[llvm] [RISCV][GISEL] Legalize G_EXTRACT_SUBVECTOR (PR #109426)
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 20 11:07:47 PDT 2024
================
@@ -0,0 +1,339 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=riscv32 -mattr=+v -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=CHECK,RV32
+# RUN: llc -mtriple=riscv64 -mattr=+v -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=CHECK,RV64
+
+# Special handling for i1-element vectors with non-zero index
+---
+name: extract_subvector_nxv4i1
+legalized: false
+tracksRegLiveness: true
+body: |
+ bb.0.entry:
+ ; RV32-LABEL: name: extract_subvector_nxv4i1
+ ; RV32: [[DEF:%[0-9]+]]:_(<vscale x 4 x s1>) = G_IMPLICIT_DEF
+ ; RV32-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+ ; RV32-NEXT: [[SPLAT_VECTOR:%[0-9]+]]:_(<vscale x 4 x s8>) = G_SPLAT_VECTOR [[C]](s32)
+ ; RV32-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+ ; RV32-NEXT: [[SPLAT_VECTOR1:%[0-9]+]]:_(<vscale x 4 x s8>) = G_SPLAT_VECTOR [[C1]](s32)
+ ; RV32-NEXT: [[SELECT:%[0-9]+]]:_(<vscale x 4 x s8>) = G_SELECT [[DEF]](<vscale x 4 x s1>), [[SPLAT_VECTOR1]], [[SPLAT_VECTOR]]
+ ; RV32-NEXT: [[READ_VLENB:%[0-9]+]]:_(s32) = G_READ_VLENB
+ ; RV32-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
+ ; RV32-NEXT: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[READ_VLENB]], [[C2]](s32)
+ ; RV32-NEXT: [[VMSET_VL:%[0-9]+]]:_(<vscale x 2 x s1>) = G_VMSET_VL $x0
+ ; RV32-NEXT: [[DEF1:%[0-9]+]]:_(<vscale x 4 x s8>) = G_IMPLICIT_DEF
+ ; RV32-NEXT: [[VSLIDEDOWN_VL:%[0-9]+]]:_(<vscale x 4 x s8>) = G_VSLIDEDOWN_VL [[DEF1]], [[SELECT]], [[LSHR]](s32), [[VMSET_VL]](<vscale x 2 x s1>), $x0, 3
+ ; RV32-NEXT: [[EXTRACT_SUBVECTOR:%[0-9]+]]:_(<vscale x 2 x s8>) = G_EXTRACT_SUBVECTOR [[VSLIDEDOWN_VL]](<vscale x 4 x s8>), 0
----------------
topperc wrote:
AArch64 has no "legal" G_EXTRACT_VECTOR_ELT, but the custom handler keeps any G_EXTRACT_VECTOR_ELT with a constant index.
```
getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
.unsupportedIf([=](const LegalityQuery &Query) {
const LLT &EltTy = Query.Types[1].getElementType();
return Query.Types[0] != EltTy;
})
.minScalar(2, s64)
.customIf([=](const LegalityQuery &Query) {
const LLT &VecTy = Query.Types[1];
return VecTy == v2s16 || VecTy == v4s16 || VecTy == v8s16 ||
VecTy == v4s32 || VecTy == v2s64 || VecTy == v2s32 ||
VecTy == v8s8 || VecTy == v16s8 || VecTy == v2p0;
})
.minScalarOrEltIf(
[=](const LegalityQuery &Query) {
// We want to promote to <M x s1> to <M x s64> if that wouldn't
// cause the total vec size to be > 128b.
return Query.Types[1].getNumElements() <= 2;
},
0, s64)
.minScalarOrEltIf(
[=](const LegalityQuery &Query) {
return Query.Types[1].getNumElements() <= 4;
},
0, s32)
.minScalarOrEltIf(
[=](const LegalityQuery &Query) {
return Query.Types[1].getNumElements() <= 8;
},
0, s16)
.minScalarOrEltIf(
[=](const LegalityQuery &Query) {
return Query.Types[1].getNumElements() <= 16;
},
0, s8)
.minScalarOrElt(0, s8) // Worst case, we need at least s8.
.moreElementsToNextPow2(1)
.clampMaxNumElements(1, s64, 2)
.clampMaxNumElements(1, s32, 4)
.clampMaxNumElements(1, s16, 8)
.clampMaxNumElements(1, s8, 16)
.clampMaxNumElements(1, p0, 2);
```
https://github.com/llvm/llvm-project/pull/109426
More information about the llvm-commits
mailing list