[llvm] [AArch64] Add sanitazation when custom lowering fp partial reduce (PR #184811)

via llvm-commits llvm-commits at lists.llvm.org
Thu Mar 5 07:42:06 PST 2026


https://github.com/Lukacma created https://github.com/llvm/llvm-project/pull/184811

This patch adds sanitization of input vectors when converting from using fixed vectors into using scalable vectors for fp partial reduce. This is done to preserve behaviour correctness and avoid triggering fp exceptions on extra sve lanes. The performance with this sanitazation is still better than trying to do it in neon using fmul + fadd, at least in simple cases. 

>From f9c246936fc12681234314758860e42c06027af6 Mon Sep 17 00:00:00 2001
From: Marian Lukac <Marian.Lukac at arm.com>
Date: Thu, 5 Mar 2026 15:28:32 +0000
Subject: [PATCH 1/2] [AArch64] Add sanitazation when custom lowering fp
 partial reduce

---
 .../Target/AArch64/AArch64ISelLowering.cpp    |  18 ++-
 llvm/test/CodeGen/AArch64/clmul-fixed.ll      |  20 ++--
 llvm/test/CodeGen/AArch64/clmul.ll            |  18 +--
 .../partial-reduction-add-predicated.ll       |  16 ++-
 .../AArch64/sve-fixed-length-frame-offests.ll |   8 +-
 .../AArch64/sve2p1-fixed-length-fdot.ll       | 107 +++++++++++++-----
 6 files changed, 130 insertions(+), 57 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 0201d45010195..e713453fc9792 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -3591,7 +3591,8 @@ MachineBasicBlock *AArch64TargetLowering::EmitInstrWithCustomInserter(
 
 // Forward declarations of SVE fixed length lowering helpers
 static EVT getContainerForFixedLengthVector(SelectionDAG &DAG, EVT VT);
-static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V);
+static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V,
+                                       bool clearUnused = false);
 static SDValue convertFromScalableVector(SelectionDAG &DAG, EVT VT, SDValue V);
 static SDValue convertFixedMaskToScalableVector(SDValue Mask,
                                                 SelectionDAG &DAG);
@@ -31416,14 +31417,18 @@ static SDValue getPredicateForVector(SelectionDAG &DAG, SDLoc &DL, EVT VT) {
 }
 
 // Grow V to consume an entire SVE register.
-static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V) {
+static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V,
+                                       bool clearUnused) {
   assert(VT.isScalableVector() &&
          "Expected to convert into a scalable vector!");
   assert(V.getValueType().isFixedLengthVector() &&
          "Expected a fixed length vector operand!");
   SDLoc DL(V);
   SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
-  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getPOISON(VT), V, Zero);
+  SDValue ZeroVec = VT.isFloatingPoint() ? DAG.getConstantFP(0, DL, VT)
+                                         : DAG.getConstant(0, DL, VT);
+  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
+                     clearUnused ? ZeroVec : DAG.getPOISON(VT), V, Zero);
 }
 
 // Shrink V so it's just big enough to maintain a VT's worth of data.
@@ -32590,9 +32595,10 @@ AArch64TargetLowering::LowerPARTIAL_REDUCE_MLA(SDValue Op,
   if (ConvertToScalable) {
     ResultVT = getContainerForFixedLengthVector(DAG, ResultVT);
     OpVT = getContainerForFixedLengthVector(DAG, LHS.getValueType());
-    Acc = convertToScalableVector(DAG, ResultVT, Acc);
-    LHS = convertToScalableVector(DAG, OpVT, LHS);
-    RHS = convertToScalableVector(DAG, OpVT, RHS);
+    Acc =
+        convertToScalableVector(DAG, ResultVT, Acc, ResultVT.isFloatingPoint());
+    LHS = convertToScalableVector(DAG, OpVT, LHS, OpVT.isFloatingPoint());
+    RHS = convertToScalableVector(DAG, OpVT, RHS, OpVT.isFloatingPoint());
     Op = DAG.getNode(Op.getOpcode(), DL, ResultVT, {Acc, LHS, RHS});
   }
 
