move scratch surface creation to IGC

fix ip scanning issue

Author: Gang Y Chen

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -3366,10 +3366,68 @@ namespace IGC
                 V(vKernel->GetPredefinedSurface(surfacevar, PREDEFINED_SURFACE_T255));
                 break;
             case ESURFACE_SCRATCH:
-                // NOTE: For scratch surface, we need to shr the surface state offset coming in R0.5 by 4.
-                //       This shr operation is generated by vISA in HDC path
+            {
+                // For scratch surface, we need to shr the surface state offset coming in R0.5 by 4
+                //      This is because the scratch offset is passed in via r0.5[31:10],
+                //      but the BSS/SS descriptor expects the offset in [31:6] bits, thus we must shift it right by 4
+                //      We also need to and r0.5 with 0xFFFFFC00 to retrieve bits 31:10
+
+                // TBD is it needed or we will have the bits 9:0 are reset already in the payload?
+                //     (W) and (1) sso r0.5 0xFFFFC00, placed at kernel entry
+                VISA_GenVar* r0Var = nullptr;
+                VISA_VectorOpnd* surfOpnd_r0_5 = nullptr;
+                uint32_t imm_data_dw = 0xFFFFFC00;
+                VISA_VectorOpnd* andOpnd = nullptr;
+                VISA_VectorOpnd* surfOpAndDst = nullptr;
+
                 V(vKernel->GetPredefinedSurface(surfacevar, PREDEFINED_SURFACE_SCRATCH));
+                V(vKernel->GetPredefinedVar(r0Var, PREDEFINED_R0));
+
+                CVariable* surfOpAndVar = m_program->GetNewVariable(1, ISA_TYPE_UD, EALIGN_DWORD, true, "SurfaceOpnd");
+                V(vKernel->CreateVISADstOperand(surfOpAndDst, GetVISAVariable(surfOpAndVar), 1, 0, 0));
+
+                V(vKernel->CreateVISASrcOperand(surfOpnd_r0_5, r0Var, MODIFIER_NONE, 0, 1, 0, 0, 5));
+                V(vKernel->CreateVISAImmediate(andOpnd, &imm_data_dw, ISA_TYPE_UD));
+                V(vKernel->AppendVISAArithmeticInst(
+                    ISA_AND,
+                    nullptr,
+                    false,
+                    vISA_EMASK_M1_NM,
+                    EXEC_SIZE_1,
+                    surfOpAndDst,
+                    surfOpnd_r0_5,
+                    andOpnd));
+
+                // if use new extend message descriptor format
+                // (W) shr (1) a0.0 ss0 0x4
+                // else
+                // (W) shl (1) a0.0 ss0 0x2
+
+                VISA_VectorOpnd* surfOpndShiftDst = nullptr;
+                VISA_VectorOpnd* surfOpnd_r0_5_bits_31_10 = nullptr;
+                VISA_VectorOpnd* shiftOpnd = nullptr;
+                bool useNewExtMsgFormat = m_program->m_Platform->hasScratchSurface();
+                // Not sure that we have a case of using bindless scratch surface on
+                // platforms that use old ExtMsgFormat a0[12:31] as surface-state
+                // heap offset. However, the logic is maintained when moving from
+                // finalizer to IGC
+                uint16_t imm_data_w = useNewExtMsgFormat ? 4 : 2;
+                ISA_Opcode shiftOpcode = useNewExtMsgFormat ? ISA_SHR : ISA_SHL;
+
+                surfOpnd_r0_5_bits_31_10 = GetSourceOperandNoModifier(surfOpAndVar);
+                V(vKernel->CreateVISAImmediate(shiftOpnd, &imm_data_w, ISA_TYPE_UW));
+                V(vKernel->CreateVISAStateOperand(surfOpndShiftDst, surfacevar, 0, true));
+                V(vKernel->AppendVISAArithmeticInst(
+                    shiftOpcode,
+                    nullptr,
+                    false,
+                    vISA_EMASK_M1_NM,
+                    EXEC_SIZE_1,
+                    surfOpndShiftDst,
+                    surfOpnd_r0_5_bits_31_10,
+                    shiftOpnd));
                 break;
+            }
             default:
                 IGC_ASSERT_MESSAGE(0, "Invalid surface");
                 break;
@@ -7304,8 +7362,6 @@ namespace IGC
             // For scratch surface, we need to shr the surface state offset coming in R0.5 by 4
             //      This is because the scratch offset is passed in via r0.5[31:10],
             //      but the BSS/SS descriptor expects the offset in [31:6] bits, thus we must shift it right by 4
-
-            // NOTE: This shr operation is generated by IGC in LSC path, where as in HDC path it is done by vISA
             VISA_GenVar* r0Var = nullptr;
             V(vKernel->GetPredefinedVar(r0Var, PREDEFINED_R0));
             VISA_VectorOpnd* surfOpnd = nullptr;

