Restrict IMAD slicing to newer platforms

The MadLoopSlice and AdvCodeMotion passes applied integer MAD
slicing too broadly, causing regressions on some platforms.
Narrow the platform guard so IMAD slicing only fires on
platforms that benefit from the optimization.

Author: agrabezh

IGC/Compiler/CISACodeGen/AdvCodeMotion.cpp

@@ -730,8 +730,8 @@ static bool isIMAD(const Instruction *I) {
 static bool isCandidateMAD(const Instruction *I, const CodeGenContext *CGC) {
   if (isDMAD(I))
     return true;
-  // Keep behavior consistent with MadLoopSlice: only touch IMAD on PVC+.
-  return (CGC && CGC->platform.isProductChildOf(IGFX_PVC) && isIMAD(I));
+  // Keep behavior consistent with MadLoopSlice: only touch IMAD on XE3P+.
+  return (CGC && CGC->platform.isCoreChildOf(IGFX_XE3P_CORE) && isIMAD(I));
 }
 
 static bool clusterByEquivalenceClasses(ArrayRef<Instruction *> Order, EquivalenceClasses<Instruction *> &ECs,
@@ -900,7 +900,7 @@ bool MadLoopSlice::sliceBlock(BasicBlock *BB) const {
     Run.clear();
   };
 
-  // Slice contiguous runs of MAD-like ops (DMAD or PVC+ IMAD) inside an acyclic BB.
+  // Slice contiguous runs of MAD-like ops (DMAD or XE3P+ IMAD) inside an acyclic BB.
   // This is less restrictive than MadLoopSlice (which requires an all-MAD loop body)
   // and matches real kernels that have setup code before/after the MAD region.
   for (Instruction &I : *BB) {

IGC/Compiler/tests/MadLoopSlice/imad.ll

@@ -8,7 +8,7 @@
 
 
 ; REQUIRES: llvm-14-plus
-; RUN: igc_opt --platformpvc --opaque-pointers -igc-madloopslice -S < %s | FileCheck %s
+; RUN: igc_opt --platformCri --opaque-pointers -igc-madloopslice -S < %s | FileCheck %s
 ; ------------------------------------------------
 ; MadLoopSlice (block slicing)
 ; ------------------------------------------------