diff --git a/llvm/test/CodeGen/AArch64/clmul-fixed.ll b/llvm/test/CodeGen/AArch64/clmul-fixed.ll
index 61081c3342889..6dbc0b4a70f37 100644
--- a/llvm/test/CodeGen/AArch64/clmul-fixed.ll
+++ b/llvm/test/CodeGen/AArch64/clmul-fixed.ll
@@ -1728,21 +1728,21 @@ define <1 x i128> @clmul_v1i128_neon(<1 x i128> %x, <1 x i128> %y) {
 ; CHECK-AES:       // %bb.0:
 ; CHECK-AES-NEXT:    rbit x8, x2
 ; CHECK-AES-NEXT:    rbit x9, x0
-; CHECK-AES-NEXT:    fmov d0, x2
-; CHECK-AES-NEXT:    fmov d1, x1
-; CHECK-AES-NEXT:    fmov d2, x3
+; CHECK-AES-NEXT:    fmov d0, x3
+; CHECK-AES-NEXT:    fmov d1, x0
+; CHECK-AES-NEXT:    fmov d2, x2
 ; CHECK-AES-NEXT:    fmov d3, x8
 ; CHECK-AES-NEXT:    fmov d4, x9
-; CHECK-AES-NEXT:    pmull v1.1q, v1.1d, v0.1d
+; CHECK-AES-NEXT:    pmull v0.1q, v1.1d, v0.1d
 ; CHECK-AES-NEXT:    pmull v3.1q, v4.1d, v3.1d
-; CHECK-AES-NEXT:    fmov d4, x0
-; CHECK-AES-NEXT:    pmull v2.1q, v4.1d, v2.1d
-; CHECK-AES-NEXT:    fmov x9, d1
+; CHECK-AES-NEXT:    fmov d4, x1
+; CHECK-AES-NEXT:    pmull v1.1q, v1.1d, v2.1d
+; CHECK-AES-NEXT:    pmull v4.1q, v4.1d, v2.1d
+; CHECK-AES-NEXT:    fmov x10, d0
 ; CHECK-AES-NEXT:    fmov x8, d3
-; CHECK-AES-NEXT:    pmull v0.1q, v4.1d, v0.1d
-; CHECK-AES-NEXT:    fmov x10, d2
+; CHECK-AES-NEXT:    fmov x0, d1
+; CHECK-AES-NEXT:    fmov x9, d4
 ; CHECK-AES-NEXT:    rbit x8, x8
-; CHECK-AES-NEXT:    fmov x0, d0
 ; CHECK-AES-NEXT:    eor x9, x10, x9
 ; CHECK-AES-NEXT:    eor x1, x9, x8, lsr #1
 ; CHECK-AES-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/clmul.ll b/llvm/test/CodeGen/AArch64/clmul.ll
index 7a94696711882..ee182100fdfd8 100644
--- a/llvm/test/CodeGen/AArch64/clmul.ll
+++ b/llvm/test/CodeGen/AArch64/clmul.ll
@@ -179,9 +179,9 @@ define i32 @clmul_i32(i32 %x, i32 %y) {
 ;
 ; CHECK-AES-LABEL: clmul_i32:
 ; CHECK-AES:       // %bb.0:
-; CHECK-AES-NEXT:    fmov s0, w1
-; CHECK-AES-NEXT:    fmov s1, w0
-; CHECK-AES-NEXT:    pmull v0.1q, v1.1d, v0.1d
+; CHECK-AES-NEXT:    fmov s0, w0
+; CHECK-AES-NEXT:    fmov s1, w1
+; CHECK-AES-NEXT:    pmull v0.1q, v0.1d, v1.1d
 ; CHECK-AES-NEXT:    fmov w0, s0
 ; CHECK-AES-NEXT:    ret
   %a = call i32 @llvm.clmul.i32(i32 %x, i32 %y)
@@ -573,9 +573,9 @@ define i32 @clmul_i32_zext(i16 %x, i16 %y) {
 ; CHECK-AES:       // %bb.0:
 ; CHECK-AES-NEXT:    and w8, w0, #0xffff
 ; CHECK-AES-NEXT:    and w9, w1, #0xffff
-; CHECK-AES-NEXT:    fmov s0, w9
-; CHECK-AES-NEXT:    fmov s1, w8
-; CHECK-AES-NEXT:    pmull v0.1q, v1.1d, v0.1d
+; CHECK-AES-NEXT:    fmov s0, w8
+; CHECK-AES-NEXT:    fmov s1, w9
+; CHECK-AES-NEXT:    pmull v0.1q, v0.1d, v1.1d
 ; CHECK-AES-NEXT:    fmov w0, s0
 ; CHECK-AES-NEXT:    ret
   %zextx = zext i16 %x to i32
@@ -689,9 +689,9 @@ define i64 @clmul_i64_zext(i32 %x, i32 %y) {
 ; CHECK-AES:       // %bb.0:
 ; CHECK-AES-NEXT:    mov w8, w0
 ; CHECK-AES-NEXT:    mov w9, w1
-; CHECK-AES-NEXT:    fmov d0, x9
-; CHECK-AES-NEXT:    fmov d1, x8
-; CHECK-AES-NEXT:    pmull v0.1q, v1.1d, v0.1d
+; CHECK-AES-NEXT:    fmov d0, x8
+; CHECK-AES-NEXT:    fmov d1, x9
+; CHECK-AES-NEXT:    pmull v0.1q, v0.1d, v1.1d
 ; CHECK-AES-NEXT:    fmov x0, d0
 ; CHECK-AES-NEXT:    ret
   %zextx = zext i32 %x to i64
