Set extra cost for some math intrinsics

We need to set `TCC_Expensive` cost for some math intrinsics
to reduce the speculative execution of `sqrt`, `sin`, `cos`.

Author: agrabezh

IGC/Compiler/GenTTI.cpp

@@ -589,6 +589,20 @@ namespace llvm {
         , const User* U
 #endif
     ) {
+        // The extra cost of speculative execution for math intrinsics
+        if (auto *II = dyn_cast_or_null<IntrinsicInst>(U)) {
+            if (Intrinsic::ID IID = II->getIntrinsicID()) {
+                switch (IID) {
+                    case Intrinsic::cos:
+                    case Intrinsic::sin:
+                    case Intrinsic::sqrt:
+                        return TTI::TCC_Expensive;
+                    default:
+                        break;
+                }
+            }
+        }
+
         IGC::CodeGenContext* CGC = this->ctx;
         if (!CGC->enableFunctionCall() && !GenISAIntrinsic::isIntrinsic(F) &&
             !F->isIntrinsic()) {
@@ -621,6 +635,20 @@ namespace llvm {
     llvm::InstructionCost GenIntrinsicsTTIImpl::getUserCost(const User* U, ArrayRef<const Value*> Operands, TTI::TargetCostKind CostKind)
 #endif
     {
+        // The extra cost of speculative execution for math intrinsics
+        if (auto *II = dyn_cast_or_null<IntrinsicInst>(U)) {
+            if (Intrinsic::ID IID = II->getIntrinsicID()) {
+                switch (IID) {
+                    case Intrinsic::cos:
+                    case Intrinsic::sin:
+                    case Intrinsic::sqrt:
+                        return TTI::TCC_Expensive;
+                    default:
+                        break;
+                }
+            }
+        }
+
         const Function* F = dyn_cast<Function>(U);
         if (F != nullptr)
         {

IGC/Compiler/tests/SimplifyCFG/expensive_math_intrinsics.ll

@@ -0,0 +1,74 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2024 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+
+; REQUIRES: llvm-14-plus
+; RUN: igc_opt %s -S -o - -simplifycfg -gen-tti | FileCheck -check-prefix=CHECK %s
+
+; This test checks that SimplifyCFG do not merge nodes with expensive math intrinsics
+; such as sqrt, sin, cos
+
+define internal spir_func i8 @intersectWorldBoolean(i1 %cmp.i, float %sub7.i) {
+; CHECK-LABEL: @intersectWorldBoolean(
+; CHECK:       if.end.i:
+; CHECK-NEXT:    [[CALL_I_I_I2:%.*]] = call float @llvm.sqrt.f32
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  br label %for.body
+
+for.body:                                         ; preds = %for.cond
+  br i1 undef, label %if.then9.i.i, label %if.else11.i.i
+
+if.then9.i.i:                                     ; preds = %for.body
+  ret i8 0
+
+if.else11.i.i:                                    ; preds = %for.body
+  br label %_Z15__spirv_ocl_cosf.exit
+
+_Z15__spirv_ocl_cosf.exit:                        ; preds = %if.else11.i.i
+  br i1 %cmp.i, label %if.then.i, label %if.end.i
+
+if.then.i:                                        ; preds = %_Z15__spirv_ocl_cosf.exit
+  br label %intersectRaySphere.exit
+
+if.end.i:                                         ; preds = %_Z15__spirv_ocl_cosf.exit
+  %call.i.i.i2 = call float @llvm.sqrt.f32(float %sub7.i)
+  %i57 = fadd fast float 0.000000e+00, %call.i.i.i2
+  %div.i = fmul fast float %i57, 0.000000e+00
+  br label %intersectRaySphere.exit
+
+intersectRaySphere.exit:                          ; preds = %if.end.i, %if.then.i
+  %tRay.1 = phi float [ 0.000000e+00, %if.then.i ], [ %div.i, %if.end.i ]
+  br i1 undef, label %land.lhs.true, label %if.end
+
+land.lhs.true:                                    ; preds = %intersectRaySphere.exit
+  %cmp10 = fcmp olt float 0.000000e+00, 0.000000e+00
+  br i1 %cmp10, label %if.then, label %if.end
+
+if.then:                                          ; preds = %land.lhs.true
+  br i1 undef, label %if.end.i3, label %_Z21__spirv_ocl_normalizeDv4_f.exit
+
+if.end.i3:                                        ; preds = %if.then
+  ret i8 0
+
+_Z21__spirv_ocl_normalizeDv4_f.exit:              ; preds = %if.then
+  %i98 = insertelement <4 x float> zeroinitializer, float 0.000000e+00, i32 0
+  br label %if.end
+
+if.end:                                           ; preds = %_Z21__spirv_ocl_normalizeDv4_f.exit, %land.lhs.true, %intersectRaySphere.exit
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  br label %for.cond
+}
+
+; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
+declare float @llvm.sqrt.f32(float) #0
+
+attributes #0 = { nofree nosync nounwind readnone speculatable willreturn }