[llvm] [AArch64][SVE] Don't require 16-byte aligned SVE loads/stores with +strict-align (PR #119732)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-aarch64

Author: Benjamin Maxwell (MacDue)

<details>
<summary>Changes</summary>

Instead, allow any alignment >= the element size (in bytes). This is all that is needed for vector loads even if unaligned accesses are disabled.

See: https://developer.arm.com/documentation/ddi0602/2024-09/Shared-Pseudocode/aarch64-functions-memory?lang=en#impl-aarch64.Mem.read.3

Specifically:
```
// Check alignment on size of element accessed, not overall access size.
constant integer alignment = if accdesc.ispair then size DIV 2 else size;
```

The `size` passed to `Mem` by SVE load/store instructions is the element size.

---
Full diff: https://github.com/llvm/llvm-project/pull/119732.diff


2 Files Affected:

- (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+10) 
- (added) llvm/test/CodeGen/AArch64/sve-load-store-strict-align.ll (+64) 


``````````diff
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index e61dedb2477560..ee882dbae06537 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -2569,6 +2569,16 @@ MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(
     EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
     unsigned *Fast) const {
+
+  // Allow SVE loads/stores where the alignment >= the size of the element type,
+  // even with +strict-align. The SVE loads/stores do not require memory to be
+  // aligned more than the element type even without unaligned accesses.
+  // Without already aligned loads/stores are forced to have 16-byte alignment,
+  // which is unnecessary and fails to build as TLI.expandUnalignedLoad() and
+  // TLI.expandUnalignedStore() don't yet support scalable vectors.
+  if (VT.isScalableVector() && Alignment >= Align(VT.getScalarSizeInBits() / 8))
+      return true;
+
   if (Subtarget->requiresStrictAlign())
     return false;
 
diff --git a/llvm/test/CodeGen/AArch64/sve-load-store-strict-align.ll b/llvm/test/CodeGen/AArch64/sve-load-store-strict-align.ll
new file mode 100644
index 00000000000000..d9c7ef5ddc88c0
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sve-load-store-strict-align.ll
@@ -0,0 +1,64 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve,+strict-align < %s | FileCheck %s
+
+define void @nxv16i8(ptr %ldptr, ptr %stptr) {
+; CHECK-LABEL: nxv16i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; CHECK-NEXT:    st1b { z0.b }, p0, [x1]
+; CHECK-NEXT:    ret
+  %l3 = load <vscale x 16 x i8>, ptr %ldptr, align 1
+  store <vscale x 16 x i8> %l3, ptr %stptr, align 1
+  ret void
+}
+
+define void @nxv8i16(ptr %ldptr, ptr %stptr) {
+; CHECK-LABEL: nxv8i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; CHECK-NEXT:    st1h { z0.h }, p0, [x1]
+; CHECK-NEXT:    ret
+  %l3 = load <vscale x 8 x i16>, ptr %ldptr, align 2
+  store <vscale x 8 x i16> %l3, ptr %stptr, align 2
+  ret void
+}
+
+define void @nxv4i32(ptr %ldptr, ptr %stptr) {
+; CHECK-LABEL: nxv4i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-NEXT:    st1w { z0.s }, p0, [x1]
+; CHECK-NEXT:    ret
+  %l3 = load <vscale x 4 x i32>, ptr %ldptr, align 4
+  store <vscale x 4 x i32> %l3, ptr %stptr, align 4
+  ret void
+}
+
+define void @nxv2i64(ptr %ldptr, ptr %stptr) {
+; CHECK-LABEL: nxv2i64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-NEXT:    st1d { z0.d }, p0, [x1]
+; CHECK-NEXT:    ret
+  %l3 = load <vscale x 2 x i64>, ptr %ldptr, align 8
+  store <vscale x 2 x i64> %l3, ptr %stptr, align 8
+  ret void
+}
+
+; FIXME: Support TLI.expandUnalignedLoad()/TLI.expandUnalignedStore() for SVE.
+; define void @unaligned_nxv2i64(ptr %ldptr, ptr %stptr) {
+; ; CHECK-LABEL: nxv2i64:
+; ; CHECK:       // %bb.0:
+; ; CHECK-NEXT:    ptrue p0.d
+; ; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; ; CHECK-NEXT:    st1d { z0.d }, p0, [x1]
+; ; CHECK-NEXT:    ret
+;   %l3 = load <vscale x 2 x i64>, ptr %ldptr, align 4
+;   store <vscale x 2 x i64> %l3, ptr %stptr, align 4
+;   ret void
+; }

``````````

</details>


https://github.com/llvm/llvm-project/pull/119732