diff --git a/llvm/test/CodeGen/AArch64/partial-reduction-add-predicated.ll b/llvm/test/CodeGen/AArch64/partial-reduction-add-predicated.ll
index 24cdd0a852222..de6ecba1d878f 100644
--- a/llvm/test/CodeGen/AArch64/partial-reduction-add-predicated.ll
+++ b/llvm/test/CodeGen/AArch64/partial-reduction-add-predicated.ll
@@ -92,11 +92,17 @@ define <4 x float> @predicated_fdot_fixed_length(<4 x float> %acc, <8 x i1> %p,
 ; CHECK-LABEL: predicated_fdot_fixed_length:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ushll v1.8h, v1.8b, #0
-; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-NEXT:    movi v4.2d, #0000000000000000
 ; CHECK-NEXT:    // kill: def $q2 killed $q2 def $z2
+; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-NEXT:    ptrue p0.h, vl8
+; CHECK-NEXT:    ptrue p1.s, vl4
 ; CHECK-NEXT:    shl v1.8h, v1.8h, #15
+; CHECK-NEXT:    sel z2.h, p0, z2.h, z4.h
+; CHECK-NEXT:    sel z0.s, p1, z0.s, z4.s
 ; CHECK-NEXT:    cmlt v1.8h, v1.8h, #0
 ; CHECK-NEXT:    and v1.16b, v1.16b, v3.16b
+; CHECK-NEXT:    sel z1.h, p0, z1.h, z4.h
 ; CHECK-NEXT:    fdot z0.s, z2.h, z1.h
 ; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
 ; CHECK-NEXT:    ret
@@ -128,11 +134,17 @@ define <4 x float> @predicated_fpext_fmul_fixed_length(<4 x float> %acc, <8 x i1
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    ushll v1.8h, v1.8b, #0
 ; CHECK-NEXT:    movi v3.8h, #60, lsl #8
-; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
 ; CHECK-NEXT:    // kill: def $q2 killed $q2 def $z2
+; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-NEXT:    movi v4.2d, #0000000000000000
+; CHECK-NEXT:    ptrue p0.h, vl8
+; CHECK-NEXT:    ptrue p1.s, vl4
 ; CHECK-NEXT:    shl v1.8h, v1.8h, #15
+; CHECK-NEXT:    sel z2.h, p0, z2.h, z4.h
+; CHECK-NEXT:    sel z0.s, p1, z0.s, z4.s
 ; CHECK-NEXT:    cmlt v1.8h, v1.8h, #0
 ; CHECK-NEXT:    and v1.16b, v1.16b, v3.16b
+; CHECK-NEXT:    sel z1.h, p0, z1.h, z4.h
 ; CHECK-NEXT:    fdot z0.s, z2.h, z1.h
 ; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
 ; CHECK-NEXT:    ret
diff --git a/llvm/test/CodeGen/AArch64/sve-fixed-length-frame-offests.ll b/llvm/test/CodeGen/AArch64/sve-fixed-length-frame-offests.ll
index 5e940078695f0..8fe52fa0dcb10 100644
--- a/llvm/test/CodeGen/AArch64/sve-fixed-length-frame-offests.ll
+++ b/llvm/test/CodeGen/AArch64/sve-fixed-length-frame-offests.ll
@@ -12,12 +12,12 @@ define void @foo(ptr %a) #0 {
 ; CHECK:       SelectionDAG has 13 nodes:
 ; CHECK-NEXT:    t0: ch,glue = EntryToken
 ; CHECK-NEXT:    t2: i64,ch = CopyFromReg t0, Register:i64 %0
-; CHECK-NEXT:    t22: nxv2i64,ch = LDR_ZXI<Mem:(volatile load (<vscale x 1 x s128>) from %ir.a, align 64)> t2, TargetConstant:i64<0>, t0
+; CHECK-NEXT:    t23: nxv2i64,ch = LDR_ZXI<Mem:(volatile load (<vscale x 1 x s128>) from %ir.a, align 64)> t2, TargetConstant:i64<0>, t0
 ; CHECK-NEXT:    t8: i64 = ADDXri TargetFrameIndex:i64<1>, TargetConstant:i32<0>, TargetConstant:i32<0>
 ; CHECK-NEXT:    t6: i64 = ADDXri TargetFrameIndex:i64<0>, TargetConstant:i32<0>, TargetConstant:i32<0>
-; CHECK-NEXT:    t23: ch = STR_ZXI<Mem:(volatile store (<vscale x 1 x s128>) into %ir.r0, align 64)> t22, t6, TargetConstant:i64<0>, t22:1
