Improve handling of half element types

- Add tests for half element types, and only sink operands when
  subtargtet has fullfp16
- Refactor scalable test to use target-features attribute, rather than
  -mattr on the RUN line

Author: hazzlim

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -5244,6 +5244,10 @@ bool AArch64TTIImpl::isProfitableToSinkOperands(
     if (I->getType()->isScalableTy())
       return false;
 
+    if (cast<VectorType>(I->getType())->getElementType()->isHalfTy() &&
+        !ST->hasFullFP16())
+      return false;
+
     // Sink splats for index lane variants
     if (isSplatShuffle(I->getOperand(0)))
       Ops.push_back(&I->getOperandUse(0));

llvm/test/CodeGen/AArch64/sinksplat.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve -o - %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -o - %s | FileCheck %s
 
 define <4 x i32> @smull(<4 x i16> %x, ptr %y) {
 ; CHECK-LABEL: smull:
@@ -422,8 +422,91 @@ l2:
   ret <4 x i32> %r
 }
 
+; We shouldn't sink without fullfp16.
+define <4 x half> @fmul_half(ptr %x, ptr %y) {
+; CHECK-LABEL: fmul_half:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    ld1r { v1.4h }, [x0]
+; CHECK-NEXT:    movi d0, #0000000000000000
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:    fcvtl v1.4s, v1.4h
+; CHECK-NEXT:  .LBB13_1: // %l1
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr d2, [x1, x8]
+; CHECK-NEXT:    fcvtl v0.4s, v0.4h
+; CHECK-NEXT:    add x8, x8, #8
+; CHECK-NEXT:    cmp w8, #8
+; CHECK-NEXT:    fcvtl v2.4s, v2.4h
+; CHECK-NEXT:    fmul v2.4s, v2.4s, v1.4s
+; CHECK-NEXT:    fcvtn v2.4h, v2.4s
+; CHECK-NEXT:    fcvtl v2.4s, v2.4h
+; CHECK-NEXT:    fadd v0.4s, v2.4s, v0.4s
+; CHECK-NEXT:    fcvtn v0.4h, v0.4s
+; CHECK-NEXT:    b.eq .LBB13_1
+; CHECK-NEXT:  // %bb.2: // %l2
+; CHECK-NEXT:    ret
+entry:
+  %x.val = load half, ptr %x
+  %x.ins = insertelement <4 x half> poison, half %x.val, i64 0
+  %a = shufflevector <4 x half> %x.ins, <4 x half> undef, <4 x i32> zeroinitializer
+  br label %l1
+
+l1:
+  %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]
+  %q = phi <4 x half> [ zeroinitializer, %entry ], [ %c, %l1 ]
+  %idx.y = mul nuw nsw i32 %p, 4
+  %ptr.y = getelementptr half, ptr %y, i32 %idx.y
+  %l = load <4 x half>, ptr %ptr.y
+  %b = fmul <4 x half> %l, %a
+  %c = fadd <4 x half> %b, %q
+  %pa = add i32 %p, 1
+  %c1 = icmp eq i32 %p, 0
+  br i1 %c1, label %l1, label %l2
+
+l2:
+  ret <4 x half> %c
+}
+
+define <4 x half> @fmul_half_fullfp16(ptr %x, ptr %y) "target-features"="+fullfp16" {
+; CHECK-LABEL: fmul_half_fullfp16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    movi d0, #0000000000000000
+; CHECK-NEXT:    ldr h1, [x0]
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  .LBB14_1: // %l1
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr d2, [x1, x8]
+; CHECK-NEXT:    add x8, x8, #8
+; CHECK-NEXT:    cmp w8, #8
+; CHECK-NEXT:    fmul v2.4h, v2.4h, v1.h[0]
+; CHECK-NEXT:    fadd v0.4h, v2.4h, v0.4h
+; CHECK-NEXT:    b.eq .LBB14_1
+; CHECK-NEXT:  // %bb.2: // %l2
+; CHECK-NEXT:    ret
+entry:
+  %x.val = load half, ptr %x
+  %x.ins = insertelement <4 x half> poison, half %x.val, i64 0
+  %a = shufflevector <4 x half> %x.ins, <4 x half> undef, <4 x i32> zeroinitializer
+  br label %l1
+
+l1:
+  %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]
+  %q = phi <4 x half> [ zeroinitializer, %entry ], [ %c, %l1 ]
+  %idx.y = mul nuw nsw i32 %p, 4
+  %ptr.y = getelementptr half, ptr %y, i32 %idx.y
+  %l = load <4 x half>, ptr %ptr.y
+  %b = fmul <4 x half> %l, %a
+  %c = fadd <4 x half> %b, %q
+  %pa = add i32 %p, 1
+  %c1 = icmp eq i32 %p, 0
+  br i1 %c1, label %l1, label %l2
+
+l2:
+  ret <4 x half> %c
+}
+
 ; We shouldn't sink the splat operand for scalable vectors.
-define <vscale x 4 x float> @fmul_scalable(ptr %x, ptr %y) {
+define <vscale x 4 x float> @fmul_scalable(ptr %x, ptr %y) "target-features"="+sve" {
 ; CHECK-LABEL: fmul_scalable:
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    ptrue p0.s
@@ -433,14 +516,14 @@ define <vscale x 4 x float> @fmul_scalable(ptr %x, ptr %y) {
 ; CHECK-NEXT:    mov w9, #1 // =0x1
 ; CHECK-NEXT:    ld1rw { z1.s }, p0/z, [x0]
 ; CHECK-NEXT:    lsl x8, x8, #2
-; CHECK-NEXT:  .LBB13_1: // %l1
+; CHECK-NEXT:  .LBB15_1: // %l1
 ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    ld1w { z2.s }, p0/z, [x1]
 ; CHECK-NEXT:    subs w9, w9, #1
 ; CHECK-NEXT:    add x1, x1, x8
 ; CHECK-NEXT:    fmul z2.s, z2.s, z1.s
 ; CHECK-NEXT:    fadd z0.s, z2.s, z0.s
-; CHECK-NEXT:    b.eq .LBB13_1
+; CHECK-NEXT:    b.eq .LBB15_1
 ; CHECK-NEXT:  // %bb.2: // %l2
 ; CHECK-NEXT:    ret
 entry: