Apply remat to function arguments

Try to reduce impact of MergeAllocas on register pressure.

Author: stefan-il

IGC/Compiler/CISACodeGen/RematAddressArithmetic.cpp

@@ -527,14 +527,23 @@ void CloneAddressArithmetic::collectInstToProcess(RematSet& ToProcess, Function&
 
             bool IsLoad = llvm::isa<LoadInst>(I);
             bool IsStore = llvm::isa<StoreInst>(I);
-            if (!IsLoad && !IsStore) continue;
+            bool IsCall = llvm::isa<CallInst>(I);
+            if (!IsLoad && !IsStore && !IsCall) continue;
 
             llvm::Value* V =
                 IsLoad ? static_cast<LoadInst *>(&I)->getPointerOperand() :
                 static_cast<StoreInst *>(&I)->getPointerOperand();
 
             if(isRematInstruction(V))
                 ToProcess.insert(static_cast<Instruction *>(V));
+
+            if (IsCall && IGC_IS_FLAG_ENABLED(RematCollectCallArgs)) {
+                for (auto &Arg : cast<CallInst>(I).args()) {
+                    if(isRematInstruction(Arg)) {
+                        ToProcess.insert(cast<Instruction>(&Arg));
+                    }
+                }
+            }
         }
     }
 }

IGC/Compiler/tests/RematAddressArithmetic/callParametarsRemat.ll

@@ -0,0 +1,36 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2025 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; REQUIRES: regkeys
+; UNSUPPORTED: system-windows
+; RUN: igc_opt --typed-pointers %s -S -o - -igc-clone-address-arithmetic --regkey=RematChainLimit=10 --regkey=RematFlowThreshold=100 --regkey=RematRPELimit=0 --dce | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024"
+target triple = "spir64-unknown-unknown"
+
+%"struct.ONEAPIKernelContext::LightSample" = type { %struct.packed_float3, %struct.packed_float3, %struct.packed_float3, float, float, float, float, float, float, i32, i32, i32, i32, i32, i32 }
+%struct.packed_float3 = type { float, float, float }
+
+declare spir_func void @target_func(%"struct.ONEAPIKernelContext::LightSample" addrspace(4)*)
+
+define spir_kernel void @main() {
+entry:
+  %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)
+  ret void
+}

IGC/common/igc_flags.h

@@ -861,7 +861,8 @@ DECLARE_IGC_REGKEY(bool, RematAllowExtractElement, false, "Allow Extract Element
 DECLARE_IGC_REGKEY(bool, RematReassocBefore, false, "Enable short sequence of passes before clone address arithmetic pass to potentially decrese amount of operations that will be rematerialized", false)
 DECLARE_IGC_REGKEY(bool, RematInstCombineBefore, false, "Enable short sequence of passes before clone address arithmetic pass to potentially decrese amount of operations that will be rematerialized", false)
 DECLARE_IGC_REGKEY(bool, RematAddrSpaceCastToUse, true, "Allow rematerialization of inttoptr that are used inside AddrSpaceCastInst", false)
-DECLARE_IGC_REGKEY(bool, RematCallsOperand, false, "Allow rematerialization of inttoptr that are used as call's operand", false)
+DECLARE_IGC_REGKEY(bool, RematCallsOperand, true, "Allow rematerialization of inttoptr that are used as call's operand", false)
+DECLARE_IGC_REGKEY(bool, RematCollectCallArgs, true, "Allow collection of call arguments for rematerialization", false)
 DECLARE_IGC_REGKEY(bool, RematAllowOneUseLoad, false, "Remat allow to move loads that have one use and it's inside the chain", false)
 DECLARE_IGC_REGKEY(bool, RematAllowLoads, false, "Remat allow to move loads, no checks, exclusively for testing purposes", false)
 DECLARE_IGC_REGKEY(bool, DumpRegPressureEstimate, false,  "Dump RegPressureEstimate to a file", false)