[PATCH] D12224: AMDGPU: Fix not moving users of s_bfe_i64 to VALU

[Matt Arsenault via llvm-commits]

arsenm created this revision.
arsenm added a reviewer: tstellarAMD.
arsenm added a subscriber: llvm-commits.
Herald added a subscriber: arsenm.

This wouldn't propagate to users of the original BFE and would hit a verifier error.

http://reviews.llvm.org/D12224

Files:
  lib/Target/AMDGPU/SIInstrInfo.cpp
  test/CodeGen/AMDGPU/sext-in-reg.ll

Index: test/CodeGen/AMDGPU/sext-in-reg.ll
===================================================================
--- test/CodeGen/AMDGPU/sext-in-reg.ll
+++ test/CodeGen/AMDGPU/sext-in-reg.ll
@@ -609,3 +609,53 @@
   store i32 %bfe, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; Make sure we propagate the VALUness to users of a moved scalar BFE.
+
+; FUNC-LABEL: {{^}}v_sext_in_reg_i1_to_i64_move_use:
+; SI: buffer_load_dwordx2
+; SI: v_lshl_b64 v{{\[}}[[VAL_LO:[0-9]+]]:[[VAL_HI:[0-9]+]]{{\]}}
+; SI-DAG: v_bfe_i32 v[[LO:[0-9]+]], v[[VAL_LO]], 0, 1
+; SI-DAG: v_ashrrev_i32_e32 v[[HI:[0-9]+]], 31, v[[LO]]
+; SI-DAG: v_and_b32_e32 v[[RESULT_LO:[0-9]+]], s{{[0-9]+}}, v[[LO]]
+; SI-DAG: v_and_b32_e32 v[[RESULT_HI:[0-9]+]], s{{[0-9]+}}, v[[HI]]
+; SI: buffer_store_dwordx2 v{{\[}}[[RESULT_LO]]:[[RESULT_HI]]{{\]}}
+define void @v_sext_in_reg_i1_to_i64_move_use(i64 addrspace(1)* %out, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr, i64 %s.val) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %a.gep = getelementptr i64, i64 addrspace(1)* %aptr, i32 %tid
+  %b.gep = getelementptr i64, i64 addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %tid
+  %a = load i64, i64 addrspace(1)* %a.gep, align 8
+  %b = load i64, i64 addrspace(1)* %b.gep, align 8
+
+  %c = shl i64 %a, %b
+  %shl = shl i64 %c, 63
+  %ashr = ashr i64 %shl, 63
+
+  %and = and i64 %ashr, %s.val
+  store i64 %and, i64 addrspace(1)* %out.gep, align 8
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_sext_in_reg_i32_to_i64_move_use:
+; SI: buffer_load_dwordx2
+; SI: v_lshl_b64 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}},
+; SI-DAG: v_ashrrev_i32_e32 v[[SHR:[0-9]+]], 31, v[[LO]]
+; SI-DAG: v_and_b32_e32 v[[RESULT_LO:[0-9]+]], s{{[0-9]+}}, v[[LO]]
+; SI-DAG: v_and_b32_e32 v[[RESULT_HI:[0-9]+]], s{{[0-9]+}}, v[[SHR]]
+; SI: buffer_store_dwordx2 v{{\[}}[[RESULT_LO]]:[[RESULT_HI]]{{\]}}
+define void @v_sext_in_reg_i32_to_i64_move_use(i64 addrspace(1)* %out, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr, i64 %s.val) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %a.gep = getelementptr i64, i64 addrspace(1)* %aptr, i32 %tid
+  %b.gep = getelementptr i64, i64 addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %tid
+  %a = load i64, i64 addrspace(1)* %a.gep, align 8
+  %b = load i64, i64 addrspace(1)* %b.gep, align 8
+
+  %c = shl i64 %a, %b
+  %shl = shl i64 %c, 32
+  %ashr = ashr i64 %shl, 32
+  %and = and i64 %ashr, %s.val
+  store i64 %and, i64 addrspace(1)* %out.gep, align 8
+  ret void
+}
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2568,6 +2568,7 @@
       .addImm(AMDGPU::sub1);
 
     MRI.replaceRegWith(Dest.getReg(), ResultReg);
+    addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
     return;
   }
 
@@ -2586,6 +2587,7 @@
     .addImm(AMDGPU::sub1);
 
   MRI.replaceRegWith(Dest.getReg(), ResultReg);
+  addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
 }
 
 void SIInstrInfo::addUsersToMoveToVALUWorklist(


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D12224.32800.patch
Type: text/x-patch
Size: 3155 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150821/7a422d12/attachment.bin>