-; CHECK-NEXT:    t24: ch = STR_ZXI<Mem:(volatile store (<vscale x 1 x s128>) into %ir.r1, align 64)> t22, t8, TargetConstant:i64<0>, t23
-; CHECK-NEXT:    t10: ch = RET_ReallyLR t24
+; CHECK-NEXT:    t24: ch = STR_ZXI<Mem:(volatile store (<vscale x 1 x s128>) into %ir.r0, align 64)> t23, t6, TargetConstant:i64<0>, t23:1
+; CHECK-NEXT:    t25: ch = STR_ZXI<Mem:(volatile store (<vscale x 1 x s128>) into %ir.r1, align 64)> t23, t8, TargetConstant:i64<0>, t24
+; CHECK-NEXT:    t10: ch = RET_ReallyLR t25
 ; CHECK-EMPTY:
 entry:
   %r0 = alloca <8 x i64>
diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
index 864c66caf5f6c..98dfd7aa1b316 100644
--- a/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
+++ b/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
@@ -23,11 +23,17 @@ define void @fdot_v4f32(ptr %accptr, ptr %aptr, ptr %bptr) {
 ;
 ; SVE2P1-LABEL: fdot_v4f32:
 ; SVE2P1:       // %bb.0: // %entry
-; SVE2P1-NEXT:    ldr q0, [x0]
-; SVE2P1-NEXT:    ldr q1, [x1]
-; SVE2P1-NEXT:    ldr q2, [x2]
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    str q0, [x0]
+; SVE2P1-NEXT:    movi v0.2d, #0000000000000000
+; SVE2P1-NEXT:    ldr q1, [x0]
+; SVE2P1-NEXT:    ptrue p0.s, vl4
+; SVE2P1-NEXT:    ldr q2, [x1]
+; SVE2P1-NEXT:    ptrue p1.h, vl8
+; SVE2P1-NEXT:    ldr q3, [x2]
+; SVE2P1-NEXT:    sel z1.s, p0, z1.s, z0.s
+; SVE2P1-NEXT:    sel z2.h, p1, z2.h, z0.h
+; SVE2P1-NEXT:    mov z0.h, p1/m, z3.h
+; SVE2P1-NEXT:    fdot z1.s, z2.h, z0.h
+; SVE2P1-NEXT:    str q1, [x0]
 ; SVE2P1-NEXT:    ret
 entry:
   %acc = load <4 x float>, ptr %accptr
@@ -65,12 +71,16 @@ define void @fdot_wide_v8f32(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(2,0
 ; SVE2P1-LABEL: fdot_wide_v8f32:
 ; SVE2P1:       // %bb.0: // %entry
 ; SVE2P1-NEXT:    ptrue p0.s, vl8
+; SVE2P1-NEXT:    movi v0.2d, #0000000000000000
 ; SVE2P1-NEXT:    ptrue p1.h, vl16
-; SVE2P1-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; SVE2P1-NEXT:    ld1h { z1.h }, p1/z, [x1]
-; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x2]
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    st1w { z0.s }, p0, [x0]
+; SVE2P1-NEXT:    ld1w { z1.s }, p0/z, [x0]
+; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x1]
+; SVE2P1-NEXT:    ld1h { z3.h }, p1/z, [x2]
+; SVE2P1-NEXT:    sel z1.s, p0, z1.s, z0.s
+; SVE2P1-NEXT:    sel z2.h, p1, z2.h, z0.h
+; SVE2P1-NEXT:    mov z0.h, p1/m, z3.h
+; SVE2P1-NEXT:    fdot z1.s, z2.h, z0.h
+; SVE2P1-NEXT:    st1w { z1.s }, p0, [x0]
 ; SVE2P1-NEXT:    ret
 entry:
   %acc = load <8 x float>, ptr %accptr
@@ -108,12 +118,16 @@ define void @fdot_wide_v16f32(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(4,
 ; SVE2P1-LABEL: fdot_wide_v16f32:
 ; SVE2P1:       // %bb.0: // %entry
 ; SVE2P1-NEXT:    ptrue p0.s, vl16
+; SVE2P1-NEXT:    movi v0.2d, #0000000000000000
 ; SVE2P1-NEXT:    ptrue p1.h, vl32
-; SVE2P1-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; SVE2P1-NEXT:    ld1h { z1.h }, p1/z, [x1]
-; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x2]
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    st1w { z0.s }, p0, [x0]
+; SVE2P1-NEXT:    ld1w { z1.s }, p0/z, [x0]
+; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x1]
+; SVE2P1-NEXT:    ld1h { z3.h }, p1/z, [x2]
+; SVE2P1-NEXT:    sel z1.s, p0, z1.s, z0.s
+; SVE2P1-NEXT:    sel z2.h, p1, z2.h, z0.h
+; SVE2P1-NEXT:    mov z0.h, p1/m, z3.h
+; SVE2P1-NEXT:    fdot z1.s, z2.h, z0.h
+; SVE2P1-NEXT:    st1w { z1.s }, p0, [x0]
 ; SVE2P1-NEXT:    ret
 entry:
   %acc = load <16 x float>, ptr %accptr
@@ -151,12 +165,16 @@ define void @fdot_wide_v32f32(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(8,
 ; SVE2P1-LABEL: fdot_wide_v32f32:
 ; SVE2P1:       // %bb.0: // %entry
 ; SVE2P1-NEXT:    ptrue p0.s, vl32
+; SVE2P1-NEXT:    movi v0.2d, #0000000000000000
 ; SVE2P1-NEXT:    ptrue p1.h, vl64
-; SVE2P1-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; SVE2P1-NEXT:    ld1h { z1.h }, p1/z, [x1]
-; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x2]
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    st1w { z0.s }, p0, [x0]
+; SVE2P1-NEXT:    ld1w { z1.s }, p0/z, [x0]
+; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x1]
+; SVE2P1-NEXT:    ld1h { z3.h }, p1/z, [x2]
+; SVE2P1-NEXT:    sel z1.s, p0, z1.s, z0.s
+; SVE2P1-NEXT:    sel z2.h, p1, z2.h, z0.h
+; SVE2P1-NEXT:    mov z0.h, p1/m, z3.h
+; SVE2P1-NEXT:    fdot z1.s, z2.h, z0.h
+; SVE2P1-NEXT:    st1w { z1.s }, p0, [x0]
 ; SVE2P1-NEXT:    ret
 entry:
   %acc = load <32 x float>, ptr %accptr
@@ -194,12 +212,16 @@ define void @fdot_wide_v64f32(ptr %accptr, ptr %aptr, ptr %bptr) vscale_range(16
 ; SVE2P1-LABEL: fdot_wide_v64f32:
 ; SVE2P1:       // %bb.0: // %entry
 ; SVE2P1-NEXT:    ptrue p0.s, vl64
+; SVE2P1-NEXT:    movi v0.2d, #0000000000000000
 ; SVE2P1-NEXT:    ptrue p1.h, vl128
-; SVE2P1-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; SVE2P1-NEXT:    ld1h { z1.h }, p1/z, [x1]
-; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x2]
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    st1w { z0.s }, p0, [x0]
+; SVE2P1-NEXT:    ld1w { z1.s }, p0/z, [x0]
+; SVE2P1-NEXT:    ld1h { z2.h }, p1/z, [x1]
+; SVE2P1-NEXT:    ld1h { z3.h }, p1/z, [x2]
+; SVE2P1-NEXT:    sel z1.s, p0, z1.s, z0.s
+; SVE2P1-NEXT:    sel z2.h, p1, z2.h, z0.h
+; SVE2P1-NEXT:    mov z0.h, p1/m, z3.h
+; SVE2P1-NEXT:    fdot z1.s, z2.h, z0.h
+; SVE2P1-NEXT:    st1w { z1.s }, p0, [x0]
 ; SVE2P1-NEXT:    ret
 entry:
   %acc = load <64 x float>, ptr %accptr
@@ -228,9 +250,15 @@ define <4 x float> @fixed_fdot_wide(<4 x float> %acc, <8 x half> %a, <8 x half>
 ;
 ; SVE2P1-LABEL: fixed_fdot_wide:
 ; SVE2P1:       // %bb.0: // %entry
-; SVE2P1-NEXT:    // kill: def $q0 killed $q0 def $z0
+; SVE2P1-NEXT:    movi v3.2d, #0000000000000000
+; SVE2P1-NEXT:    ptrue p0.h, vl8
 ; SVE2P1-NEXT:    // kill: def $q2 killed $q2 def $z2
 ; SVE2P1-NEXT:    // kill: def $q1 killed $q1 def $z1
+; SVE2P1-NEXT:    // kill: def $q0 killed $q0 def $z0
+; SVE2P1-NEXT:    ptrue p1.s, vl4
+; SVE2P1-NEXT:    sel z2.h, p0, z2.h, z3.h
+; SVE2P1-NEXT:    sel z1.h, p0, z1.h, z3.h
+; SVE2P1-NEXT:    sel z0.s, p1, z0.s, z3.s
 ; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
 ; SVE2P1-NEXT:    // kill: def $q0 killed $q0 killed $z0
 ; SVE2P1-NEXT:    ret
@@ -242,6 +270,33 @@ entry:
   ret <4 x float> %partial.reduce
 }
 
