Enable Dot and Add Ops to Dp4a Ops Optimization

Enable Dot and Add Ops to Dp4a Ops Optimization.

Author: ichenkai

IGC/Compiler/CustomSafeOptPass.cpp

@@ -378,7 +378,7 @@ void CustomSafeOptPass::visitShuffleIndex(llvm::CallInst* I)
 // %id317- = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 %id213-, i32 %267, i32 %321)
 void CustomSafeOptPass::mergeDotAddToDp4a(llvm::CallInst* I)
 {
-    if (!IGC_IS_FLAG_ENABLED(EnableDotAddToDp4aMerge))
+    if (IGC_IS_FLAG_ENABLED(DisableDotAddToDp4aMerge))
       return;
 
     // found %id213- = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 0, i32 %305, i32 %345)

IGC/Compiler/tests/CustomSafeOptPass/DotAddToDp4a.ll

@@ -5,15 +5,16 @@
 ; SPDX-License-Identifier: MIT
 ;
 ;============================ end_copyright_notice =============================
-; REQUIRES: regkeys
 ;
-; RUN: igc_opt --platformdg2 --regkey EnableDotAddToDp4aMerge=1 -igc-custom-safe-opt -S < %s --dce | FileCheck %s
+; RUN: igc_opt --platformdg2 -igc-custom-safe-opt -S < %s --dce | FileCheck %s
 ; ------------------------------------------------
 ; CustomSafeOptPass: DotAddToDp4a
 ; ------------------------------------------------
 define void @test_DotAddToDp4a(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f) {
 ; CHECK-LABEL: @test_DotAddToDp4a(
 ; CHECK:    [[TMP1:%.*]] = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 0, i32 %a, i32 %b)
+; CHECK:    [[TMP2:%.*]] = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 [[TMP1]], i32 %c, i32 %d)
+; CHECK:    [[TMP3:%.*]] = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 [[TMP2]], i32 %e, i32 %f)
 ;
   %1 = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 0, i32 %a, i32 %b)
   %2 = call i32 @llvm.genx.GenISA.dp4a.ss.i32(i32 0, i32 %c, i32 %d)

IGC/common/igc_flags.h

@@ -289,7 +289,7 @@ DECLARE_IGC_REGKEY(bool, EnableGEPLSR, true, "Enables GEP Loop Strength Reductio
 DECLARE_IGC_REGKEY(DWORD, GEPLSRThresholdRatio, 100, "Ratio for register pressure threshold in GEP Loop Strength Reduction pass", false)
 DECLARE_IGC_REGKEY(bool, EnableGEPLSRToPreheader, true, "Enables reduction to loop's preheader in GEP Loop Strength Reduction pass", false)
 DECLARE_IGC_REGKEY(DWORD, FPRoundingModeCoalescingMaxDistance, 20, "Max distance in instructions for reordering FP instructions with common rounding mode", false)
-DECLARE_IGC_REGKEY(bool, EnableDotAddToDp4aMerge, false, "Enable Dot and Add ops to Dp4a merge optimization.", false)
+DECLARE_IGC_REGKEY(bool, DisableDotAddToDp4aMerge, false, "Disable Dot and Add ops to Dp4a merge optimization.", false)
 DECLARE_IGC_REGKEY(bool, DisableInvalidateRTStackAfterLastRead, false, "Disables L1 cache invalidation after the last read of the RT stack. Affects rayqueries only", false)
 
 DECLARE_IGC_GROUP("Shader debugging")