXT: movi v3.2d, #0000000000000000
+; DOT-NEXT: udot v3.4s, v1.16b, v2.16b
+; DOT-NEXT: udot v3.4s, v0.16b, v2.16b
+; DOT-NEXT: addv s0, v3.4s
+; DOT-NEXT: fmov w0, s0
+; DOT-NEXT: ret
+;
+; STREAMING-SVE-LABEL: reduce_uaddv_v16i8:
+; STREAMING-SVE: // %bb.0:
+; STREAMING-SVE-NEXT: mov z2.b, #1 // =0x1
+; STREAMING-SVE-NEXT: mov z3.s, #0 // =0x0
+; STREAMING-SVE-NEXT: ptrue p0.s, vl4
+; STREAMING-SVE-NEXT: udot v3.4s, v1.16b, v2.16b
+; STREAMING-SVE-NEXT: udot v3.4s, v0.16b, v2.16b
+; STREAMING-SVE-NEXT: uaddv d0, p0, z3.s
+; STREAMING-SVE-NEXT: fmov x0, d0
+; STREAMING-SVE-NEXT: // kill: def $w0 killed $w0 killed $x0
+; STREAMING-SVE-NEXT: ret
+ %1 = zext <32 x i8> %a to <32 x i32>
+ %2 = call i32 @llvm.vector.reduce.add.v32i32(<32 x i32> %1)
+ ret i32 %2
+}
+
+define i32 @reduce_saddv_v16i8(<32 x i8> %a) {
+; CHECK-LABEL: reduce_saddv_v16i8:
+; CHECK: // %bb.0:
+; CHECK-NEXT: sshll2 v2.8h, v1.16b, #0
+; CHECK-NEXT: sshll2 v3.8h, v0.16b, #0
+; CHECK-NEXT: sshll v1.8h, v1.8b, #0
+; CHECK-NEXT: sshll v0.8h, v0.8b, #0
+; CHECK-NEXT: saddl2 v4.4s, v3.8h, v2.8h
+; CHECK-NEXT: saddl v2.4s, v3.4h, v2.4h
+; CHECK-NEXT: saddl2 v5.4s, v0.8h, v1.8h
+; CHECK-NEXT: saddl v0.4s, v0.4h, v1.4h
+; CHECK-NEXT: add v1.4s, v5.4s, v4.4s
+; CHECK-NEXT: add v0.4s, v0.4s, v2.4s
+; CHECK-NEXT: add v0.4s, v0.4s, v1.4s
+; CHECK-NEXT: addv s0, v0.4s
+; CHECK-NEXT: fmov w0, s0
+; CHECK-NEXT: ret
+;
+; DOT-LABEL: reduce_saddv_v16i8:
+; DOT: // %bb.0:
+; DOT-NEXT: movi v2.16b, #1
+; DOT-NEXT: movi v3.2d, #0000000000000000
+; DOT-NEXT: sdot v3.4s, v1.16b, v2.16b
+; DOT-NEXT: sdot v3.4s, v0.16b, v2.16b
+; DOT-NEXT: addv s0, v3.4s
+; DOT-NEXT: fmov w0, s0
+; DOT-NEXT: ret
+;
+; STREAMING-SVE-LABEL: reduce_saddv_v16i8:
+; STREAMING-SVE: // %bb.0:
+; STREAMING-SVE-NEXT: mov z2.b, #1 // =0x1
+; STREAMING-SVE-NEXT: mov z3.s, #0 // =0x0
+; STREAMING-SVE-NEXT: ptrue p0.s, vl4
+; STREAMING-SVE-NEXT: sdot v3.4s, v1.16b, v2.16b
+; STREAMING-SVE-NEXT: sdot v3.4s, v0.16b, v2.16b
+; STREAMING-SVE-NEXT: uaddv d0, p0, z3.s
+; STREAMING-SVE-NEXT: fmov x0, d0
+; STREAMING-SVE-NEXT: // kill: def $w0 killed $w0 killed $x0
+; STREAMING-SVE-NEXT: ret
+ %1 = sext <32 x i8> %a to <32 x i32>
+ %2 = call i32 @llvm.vector.reduce.add.v32i32(<32 x i32> %1)
+ ret i32 %2
+}
>From 0cc64169015f469e174506c7eb92ee01032c5d11 Mon Sep 17 00:00:00 2001
From: Sander de Smalen <sander.desmalen at arm.com>
Date: Fri, 2 Aug 2024 14:24:47 +0100
Subject: [PATCH 2/2] [AArch64] Avoid NEON dot product in
streaming[-compatible] functions.