+define <2 x float> @fixed_fdot(<2 x float> %acc, <4 x half> %a, <4 x half> %b) {
+; SVE2-LABEL: fixed_fdot:
+; SVE2:       // %bb.0: // %entry
+; SVE2-NEXT:    fcvtl v1.4s, v1.4h
+; SVE2-NEXT:    fcvtl v2.4s, v2.4h
+; SVE2-NEXT:    fmul v1.4s, v1.4s, v2.4s
+; SVE2-NEXT:    fadd v0.2s, v0.2s, v1.2s
+; SVE2-NEXT:    ext v1.16b, v1.16b, v1.16b, #8
+; SVE2-NEXT:    fadd v0.2s, v1.2s, v0.2s
+; SVE2-NEXT:    ret
+;
+; SVE2P1-LABEL: fixed_fdot:
+; SVE2P1:       // %bb.0: // %entry
+; SVE2P1-NEXT:    // kill: def $d0 killed $d0 def $z0
+; SVE2P1-NEXT:    // kill: def $d2 killed $d2 def $z2
+; SVE2P1-NEXT:    // kill: def $d1 killed $d1 def $z1
+; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
+; SVE2P1-NEXT:    // kill: def $d0 killed $d0 killed $z0
+; SVE2P1-NEXT:    ret
+entry:
+  %a.wide = fpext <4 x half> %a to <4 x float>
+  %b.wide = fpext <4 x half> %b to <4 x float>
+  %mult = fmul <4 x float> %a.wide, %b.wide
+  %partial.reduce = call <2 x float> @llvm.vector.partial.reduce.fadd(<2 x float> %acc, <4 x float> %mult)
+  ret <2 x float> %partial.reduce
+}
+
 define <8 x half> @partial_reduce_half(<8 x half> %acc, <16 x half> %a) {
 ; CHECK-LABEL: partial_reduce_half:
 ; CHECK:       // %bb.0: // %entry

