[RISCV] Disable fixed vectors in getOptimalMemOpType if minimum VLEN is 32. #102974
Conversation
…is 32. This is needed to match llvm#102405 which disabled fixed to scalable vector lowering for VLEN=32. Fixes llvm#102566 and 102568.
|
@llvm/pr-subscribers-backend-risc-v Author: Craig Topper (topperc) ChangesThis is needed to match #102405 which disabled fixed to scalable vector lowering for VLEN=32. Fixes #102566 and #102568. Full diff: https://github.com/llvm/llvm-project/pull/102974.diff 2 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 3beaa5198259d0..6fde09d89e4839 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -21024,6 +21024,11 @@ EVT RISCVTargetLowering::getOptimalMemOpType(const MemOp &Op,
// which ends up using scalar sequences.
return MVT::Other;
+ // If the minimum VLEN is less than RISCV::RVVBitsPerBlock we don't support
+ // fixed vectors.
+ if (MinVLenInBytes <= RISCV::RVVBitsPerBlock / 8)
+ return MVT::Other;
+
// Prefer i8 for non-zero memset as it allows us to avoid materializing
// a large scalar constant and instead use vmv.v.x/i to do the
// broadcast. For everything else, prefer ELenVT to minimize VL and thus
diff --git a/llvm/test/CodeGen/RISCV/rvv/memcpy-crash-zvl32b.ll b/llvm/test/CodeGen/RISCV/rvv/memcpy-crash-zvl32b.ll
new file mode 100644
index 00000000000000..e020fe1a0aa1ac
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/rvv/memcpy-crash-zvl32b.ll
@@ -0,0 +1,17 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=riscv64 -mattr=+zve32x | FileCheck %s
+
+; Make sure we don't with VLEN=32.
+
+define void @c() {
+; CHECK-LABEL: c:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: lw a0, 0(zero)
+; CHECK-NEXT: sw a0, 0(zero)
+; CHECK-NEXT: ret
+entry:
+ call void @llvm.memcpy.p0.p0.i64(ptr null, ptr null, i64 4, i1 false)
+ ret void
+}
+
+declare void @llvm.memcpy.p0.p0.i64(ptr noalias nocapture writeonly, ptr noalias nocapture readonly, i64, i1 immarg) #1
|
I have no particular objection to this being backported, but I do want to clearly note that a minimum VLEN of 32 (i.e. zve32x,vze32f) is not a supported configuration. There are probably other problems lurking around this area, and I don't want to give anyone the expectation that this is generally expected to work. At the moment, trying to use zve32x/zve32f is living very dangerously. |
zve32x/zve32f are supposed to work as long as zvl64b or higher is also used. |
Correct, sorry I should have been more clear about the dangerous configuration. This review is specifically about the zvl32b case. |
Yes, I tested with my case, when I pass |
-fno-vectorize |
Thanks |
…is 32. (llvm#102974) This is needed to match llvm#102405 which disabled fixed to scalable vector lowering for VLEN=32. Fixes llvm#102566 and llvm#102568.
This is needed to match #102405 which disabled fixed to scalable vector lowering for VLEN=32.
Fixes #102566 and #102568.