A follow-up patch will improve codegen for these functions.
---
.../Target/AArch64/AArch64ISelLowering.cpp | 3 +
...-streaming-mode-fixed-length-reductions.ll | 68 ++++++++++++++++---
2 files changed, 61 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 01079a95b4746..2e869f11b8431 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -17866,6 +17866,9 @@ static SDValue performVecReduceAddCombineWithUADDLP(SDNode *N,
// and generate vecreduce.add(concat_vector(DOT, DOT2, ..)).
static SDValue performVecReduceAddCombine(SDNode *N, SelectionDAG &DAG,
const AArch64Subtarget *ST) {
+ if (!ST->isNeonAvailable())
+ return SDValue();
+
if (!ST->hasDotProd())
return performVecReduceAddCombineWithUADDLP(N, DAG);
diff --git a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
index 7d0918c7023cc..00a15f4bcd639 100644
--- a/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
+++ b/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-reductions.ll
@@ -36,12 +36,36 @@ define i32 @reduce_uaddv_v16i8(<32 x i8> %a) {
;
; STREAMING-SVE-LABEL: reduce_uaddv_v16i8:
; STREAMING-SVE: // %bb.0:
-; STREAMING-SVE-NEXT: mov z2.b, #1 // =0x1
-; STREAMING-SVE-NEXT: mov z3.s, #0 // =0x0
+; STREAMING-SVE-NEXT: // kill: def $q1 killed $q1 def $z1
+; STREAMING-SVE-NEXT: uunpklo z2.h, z1.b
+; STREAMING-SVE-NEXT: // kill: def $q0 killed $q0 def $z0
+; STREAMING-SVE-NEXT: uunpklo z3.h, z0.b
; STREAMING-SVE-NEXT: ptrue p0.s, vl4
-; STREAMING-SVE-NEXT: udot v3.4s, v1.16b, v2.16b
-; STREAMING-SVE-NEXT: udot v3.4s, v0.16b, v2.16b
-; STREAMING-SVE-NEXT: uaddv d0, p0, z3.s
+; STREAMING-SVE-NEXT: ext z1.b, z1.b, z1.b, #8
+; STREAMING-SVE-NEXT: ext z0.b, z0.b, z0.b, #8
+; STREAMING-SVE-NEXT: uunpklo z1.h, z1.b
+; STREAMING-SVE-NEXT: uunpklo z0.h, z0.b
+; STREAMING-SVE-NEXT: uunpklo z4.s, z2.h
+; STREAMING-SVE-NEXT: ext z2.b, z2.b, z2.b, #8
+; STREAMING-SVE-NEXT: uunpklo z6.s, z3.h
+; STREAMING-SVE-NEXT: ext z3.b, z3.b, z3.b, #8
+; STREAMING-SVE-NEXT: mov z5.d, z1.d
+; STREAMING-SVE-NEXT: uunpklo z7.s, z0.h
+; STREAMING-SVE-NEXT: ext z0.b, z0.b, z0.b, #8
+; STREAMING-SVE-NEXT: uunpklo z2.s, z2.h
+; STREAMING-SVE-NEXT: uunpklo z3.s, z3.h
+; STREAMING-SVE-NEXT: add z4.s, z6.s, z4.s
+; STREAMING-SVE-NEXT: ext z5.b, z5.b, z1.b, #8
+; STREAMING-SVE-NEXT: uunpklo z1.s, z1.h
+; STREAMING-SVE-NEXT: uunpklo z0.s, z0.h
