CustomLoopOpt fp type mismatch fix

CustomLoopOpt attempts to create Fcmp inst with first operand being a
floating point type and second operand being unconditionally f32. This
can cause a type mismatch as first operand is not necessarily f32.

Author: smilczek

IGC/Compiler/CustomLoopOpt.cpp

@@ -586,7 +586,7 @@ bool CustomLoopVersioning::processLoop(Loop* loop)
     FMF.setFast();
     irb.setFastMathFlags(FMF);
     Value* cond0 = irb.CreateFCmpOGT(
-        var_CBLoad_preHdr, ConstantFP::get(irb.getFloatTy(), 1.0));
+        var_CBLoad_preHdr, ConstantFP::get(var_CBLoad_preHdr->getType(), 1.0));
 
     Value* cond1 = irb.CreateFCmpOLT(
         irb.CreateFMul(var_range_x, var_CBLoad_preHdr),

IGC/Compiler/tests/CustomLoopVersioning/double-fp-type.ll

@@ -0,0 +1,54 @@
+;=========================== begin_copyright_notice ============================
+;
+; Copyright (C) 2024 Intel Corporation
+;
+; SPDX-License-Identifier: MIT
+;
+;============================ end_copyright_notice =============================
+;
+; REQUIRES: llvm-14-plus
+; RUN: igc_opt --opaque-pointers -debugify -igc-custom-loop-opt -S < %s 2>&1 | FileCheck %s
+; ------------------------------------------------
+; CustomLoopVersioning
+; ------------------------------------------------
+
+; This test checks that we're not creating a cmp inst with mismatched fp types.
+
+define spir_kernel void @test_customloop(ptr addrspace(65549) %a, double %b, double %c, double %d) {
+entry:
+  %aa = inttoptr i32 42 to ptr addrspace(65549)
+  %cc = call double @llvm.maxnum.f64(double %c, double 3.000000e+00)
+  %dd = call double @llvm.minnum.f64(double %d, double 2.000000e+00)
+  br label %pre_header
+
+pre_header:
+  ; CHECK: [[TMP:%.*]] = load double
+  %0 = load double, ptr addrspace(65549) %aa, align 4
+  %1 = fmul double %b, %0
+  ; CHECK: fcmp fast ogt double [[TMP]], 1.000000e+00
+  br label %loop_body
+
+loop_body:
+  %2 = phi double [ %b, %pre_header ], [ %3, %loop_body ]
+  %3 = phi double [ %1, %pre_header ], [ %7, %loop_body ]
+  %4 = call double @llvm.maxnum.f64(double %cc, double %2)
+  %5 = call double @llvm.minnum.f64(double %dd, double %3)
+  %6 = load double, ptr addrspace(65549) %aa, align 4
+  %7 = fmul double %3, %6
+  %8 = fcmp ult double %3, %dd
+  br i1 %8, label %loop_body, label %end
+
+end:
+  store double %3, ptr addrspace(65549) %a, align 4
+  ret void
+}
+
+
+declare double @llvm.maxnum.f64(double, double) #0
+declare double @llvm.minnum.f64(double, double) #0
+
+!igc.functions = !{!0}
+
+!0 = !{ptr @test_customloop, !1}
+!1 = !{!2}
+!2 = !{!"function_type", i32 0}