visa/BuildIRImpl.cpp

@@ -1867,41 +1867,26 @@ G4_INST* IR_Builder::createInternalBfnInst(
     return ii;
 }
 
-//scratch surfaces, write r0.5 to message descriptor
-//exdesc holds the value of the extended message descriptor for bit [0:11]
-// kernel entry:
-//      and (1) tmp<1>:ud r0.5<0;1,0>:ud 0xFFFFFC00:ud {NoMask}
-// before send message:
-//  shl (1) a0.0<1>:ud tmp<1>:ud 0x2 {NoMask}
-//  (for old exDesc format) add (1) a0.0<1>:ud tmp<1>:ud exDesc:ud {NoMask}
-// returns a0.0<0;1,0>:ud
+// scratch surfaces, write the content of T251 to extended message descriptor
+// exdesc holds the value of the extended message descriptor for bit [0:11]
+// add (1) a0.2<1>:ud T251<1>:ud exDesc:ud {NoMask}
+// returns a0.2<0;1,0>:ud
 G4_SrcRegRegion* IR_Builder::createScratchExDesc(uint32_t exdesc)
 {
+    // virtual var for each exdesc
+    G4_SrcRegRegion* T251 = createSrcRegRegion(builtinScratchSurface, getRegionScalar());
     const char* buf = getNameString(mem, 20, "ExDesc%d", num_temp_dcl++);
     G4_Declare* exDescDecl = createDeclareNoLookup(buf, G4_ADDRESS, 1, 1, Type_UD);
     exDescDecl->setSubRegAlign(Four_Word);
 
-    // copy r0.5[10:31] to a0[12:31] or a0[6:31] for the new format
-    initScratchSurfaceOffset();
-
-    if (!useNewExtDescFormat())
+    G4_DstRegRegion* dst = createDstRegRegion(exDescDecl, 1);
+    if (useNewExtDescFormat())
     {
-
-        // (W) shl (1) a0.0 sso 0x2
-        auto shlSrc0 = createSrcRegRegion(scratchSurfaceOffset, getRegionScalar());
-        auto shlDst = createDstRegRegion(exDescDecl, 1);
-        createBinOp(G4_shl, g4::SIMD1, shlDst, shlSrc0, createImm(0x2, Type_UW), InstOpt_WriteEnable, true);
-
-        G4_DstRegRegion* dst = createDstRegRegion(exDescDecl, 1);
-        createBinOp(G4_add, g4::SIMD1, dst, createSrcRegRegion(exDescDecl, getRegionScalar()),
-            createImm(exdesc, Type_UD), InstOpt_WriteEnable, true);
+        createMov(g4::SIMD1, dst, T251, InstOpt_WriteEnable, true);
     }
     else
     {
-        // (W) shr (1) a0.0 ss0 0x4
-        auto shrSrc0 = createSrcRegRegion(scratchSurfaceOffset, getRegionScalar());
-        auto shrDst = createDstRegRegion(exDescDecl, 1);
-        createBinOp(G4_shr, g4::SIMD1, shrDst, shrSrc0, createImm(0x4, Type_UW), InstOpt_WriteEnable, true);
+        createBinOp(G4_add, g4::SIMD1, dst, T251, createImm(exdesc, Type_UD), InstOpt_WriteEnable, true);
     }
     return createSrcRegRegion(exDescDecl, getRegionScalar());
 }