[llvm] [DXIL] Add scalarization support for WaveReadLaneAt (PR #112570)

[Finn Plummer via llvm-commits]

https://github.com/inbelic created https://github.com/llvm/llvm-project/pull/112570

None

>From b48e3eca63d1c0fff217ec833dd8c9619da766dd Mon Sep 17 00:00:00 2001
From: Finn Plummer <canadienfinn at gmail.com>
Date: Wed, 16 Oct 2024 08:57:34 -0700
Subject: [PATCH] [DXIL] Add scalarization support for WaveReadLaneAt

---
 .../DirectX/DirectXTargetTransformInfo.cpp    |  3 ++
 .../CodeGen/DirectX/WaveReadLaneAt-vec.ll     | 35 +++++++++++++++++++
 2 files changed, 38 insertions(+)
 create mode 100644 llvm/test/CodeGen/DirectX/WaveReadLaneAt-vec.ll

diff --git a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
index be714b5c87895a..8ea31401121bce 100644
--- a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
+++ b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp
@@ -18,6 +18,8 @@ using namespace llvm;
 bool DirectXTTIImpl::isTargetIntrinsicWithScalarOpAtArg(Intrinsic::ID ID,
                                                         unsigned ScalarOpdIdx) {
   switch (ID) {
+  case Intrinsic::dx_wave_readlane:
+    return ScalarOpdIdx == 1;
   default:
     return false;
   }
@@ -28,6 +30,7 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable(
   switch (ID) {
   case Intrinsic::dx_frac:
   case Intrinsic::dx_rsqrt:
+  case Intrinsic::dx_wave_readlane:
     return true;
   default:
     return false;
diff --git a/llvm/test/CodeGen/DirectX/WaveReadLaneAt-vec.ll b/llvm/test/CodeGen/DirectX/WaveReadLaneAt-vec.ll
new file mode 100644
index 00000000000000..8c2a11a3557af3
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/WaveReadLaneAt-vec.ll
@@ -0,0 +1,35 @@
+; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-compute %s | FileCheck %s
+
+; Test that for vector values, WaveReadLaneAt scalarizes and maps down to the
+; DirectX op
+
+define noundef <2 x half> @wave_read_lane_v2half(<2 x half> noundef %expr, i32 %idx) {
+entry:
+; CHECK: call half @dx.op.waveReadLaneAt.f16(i32 117, half %expr.i0, i32 %idx)
+; CHECK: call half @dx.op.waveReadLaneAt.f16(i32 117, half %expr.i1, i32 %idx)
+  %ret = call <2 x half> @llvm.dx.wave.readlane.f16(<2 x half> %expr, i32 %idx)
+  ret <2 x half> %ret
+}
+
+define noundef <3 x i32> @wave_read_lane_v3i32(<3 x i32> noundef %expr, i32 %idx) {
+entry:
+; CHECK: call i32 @dx.op.waveReadLaneAt.i32(i32 117, i32 %expr.i0, i32 %idx)
+; CHECK: call i32 @dx.op.waveReadLaneAt.i32(i32 117, i32 %expr.i1, i32 %idx)
+; CHECK: call i32 @dx.op.waveReadLaneAt.i32(i32 117, i32 %expr.i2, i32 %idx)
+  %ret = call <3 x i32> @llvm.dx.wave.readlane(<3 x i32> %expr, i32 %idx)
+  ret <3 x i32> %ret
+}
+
+define noundef <4 x double> @wave_read_lane_v4f64(<4 x double> noundef %expr, i32 %idx) {
+entry:
+; CHECK: call double @dx.op.waveReadLaneAt.f64(i32 117, double %expr.i0, i32 %idx)
+; CHECK: call double @dx.op.waveReadLaneAt.f64(i32 117, double %expr.i1, i32 %idx)
+; CHECK: call double @dx.op.waveReadLaneAt.f64(i32 117, double %expr.i2, i32 %idx)
+; CHECK: call double @dx.op.waveReadLaneAt.f64(i32 117, double %expr.i3, i32 %idx)
+  %ret = call <4 x double> @llvm.dx.wave.readlane(<4 x double> %expr, i32 %idx)
+  ret <4 x double> %ret
+}
+
+declare <2 x half> @llvm.dx.wave.readlane.v2f16(<2 x half>, i32)
+declare <3 x i32> @llvm.dx.wave.readlane.v3i32(<3 x i32>, i32)
+declare <4 x double> @llvm.dx.wave.readlane.v4f64(<4 x double>, i32)