Only Rematerialize ptr bitcasts for function calls inside

CloneAddressArithmetic

Only Rematerialize ptr bitcasts for function calls inside
CloneAddressArithmetic

Author: esukhov

IGC/Compiler/CISACodeGen/RematAddressArithmetic.cpp

@@ -515,7 +515,8 @@ bool isRematInstruction(llvm::Value* V) {
 
     bool IntToPtr = llvm::isa<IntToPtrInst>(V);
     bool AddrSpCast = llvm::isa<AddrSpaceCastInst>(V);
-    bool BitCast = llvm::isa<BitCastInst>(V);
+    // use only bitcasts on pointers as a seed instruction
+    bool BitCast = llvm::isa<BitCastInst>(V) && V->getType()->isPointerTy();
     bool GEP = llvm::isa<GetElementPtrInst>(V);
 
     bool Result = IntToPtr || AddrSpCast || BitCast || GEP;
@@ -530,6 +531,7 @@ void CloneAddressArithmetic::collectInstToProcess(RematSet& ToProcess, Function&
             bool IsLoad = llvm::isa<LoadInst>(I);
             bool IsStore = llvm::isa<StoreInst>(I);
             bool IsCall = llvm::isa<CallInst>(I);
+
             if (!IsLoad && !IsStore && !IsCall) continue;
 
             llvm::Value* V =

…AddressArithmetic/callParametarsRemat.ll …AddressArithmetic/callParametersRemat.llIGC/Compiler/tests/RematAddressArithmetic/callParametarsRemat.ll renamed to IGC/Compiler/tests/RematAddressArithmetic/callParametersRemat.ll IGC/Compiler/tests/RematAddressArithmetic/callParametarsRemat.ll renamed to IGC/Compiler/tests/RematAddressArithmetic/callParametersRemat.ll

@@ -17,6 +17,7 @@ target triple = "spir64-unknown-unknown"
 %struct.packed_float3 = type { float, float, float }
 
 declare spir_func void @target_func(%"struct.ONEAPIKernelContext::LightSample" addrspace(4)*)
+declare spir_func void @target_func_bitcast_no_ptr(double)
 
 define spir_kernel void @main() {
 entry:
@@ -34,3 +35,31 @@ cleanup.cont.i:                                   ; preds = %if.end46
   call spir_func void @target_func(%"struct.ONEAPIKernelContext::LightSample" addrspace(4)* %ls.ascast.i)
   ret void
 }
+
+define spir_func void @bar(i64 %arg1) {
+entry:
+  %val2 = mul i64 %arg1, 2
+  %val3 = mul i64 %arg1, 3
+  %cast_val2 = bitcast i64 %val2 to double
+  %cast_val3 = bitcast i64 %val3 to double
+  ; CHECK: [[VAL2:%.*]] = mul i64 [[ARG1:%.*]], 2
+  ; CHECK: [[VAL3:%.*]] = mul i64 [[ARG1]], 3
+  ; CHECK: bitcast i64 [[VAL2]] to double
+  ; CHECK: bitcast i64 [[VAL3]] to double
+  ; CHECK: br label %if.end46
+  %ls.i = alloca %"struct.ONEAPIKernelContext::LightSample", i32 0, align 4
+  br label %if.end46
+
+if.end46:
+  %ls.ascast.i = addrspacecast %"struct.ONEAPIKernelContext::LightSample"* %ls.i to %"struct.ONEAPIKernelContext::LightSample" addrspace(4)*
+  br label %cleanup.cont.i
+; CHECK-LABEL: cleanup.cont.i:
+; CHECK-NEXT: [[REMAT:%.*]] = alloca %"struct.ONEAPIKernelContext::LightSample", i32 0, align 4
+; CHECK-NEXT: [[CAST:%.*]] = addrspacecast %"struct.ONEAPIKernelContext::LightSample"* [[REMAT]] to %"struct.ONEAPIKernelContext::LightSample" addrspace(4)*
+; CHECK-NEXT: call spir_func void @target_func(%"struct.ONEAPIKernelContext::LightSample" addrspace(4)* [[CAST]])
+cleanup.cont.i:                                   ; preds = %if.end46
+  call spir_func void @target_func(%"struct.ONEAPIKernelContext::LightSample" addrspace(4)* %ls.ascast.i)
+  call spir_func void @target_func_bitcast_no_ptr(double %cast_val2)
+  call spir_func void @target_func_bitcast_no_ptr(double %cast_val3)
+  ret void
+}

IGC/Compiler/tests/RematAddressArithmetic/cloneAddressArithmetic.ll

@@ -34,3 +34,32 @@ define spir_kernel void @main(double addrspace(1)* %base, i64 %offset, i64 %I, i
   store double %rr0, double addrspace(1)* %a4
   ret void
 }
+
+
+define spir_func void @bar(double addrspace(1)* %base, i64 %offset, i64 %I, i64 %J) {
+
+; CHECK: define spir_func void @bar(double addrspace(1)* [[BASE:%.*]], i64 [[OFFSET:%.*]], i64 %I, i64 %J)
+
+  %baseArith = ptrtoint double addrspace(1)* %base to i64
+  ; CHECK: [[FIRSTBASE:%.*]] = ptrtoint double addrspace(1)* [[BASE]] to i64
+  ; CHECK: [[FIRSTOFFSET:%.*]] = mul nuw nsw i64 [[OFFSET]], 207368
+  %basePtr = mul nuw nsw i64 %offset, 207368
+  %offsetI = mul nsw i64 %I, 1288
+  %offsetJ = shl nsw i64 %J, 3
+
+  %a0 = add i64 %baseArith, 100780848
+  %a1 = add i64 %a0, %basePtr
+  %a2 = add i64 %a1, %offsetI
+  %a3 = add i64 %a2, %offsetJ
+  %a4 = inttoptr i64 %a3 to double addrspace(1)*
+  %a5 = inttoptr i64 %a3 to i64 addrspace(1)*
+  %r0 = load double, double addrspace(1)* %a4, align 8
+
+  %rr0 = fmul double %r0, 2.0
+
+  ; CHECK: [[SECONDBASE:%.*]] = ptrtoint double addrspace(1)* [[BASE]] to i64
+  ; CHECK: [[SECONDOFFSET:%.*]] = mul nuw nsw i64 [[OFFSET]], 207368
+  %cast_a5 = bitcast i64 addrspace(1)* %a5 to double addrspace(1)*
+  store double %rr0, double addrspace(1)* %cast_a5
+  ret void
+}