>From ddddebd4b2583af23b43b5a78d6ddc81316edb43 Mon Sep 17 00:00:00 2001
From: Marian Lukac <Marian.Lukac at arm.com>
Date: Thu, 5 Mar 2026 15:34:13 +0000
Subject: [PATCH 2/2] remove wrong test

---
 .../AArch64/sve2p1-fixed-length-fdot.ll       | 27 -------------------
 1 file changed, 27 deletions(-)

diff --git a/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll b/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
index 98dfd7aa1b316..1db2494753779 100644
--- a/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
+++ b/llvm/test/CodeGen/AArch64/sve2p1-fixed-length-fdot.ll
@@ -270,33 +270,6 @@ entry:
   ret <4 x float> %partial.reduce
 }
 
-define <2 x float> @fixed_fdot(<2 x float> %acc, <4 x half> %a, <4 x half> %b) {
-; SVE2-LABEL: fixed_fdot:
-; SVE2:       // %bb.0: // %entry
-; SVE2-NEXT:    fcvtl v1.4s, v1.4h
-; SVE2-NEXT:    fcvtl v2.4s, v2.4h
-; SVE2-NEXT:    fmul v1.4s, v1.4s, v2.4s
-; SVE2-NEXT:    fadd v0.2s, v0.2s, v1.2s
-; SVE2-NEXT:    ext v1.16b, v1.16b, v1.16b, #8
-; SVE2-NEXT:    fadd v0.2s, v1.2s, v0.2s
-; SVE2-NEXT:    ret
-;
-; SVE2P1-LABEL: fixed_fdot:
-; SVE2P1:       // %bb.0: // %entry
-; SVE2P1-NEXT:    // kill: def $d0 killed $d0 def $z0
-; SVE2P1-NEXT:    // kill: def $d2 killed $d2 def $z2
-; SVE2P1-NEXT:    // kill: def $d1 killed $d1 def $z1
-; SVE2P1-NEXT:    fdot z0.s, z1.h, z2.h
-; SVE2P1-NEXT:    // kill: def $d0 killed $d0 killed $z0
-; SVE2P1-NEXT:    ret
-entry:
-  %a.wide = fpext <4 x half> %a to <4 x float>
-  %b.wide = fpext <4 x half> %b to <4 x float>
-  %mult = fmul <4 x float> %a.wide, %b.wide
-  %partial.reduce = call <2 x float> @llvm.vector.partial.reduce.fadd(<2 x float> %acc, <4 x float> %mult)
-  ret <2 x float> %partial.reduce
-}
-
 define <8 x half> @partial_reduce_half(<8 x half> %acc, <16 x half> %a) {
 ; CHECK-LABEL: partial_reduce_half:
 ; CHECK:       // %bb.0: // %entry



More information about the llvm-commits mailing list