[PATCH] D39040: AMDGPU: Fix creating invalid copy when adjusting dmask

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, kzhuravl.

I'm making a guess as to how dmask works. I'm not sure if
the subregister index should always be sub0 or not.


https://reviews.llvm.org/D39040

Files:
  lib/Target/AMDGPU/SIISelLowering.cpp
  test/CodeGen/AMDGPU/adjust-writemask-invalid-copy.ll


Index: test/CodeGen/AMDGPU/adjust-writemask-invalid-copy.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/adjust-writemask-invalid-copy.ll
@@ -0,0 +1,38 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}adjust_writemask_crash_0:
+; GCN: image_get_lod v[0:1], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}} dmask:0x2
+; GCN-NOT: v1
+; GCN-NOT: v0
+; GCN: buffer_store_dword v1
+define amdgpu_ps void @adjust_writemask_crash_0() #0 {
+main_body:
+  %tmp = call <2 x float> @llvm.amdgcn.image.getlod.v2f32.v2f32.v8i32(<2 x float> undef, <8 x i32> undef, <4 x i32> undef, i32 3, i1 false, i1 false, i1 false, i1 false, i1 false)
+  %tmp1 = bitcast <2 x float> %tmp to <2 x i32>
+  %tmp2 = shufflevector <2 x i32> %tmp1, <2 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+  %tmp3 = bitcast <4 x i32> %tmp2 to <4 x float>
+  %tmp4 = extractelement <4 x float> %tmp3, i32 0
+  store volatile float %tmp4, float addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}adjust_writemask_crash_1:
+; GCN: image_get_lod v[0:1], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}} dmask:0x1
+; GCN-NOT: v1
+; GCN-NOT: v0
+; GCN: buffer_store_dword v0
+define amdgpu_ps void @adjust_writemask_crash_1() #0 {
+main_body:
+  %tmp = call <2 x float> @llvm.amdgcn.image.getlod.v2f32.v2f32.v8i32(<2 x float> undef, <8 x i32> undef, <4 x i32> undef, i32 3, i1 false, i1 false, i1 false, i1 false, i1 false)
+  %tmp1 = bitcast <2 x float> %tmp to <2 x i32>
+  %tmp2 = shufflevector <2 x i32> %tmp1, <2 x i32> undef, <4 x i32> <i32 1, i32 0, i32 undef, i32 undef>
+  %tmp3 = bitcast <4 x i32> %tmp2 to <4 x float>
+  %tmp4 = extractelement <4 x float> %tmp3, i32 1
+  store volatile float %tmp4, float addrspace(1)* undef
+  ret void
+}
+
+declare <2 x float> @llvm.amdgcn.image.getlod.v2f32.v2f32.v8i32(<2 x float>, <8 x i32>, <4 x i32>, i32, i1, i1, i1, i1, i1) #1
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readonly }
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6428,13 +6428,14 @@
 
   // If we only got one lane, replace it with a copy
   // (if NewDmask has only one bit set...)
-  if (NewDmask && (NewDmask & (NewDmask-1)) == 0) {
-    SDValue RC = DAG.getTargetConstant(AMDGPU::VGPR_32RegClassID, SDLoc(),
-                                       MVT::i32);
-    SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
-                                      SDLoc(), Users[Lane]->getValueType(0),
-                                      SDValue(Node, 0), RC);
-    DAG.ReplaceAllUsesWith(Users[Lane], Copy);
+  if (NewDmask && (NewDmask & (NewDmask - 1)) == 0) {
+    const SIRegisterInfo *TRI = getSubtarget()->getRegisterInfo();
+    unsigned SubReg = TRI->getSubRegFromChannel(countTrailingZeros(NewDmask));
+    //SDValue Copy = DAG.getTargetExtractSubreg(AMDGPU::sub0, SDLoc(Node),
+    SDValue Copy = DAG.getTargetExtractSubreg(SubReg, SDLoc(Node),
+                                              Users[Lane]->getValueType(0),
+                                              SDValue(Node, 0));
+    DAG.ReplaceAllUsesWith(Users[Lane], Copy.getNode());
     return;
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D39040.119431.patch
Type: text/x-patch
Size: 3422 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171018/2c37e964/attachment.bin>