[AArch64] Lower v1i64 and v2i64 [S|U][MIN|MAX] to SVE when available

[MacDue]

The predicate is likely to be hoisted, so in a loop, this would result in a single SVE instruction, which should have lower latency.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: Benjamin Maxwell (MacDue)

Changes

The predicate is likely to be hoisted, so in a loop, this would result in a single SVE instruction, which should have lower latency.

---

Full diff: https://github.com/llvm/llvm-project/pull/166735.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+3-1)
• (added) llvm/test/CodeGen/AArch64/vector-minmax.ll (+119)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index d08f9b94227a2..763d27716a4fa 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -11330,9 +11330,11 @@ SDValue AArch64TargetLowering::LowerMinMax(SDValue Op,
     break;
   }
 
+  // Note: This lowering is only used for v1i64 and v2i64, where we prefer using
+  // SVE if available.
   if (VT.isScalableVector() ||
       useSVEForFixedLengthVectorVT(
-          VT, /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors())) {
+          VT, /*OverrideNEON=*/Subtarget->isSVEorStreamingSVEAvailable())) {
     switch (Opcode) {
     default:
       llvm_unreachable("Wrong instruction");
diff --git a/llvm/test/CodeGen/AArch64/vector-minmax.ll b/llvm/test/CodeGen/AArch64/vector-minmax.ll
new file mode 100644
index 0000000000000..6696f94d404c5
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/vector-minmax.ll
@@ -0,0 +1,119 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc < %s -mtriple=aarch64 -mattr=+neon | FileCheck %s
+; RUN: llc < %s -mtriple=aarch64 -mattr=+neon,+sve | FileCheck %s --check-prefix=CHECK-SVE
+
+define <2 x i64> @smax_v2i64(<2 x i64> %a, <2 x i64> %b){
+; CHECK-LABEL: smax_v2i64:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    cmgt v2.2d, v0.2d, v1.2d
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: smax_v2i64:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    ptrue p0.d, vl2
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-SVE-NEXT:    // kill: def $q1 killed $q1 def $z1
+; CHECK-SVE-NEXT:    smax z0.d, p0/m, z0.d, z1.d
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 killed $z0
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <2 x i64> @llvm.smax.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %0
+}
+
+define <2 x i64> @smin_v2i64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: smin_v2i64:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    cmgt v2.2d, v1.2d, v0.2d
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: smin_v2i64:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    ptrue p0.d, vl2
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-SVE-NEXT:    // kill: def $q1 killed $q1 def $z1
+; CHECK-SVE-NEXT:    smin z0.d, p0/m, z0.d, z1.d
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 killed $z0
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <2 x i64> @llvm.smin.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %0
+}
+
+define <2 x i64> @umax_v2i64(<2 x i64> %a, <2 x i64> %b){
+; CHECK-LABEL: umax_v2i64:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    cmhi v2.2d, v0.2d, v1.2d
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: umax_v2i64:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    ptrue p0.d, vl2
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-SVE-NEXT:    // kill: def $q1 killed $q1 def $z1
+; CHECK-SVE-NEXT:    umax z0.d, p0/m, z0.d, z1.d
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 killed $z0
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <2 x i64> @llvm.umax.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %0
+}
+
+define <2 x i64> @umin_v2i64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: umin_v2i64:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    cmhi v2.2d, v1.2d, v0.2d
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: umin_v2i64:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    ptrue p0.d, vl2
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 def $z0
+; CHECK-SVE-NEXT:    // kill: def $q1 killed $q1 def $z1
+; CHECK-SVE-NEXT:    umin z0.d, p0/m, z0.d, z1.d
+; CHECK-SVE-NEXT:    // kill: def $q0 killed $q0 killed $z0
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <2 x i64> @llvm.umin.v2i64(<2 x i64> %a, <2 x i64> %b)
+  ret <2 x i64> %0
+}
+
+define <1 x i64> @smax_v1i64(<1 x i64> %a, <1 x i64> %b){
+; CHECK-LABEL: smax_v1i64:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    cmgt d2, d0, d1
+; CHECK-NEXT:    bif v0.8b, v1.8b, v2.8b
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: smax_v1i64:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    ptrue p0.d, vl1
+; CHECK-SVE-NEXT:    // kill: def $d0 killed $d0 def $z0
+; CHECK-SVE-NEXT:    // kill: def $d1 killed $d1 def $z1
+; CHECK-SVE-NEXT:    smax z0.d, p0/m, z0.d, z1.d
+; CHECK-SVE-NEXT:    // kill: def $d0 killed $d0 killed $z0
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <1 x i64> @llvm.smax.v2i64(<1 x i64> %a, <1 x i64> %b)
+  ret <1 x i64> %0
+}
+
+; This is legal for Neon, so this should use the Neon smax.
+define <4 x i32> @smax_v4i32(<4 x i32> %a, <4 x i32> %b){
+; CHECK-LABEL: smax_v4i32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    smax v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    ret
+;
+; CHECK-SVE-LABEL: smax_v4i32:
+; CHECK-SVE:       // %bb.0: // %entry
+; CHECK-SVE-NEXT:    smax v0.4s, v0.4s, v1.4s
+; CHECK-SVE-NEXT:    ret
+entry:
+  %0 = call <4 x i32> @llvm.smax.v4i32(<4 x i32> %a, <4 x i32> %b)
+  ret <4 x i32> %0
+}

[SamTebbs33]

LGTM