[PATCH] D28099: AMDGPU: Fold fneg into fp_extend

[Matt Arsenault via Phabricator via llvm-commits]

arsenm retitled this revision from "DAGCombiner: Fold fneg into fp_extend if free" to "AMDGPU: Fold fneg into fp_extend ".
arsenm updated this revision to Diff 82893.
arsenm added a comment.
Herald added subscribers: tony-tye, yaxunl, kzhuravl.

Move into target


https://reviews.llvm.org/D28099

Files:
  lib/Target/AMDGPU/AMDGPUISelLowering.cpp
  test/CodeGen/AMDGPU/fneg-combines.ll


Index: test/CodeGen/AMDGPU/fneg-combines.ll
===================================================================
--- test/CodeGen/AMDGPU/fneg-combines.ll
+++ test/CodeGen/AMDGPU/fneg-combines.ll
@@ -655,6 +655,62 @@
   ret void
 }
 
+; --------------------------------------------------------------------------------
+; fp_extend tests
+; --------------------------------------------------------------------------------
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_cvt_f64_f32_e64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+define void @v_fneg_fp_extend_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fpext = fpext float %a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store double %fneg, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_cvt_f64_f32_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+define void @v_fneg_fp_extend_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fneg.a = fsub float -0.000000e+00, %a
+  %fpext = fpext float %fneg.a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store double %fneg, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_store_use_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN-DAG: v_cvt_f64_f32_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]]
+; GCN-DAG: v_xor_b32_e32 [[FNEG_A:v[0-9]+]], 0x80000000, [[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+; GCN: buffer_store_dword [[FNEG_A]]
+define void @v_fneg_fp_extend_store_use_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fneg.a = fsub float -0.000000e+00, %a
+  %fpext = fpext float %fneg.a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store volatile double %fneg, double addrspace(1)* %out.gep
+  store volatile float %fneg.a, float addrspace(1)* undef
+  ret void
+}
 
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.fma.f32(float, float, float) #1
Index: lib/Target/AMDGPU/AMDGPUISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2828,6 +2828,21 @@
       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
     return Res;
   }
+  case ISD::FP_EXTEND: {
+    SDValue CvtSrc = N0.getOperand(0);
+    if (CvtSrc.getOpcode() == ISD::FNEG) {
+      // (fneg (fp_extend (fneg x))) -> (fp_extend x)
+      return DAG.getNode(ISD::FP_EXTEND, SL, VT, CvtSrc.getOperand(0));
+    }
+
+    // (fneg (fp_extend x)) -> (fp_extend (fneg x))
+    //
+    // We want to fold negates through extends so they can be absorbed as
+    // modifiers, so we don't need the same conditions we have for folding
+    // through rounds.
+    SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
+    return DAG.getNode(ISD::FP_EXTEND, SL, VT, Neg);
+  }
   default:
     return SDValue();
   }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D28099.82893.patch
Type: text/x-patch
Size: 4094 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170103/9afcffe3/attachment.bin>