+; STREAMING-SVE-NEXT: add z2.s, z3.s, z2.s
+; STREAMING-SVE-NEXT: uunpklo z5.s, z5.h
+; STREAMING-SVE-NEXT: add z1.s, z7.s, z1.s
+; STREAMING-SVE-NEXT: add z0.s, z0.s, z5.s
+; STREAMING-SVE-NEXT: add z1.s, z4.s, z1.s
+; STREAMING-SVE-NEXT: add z0.s, z2.s, z0.s
+; STREAMING-SVE-NEXT: add z0.s, z1.s, z0.s
+; STREAMING-SVE-NEXT: uaddv d0, p0, z0.s
; STREAMING-SVE-NEXT: fmov x0, d0
; STREAMING-SVE-NEXT: // kill: def $w0 killed $w0 killed $x0
; STREAMING-SVE-NEXT: ret
@@ -80,12 +104,36 @@ define i32 @reduce_saddv_v16i8(<32 x i8> %a) {
;
; STREAMING-SVE-LABEL: reduce_saddv_v16i8:
; STREAMING-SVE: // %bb.0:
-; STREAMING-SVE-NEXT: mov z2.b, #1 // =0x1
-; STREAMING-SVE-NEXT: mov z3.s, #0 // =0x0
+; STREAMING-SVE-NEXT: // kill: def $q1 killed $q1 def $z1
+; STREAMING-SVE-NEXT: sunpklo z2.h, z1.b
+; STREAMING-SVE-NEXT: // kill: def $q0 killed $q0 def $z0
+; STREAMING-SVE-NEXT: sunpklo z3.h, z0.b
; STREAMING-SVE-NEXT: ptrue p0.s, vl4
-; STREAMING-SVE-NEXT: sdot v3.4s, v1.16b, v2.16b
-; STREAMING-SVE-NEXT: sdot v3.4s, v0.16b, v2.16b
-; STREAMING-SVE-NEXT: uaddv d0, p0, z3.s
+; STREAMING-SVE-NEXT: ext z1.b, z1.b, z1.b, #8
+; STREAMING-SVE-NEXT: ext z0.b, z0.b, z0.b, #8
+; STREAMING-SVE-NEXT: sunpklo z1.h, z1.b
+; STREAMING-SVE-NEXT: sunpklo z0.h, z0.b
+; STREAMING-SVE-NEXT: sunpklo z4.s, z2.h
+; STREAMING-SVE-NEXT: ext z2.b, z2.b, z2.b, #8
+; STREAMING-SVE-NEXT: sunpklo z6.s, z3.h
+; STREAMING-SVE-NEXT: ext z3.b, z3.b, z3.b, #8
+; STREAMING-SVE-NEXT: mov z5.d, z1.d
+; STREAMING-SVE-NEXT: sunpklo z7.s, z0.h
+; STREAMING-SVE-NEXT: ext z0.b, z0.b, z0.b, #8
+; STREAMING-SVE-NEXT: sunpklo z2.s, z2.h
+; STREAMING-SVE-NEXT: sunpklo z3.s, z3.h
+; STREAMING-SVE-NEXT: add z4.s, z6.s, z4.s
+; STREAMING-SVE-NEXT: ext z5.b, z5.b, z1.b, #8
+; STREAMING-SVE-NEXT: sunpklo z1.s, z1.h
+; STREAMING-SVE-NEXT: sunpklo z0.s, z0.h
+; STREAMING-SVE-NEXT: add z2.s, z3.s, z2.s
+; STREAMING-SVE-NEXT: sunpklo z5.s, z5.h
+; STREAMING-SVE-NEXT: add z1.s, z7.s, z1.s
+; STREAMING-SVE-NEXT: add z0.s, z0.s, z5.s
+; STREAMING-SVE-NEXT: add z1.s, z4.s, z1.s
+; STREAMING-SVE-NEXT: add z0.s, z2.s, z0.s
+; STREAMING-SVE-NEXT: add z0.s, z1.s, z0.s
+; STREAMING-SVE-NEXT: uaddv d0, p0, z0.s
; STREAMING-SVE-NEXT: fmov x0, d0
; STREAMING-SVE-NEXT: // kill: def $w0 killed $w0 killed $x0
; STREAMING-SVE-NEXT: ret
More information about the llvm-commits
mailing list