[NVPTX] Do not mark move of global address as cheap enabling more CSE

llvm/lib/Target/NVPTX/NVPTXInstrInfo.td

@@ -1510,18 +1510,22 @@ let hasSideEffects = false in {
                                     "mov.b64 \t$d, __local_depot$num;">;
 }
 
-
-// copyPhysreg is hard-coded in NVPTXInstrInfo.cpp
-let hasSideEffects = false, isAsCheapAsAMove = true in {
-  let isMoveReg = true in
+let hasSideEffects = false in {
+  let isMoveReg = true, isAsCheapAsAMove = true in
     class MOVr<RegisterClass RC, string OpStr> :
       BasicNVPTXInst<(outs RC:$dst), (ins RC:$src), "mov." # OpStr>;
 
-  let isMoveImm = true in
+  let isMoveImm = true, isAsCheapAsAMove = true in
     class MOVi<RegTyInfo t, string suffix> :
       BasicNVPTXInst<(outs t.RC:$dst), (ins t.Imm:$src),
               "mov." # suffix,
               [(set t.Ty:$dst, t.ImmNode:$src)]>;
+
+  // We don't want to set isAsCheapAsAMove to true for these instructions as
+  // this would prevent CSE and resulted in regressions.
+  class MovSymInst<RegTyInfo t> :
+    BasicNVPTXInst<(outs t.RC:$dst), (ins Operand<t.Ty>:$src),
+                   "mov.b" # t.Size>;
 }
 
 def MOV_B1_r : MOVr<B1, "pred">;
@@ -1539,6 +1543,9 @@ def MOV_BF16_i : MOVi<BF16RT, "b16">;
 def MOV_F32_i  : MOVi<F32RT, "b32">;
 def MOV_F64_i  : MOVi<F64RT, "b64">;
 
+def MOV_B32_sym : MovSymInst<I32RT>;
+def MOV_B64_sym : MovSymInst<I64RT>;
+
 
 def to_tglobaladdr : SDNodeXForm<globaladdr, [{
   return CurDAG->getTargetGlobalAddress(N->getGlobal(), SDLoc(N),
@@ -1555,11 +1562,11 @@ def to_tframeindex : SDNodeXForm<frameindex, [{
   return CurDAG->getTargetFrameIndex(N->getIndex(), N->getValueType(0));
 }]>;
 
-def : Pat<(i32 globaladdr:$dst), (MOV_B32_i (to_tglobaladdr $dst))>;
-def : Pat<(i64 globaladdr:$dst), (MOV_B64_i (to_tglobaladdr $dst))>;
+def : Pat<(i32 globaladdr:$dst), (MOV_B32_sym (to_tglobaladdr $dst))>;
+def : Pat<(i64 globaladdr:$dst), (MOV_B64_sym (to_tglobaladdr $dst))>;
 
-def : Pat<(i32 externalsym:$dst), (MOV_B32_i (to_texternsym $dst))>;
-def : Pat<(i64 externalsym:$dst), (MOV_B64_i (to_texternsym $dst))>;
+def : Pat<(i32 externalsym:$dst), (MOV_B32_sym (to_texternsym $dst))>;
+def : Pat<(i64 externalsym:$dst), (MOV_B64_sym (to_texternsym $dst))>;
 
 //---- Copy Frame Index ----
 def LEA_ADDRi :   NVPTXInst<(outs B32:$dst), (ins ADDR:$addr),

llvm/test/CodeGen/NVPTX/cse-mov-sym.ll

@@ -0,0 +1,59 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s | FileCheck %s
+
+target triple = "nvptx64-nvidia-cuda"
+
+@global_smem = external addrspace(3) global [0 x i8], align 16
+
+
+;; Confirm the mov.b64 of global_smem is CSE'd. We need to make things a bit
+;; complex with a loop to make this interesting.
+define i32 @test_mov_sym(i32 %offset1, i32 %offset2, i1 %cond) {
+; CHECK-LABEL: test_mov_sym(
+; CHECK:       {
+; CHECK-NEXT:    .reg .pred %p<4>;
+; CHECK-NEXT:    .reg .b16 %rs<3>;
+; CHECK-NEXT:    .reg .b32 %r<8>;
+; CHECK-NEXT:    .reg .b64 %rd<7>;
+; CHECK-EMPTY:
+; CHECK-NEXT:  // %bb.0: // %entry
+; CHECK-NEXT:    ld.param.b8 %rs1, [test_mov_sym_param_2];
+; CHECK-NEXT:    and.b16 %rs2, %rs1, 1;
+; CHECK-NEXT:    setp.ne.b16 %p1, %rs2, 0;
+; CHECK-NEXT:    ld.param.b32 %r4, [test_mov_sym_param_0];
+; CHECK-NEXT:    cvt.s64.s32 %rd1, %r4;
+; CHECK-NEXT:    mov.b64 %rd2, global_smem;
+; CHECK-NEXT:    add.s64 %rd3, %rd2, %rd1;
+; CHECK-NEXT:    ld.shared.b32 %r7, [%rd3];
+; CHECK-NEXT:    not.pred %p2, %p1;
+; CHECK-NEXT:    @%p2 bra $L__BB0_4;
+; CHECK-NEXT:  // %bb.1: // %if1.preheader
+; CHECK-NEXT:    ld.param.b32 %r5, [test_mov_sym_param_1];
+; CHECK-NEXT:    setp.ne.b32 %p3, %r4, %r5;
+; CHECK-NEXT:  $L__BB0_2: // %if1
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    @%p3 bra $L__BB0_2;
+; CHECK-NEXT:  // %bb.3: // %if2
+; CHECK-NEXT:    cvt.s64.s32 %rd4, %r5;
+; CHECK-NEXT:    add.s64 %rd6, %rd2, %rd4;
+; CHECK-NEXT:    ld.shared.b32 %r6, [%rd6];
+; CHECK-NEXT:    add.s32 %r7, %r7, %r6;
+; CHECK-NEXT:  $L__BB0_4: // %end
+; CHECK-NEXT:    st.param.b32 [func_retval0], %r7;
+; CHECK-NEXT:    ret;
+entry:
+    %gep = getelementptr inbounds i8, ptr addrspace(3) @global_smem, i32 %offset1
+    %val = load i32, ptr addrspace(3) %gep
+    br i1 %cond, label %if1, label %end
+if1:
+    %cond2 = icmp eq i32 %offset1, %offset2
+    br i1 %cond2, label %if2, label %if1
+if2:
+    %gep2 = getelementptr inbounds i8, ptr addrspace(3) @global_smem, i32 %offset2
+    %val2 = load i32, ptr addrspace(3) %gep2
+    %add = add i32 %val, %val2
+    br label %end
+end:
+    %ret = phi i32 [ %add, %if2 ], [ %val, %entry ]
+    ret i32 %ret
+}