[CodeGenPrepare] Consider target memory intrinics as memory use (llvm#159638)

When deciding to sink address instructions into their uses, we check if
it is profitable to do so. The profitability check is based on the types
of uses of this address instruction -- if there are users which are not
memory instructions, then do not fold.

However, this profitability check wasn't considering target intrinsics,
which may be loads / stores.

This adds some logic to handle target memory intrinsics.

Author: jrbyrnes

llvm/lib/CodeGen/CodeGenPrepare.cpp

@@ -5596,6 +5596,19 @@ static bool FindAllMemoryUses(
       continue;
     }
 
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UserI)) {
+      SmallVector<Value *, 2> PtrOps;
+      Type *AccessTy;
+      if (!TLI.getAddrModeArguments(II, PtrOps, AccessTy))
+        return true;
+
+      if (!find(PtrOps, U.get()))
+        return true;
+
+      MemoryUses.push_back({&U, AccessTy});
+      continue;
+    }
+
     if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
       if (CI->hasFnAttr(Attribute::Cold)) {
         // If this is a cold call, we can sink the addressing calculation into

llvm/test/CodeGen/AMDGPU/sink-addr-memory-intrinsics.ll

@@ -6,25 +6,23 @@ define amdgpu_kernel void @memoryIntrinstic(ptr addrspace(3) %inptr, i1 %cond, p
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    s_and_b32 s1, s1, 1
-; CHECK-NEXT:    s_add_i32 s3, s0, 0x2000
-; CHECK-NEXT:    s_cmp_eq_u32 s1, 0
+; CHECK-NEXT:    s_bitcmp0_b32 s1, 0
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB0_2
 ; CHECK-NEXT:  ; %bb.1: ; %else
-; CHECK-NEXT:    v_mov_b32_e32 v0, s3
-; CHECK-NEXT:    ds_read_b64_tr_b16 v[2:3], v0
-; CHECK-NEXT:    s_mov_b32 s0, 0x7060302
-; CHECK-NEXT:    s_mov_b32 s1, 0x5040100
+; CHECK-NEXT:    v_mov_b32_e32 v0, s0
+; CHECK-NEXT:    ds_read_b64_tr_b16 v[2:3], v0 offset:8192
+; CHECK-NEXT:    s_mov_b32 s1, 0x7060302
+; CHECK-NEXT:    s_mov_b32 s3, 0x5040100
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    v_perm_b32 v0, v3, v2, s0
-; CHECK-NEXT:    v_perm_b32 v1, v3, v2, s1
+; CHECK-NEXT:    v_perm_b32 v0, v3, v2, s1
+; CHECK-NEXT:    v_perm_b32 v1, v3, v2, s3
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_3
 ; CHECK-NEXT:    s_branch .LBB0_4
 ; CHECK-NEXT:  .LBB0_2:
 ; CHECK-NEXT:    ; implicit-def: $vgpr1
 ; CHECK-NEXT:  .LBB0_3: ; %then
-; CHECK-NEXT:    v_mov_b32_e32 v0, s3
-; CHECK-NEXT:    ds_read_b64_tr_b16 v[2:3], v0
+; CHECK-NEXT:    v_mov_b32_e32 v0, s0
+; CHECK-NEXT:    ds_read_b64_tr_b16 v[2:3], v0 offset:8192
 ; CHECK-NEXT:    s_mov_b32 s0, 0x5040100
 ; CHECK-NEXT:    s_mov_b32 s1, 0x7060302
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)