[RISCV] Disable combineToVCPOP for illegal scalable vector types. (#140195)
This transform creates target specific instructions which must have
legal types. We were checking this for fixed vectors, but not scalable
vectors. This caused a crash with <vscale x 1 x i1> which isn't legal
for Zve32x.
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 851b0e7..5e761fc 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -19031,6 +19031,10 @@
if (!SrcMVT.isVector() || SrcMVT.getVectorElementType() != MVT::i1)
return SDValue();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!TLI.isTypeLegal(SrcMVT))
+ return SDValue();
+
// Check that destination type is large enough to hold result without
// overflow.
if (Opc == ISD::VECREDUCE_ADD) {
@@ -19047,9 +19051,6 @@
MVT ContainerVT = SrcMVT;
if (SrcMVT.isFixedLengthVector()) {
- if (!useRVVForFixedLengthVectorVT(SrcMVT, Subtarget))
- return SDValue();
-
ContainerVT = getContainerForFixedLengthVector(DAG, SrcMVT, Subtarget);
Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
}
diff --git a/llvm/test/CodeGen/RISCV/rvv/combine-reduce-add-to-vcpop.ll b/llvm/test/CodeGen/RISCV/rvv/combine-reduce-add-to-vcpop.ll
index ac1d633..88894f8 100644
--- a/llvm/test/CodeGen/RISCV/rvv/combine-reduce-add-to-vcpop.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/combine-reduce-add-to-vcpop.ll
@@ -1,13 +1,20 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc < %s -mtriple=riscv32 -mattr=+v,+zbb | FileCheck %s --check-prefixes=CHECK,RV32
-; RUN: llc < %s -mtriple=riscv64 -mattr=+v,+zbb | FileCheck %s --check-prefixes=CHECK,RV64
+; RUN: llc < %s -mtriple=riscv32 -mattr=+v,+zbb | FileCheck %s --check-prefixes=CHECK,V
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v,+zbb | FileCheck %s --check-prefixes=CHECK,V
+; RUN: llc < %s -mtriple=riscv64 -mattr=+zve32x,+zvl128b,+zbb | FileCheck %s --check-prefixes=CHECK,ZVE
define i32 @test_v2i1(<2 x i1> %x) {
-; CHECK-LABEL: test_v2i1:
-; CHECK: # %bb.0:
-; CHECK-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
-; CHECK-NEXT: vcpop.m a0, v0
-; CHECK-NEXT: ret
+; V-LABEL: test_v2i1:
+; V: # %bb.0:
+; V-NEXT: vsetivli zero, 2, e8, mf8, ta, ma
+; V-NEXT: vcpop.m a0, v0
+; V-NEXT: ret
+;
+; ZVE-LABEL: test_v2i1:
+; ZVE: # %bb.0:
+; ZVE-NEXT: vsetivli zero, 2, e8, mf4, ta, ma
+; ZVE-NEXT: vcpop.m a0, v0
+; ZVE-NEXT: ret
%a = zext <2 x i1> %x to <2 x i32>
%b = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> %a)
ret i32 %b
@@ -173,6 +180,35 @@
ret i32 %b
}
+; FIXME: Optimize this case with Zve32x. We have to use mf4 and set the VL to
+; VLEN/64.
+define i32 @test_nxv1i1(<vscale x 1 x i1> %x) {
+; V-LABEL: test_nxv1i1:
+; V: # %bb.0: # %entry
+; V-NEXT: vsetvli a0, zero, e8, mf8, ta, ma
+; V-NEXT: vcpop.m a0, v0
+; V-NEXT: ret
+;
+; ZVE-LABEL: test_nxv1i1:
+; ZVE: # %bb.0: # %entry
+; ZVE-NEXT: vsetvli a0, zero, e32, m1, ta, ma
+; ZVE-NEXT: vmv.v.i v8, 0
+; ZVE-NEXT: csrr a0, vlenb
+; ZVE-NEXT: srli a0, a0, 3
+; ZVE-NEXT: vsetvli zero, a0, e32, m1, ta, ma
+; ZVE-NEXT: vmerge.vim v8, v8, 1, v0
+; ZVE-NEXT: vsetivli zero, 1, e32, m1, ta, ma
+; ZVE-NEXT: vmv.s.x v9, zero
+; ZVE-NEXT: vsetvli zero, a0, e32, m1, ta, ma
+; ZVE-NEXT: vredsum.vs v9, v8, v9
+; ZVE-NEXT: vmv.x.s a0, v9
+; ZVE-NEXT: ret
+entry:
+ %a = zext <vscale x 1 x i1> %x to <vscale x 1 x i32>
+ %b = call i32 @llvm.vector.reduce.add.nxv1i32(<vscale x 1 x i32> %a)
+ ret i32 %b
+}
+
define i32 @test_nxv2i1(<vscale x 2 x i1> %x) {
; CHECK-LABEL: test_nxv2i1:
; CHECK: # %bb.0: # %entry
@@ -520,7 +556,3 @@
%b = call i16 @llvm.vector.reduce.add.nxv64i16(<vscale x 64 x i16> %a)
ret i16 %b
}
-
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; RV32: {{.*}}
-; RV64: {{.*}}
