[PATCH] D16556: AMDGPU: Restore AMDGPU prefixed rsq intrinsic for now

[Matt Arsenault via llvm-commits]

arsenm updated this revision to Diff 45939.
arsenm added a comment.

Fix clang build


http://reviews.llvm.org/D16556

Files:
  lib/Target/AMDGPU/AMDGPUISelLowering.cpp
  lib/Target/AMDGPU/AMDGPUIntrinsics.td
  lib/Target/AMDGPU/R600ISelLowering.cpp
  lib/Target/AMDGPU/SIISelLowering.cpp
  test/CodeGen/AMDGPU/llvm.AMDGPU.rsq.ll

Index: test/CodeGen/AMDGPU/llvm.AMDGPU.rsq.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/llvm.AMDGPU.rsq.ll
@@ -0,0 +1,33 @@
+; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+
+declare float @llvm.AMDGPU.rsq.f32(float) nounwind readnone
+
+; FUNC-LABEL: {{^}}rsq_f32:
+; SI: v_rsq_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
+; EG: RECIPSQRT_IEEE
+define void @rsq_f32(float addrspace(1)* %out, float %src) nounwind {
+  %rsq = call float @llvm.AMDGPU.rsq.f32(float %src) nounwind readnone
+  store float %rsq, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; TODO: Really these should be constant folded
+; FUNC-LABEL: {{^}}rsq_f32_constant_4.0
+; SI: v_rsq_f32_e32 {{v[0-9]+}}, 4.0
+; EG: RECIPSQRT_IEEE
+define void @rsq_f32_constant_4.0(float addrspace(1)* %out) nounwind {
+  %rsq = call float @llvm.AMDGPU.rsq.f32(float 4.0) nounwind readnone
+  store float %rsq, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}rsq_f32_constant_100.0
+; SI: v_rsq_f32_e32 {{v[0-9]+}}, 0x42c80000
+; EG: RECIPSQRT_IEEE
+define void @rsq_f32_constant_100.0(float addrspace(1)* %out) nounwind {
+  %rsq = call float @llvm.AMDGPU.rsq.f32(float 100.0) nounwind readnone
+  store float %rsq, float addrspace(1)* %out, align 4
+  ret void
+}
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1287,6 +1287,7 @@
   case Intrinsic::amdgcn_rcp:
     return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
   case Intrinsic::amdgcn_rsq:
+  case AMDGPUIntrinsic::AMDGPU_rsq: // Legacy name
     return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
   case Intrinsic::amdgcn_rsq_clamped:
   case Intrinsic::AMDGPU_rsq_clamped: { // Legacy name
Index: lib/Target/AMDGPU/R600ISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/R600ISelLowering.cpp
+++ lib/Target/AMDGPU/R600ISelLowering.cpp
@@ -805,6 +805,7 @@
       return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
 
     case Intrinsic::r600_rsq:
+    case AMDGPUIntrinsic::AMDGPU_rsq: // Legacy name
       // XXX - I'm assuming SI's RSQ_LEGACY matches R600's behavior.
       return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
     }
Index: lib/Target/AMDGPU/AMDGPUIntrinsics.td
===================================================================
--- lib/Target/AMDGPU/AMDGPUIntrinsics.td
+++ lib/Target/AMDGPU/AMDGPUIntrinsics.td
@@ -40,6 +40,11 @@
   def int_AMDGPU_bfm : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_flbit_i32 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
 
+  // Deprecated in favor of llvm.amdgcn.rsq
+  def int_AMDGPU_rsq : Intrinsic<
+    [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]
+  >;
+
   // Deprecated in favor of llvm.bitreverse
   def int_AMDGPU_brev : Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
 
Index: lib/Target/AMDGPU/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -922,10 +922,6 @@
       return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
                          Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
-    case Intrinsic::AMDGPU_rsq_clamped:
-      assert(Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS);
-      return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
-
     case Intrinsic::AMDGPU_ldexp: // Legacy name
       return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
                                                    Op.getOperand(2));


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D16556.45939.patch
Type: text/x-patch
Size: 4113 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160126/7814db73/attachment.bin>