[AMDGPU][SplitModule] Fix unintentional integer division (llvm#117586)

A static analysis tool warned that a division was always being performed
in integer division, so was either 0.0 or 1.0.

This doesn't seem intentional, so has been fixed to return a true ratio
using floating-point division. This in turn showed a bug where a
comparison against this ratio was incorrect.

Change-Id: Ia8b755339a7ddc2d3bdf5ba701bd5724b00488d2

Author: frasercrmck

llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp

@@ -1101,10 +1101,10 @@ void RecursiveSearchSplitting::pickPartition(unsigned Depth, unsigned Idx,
       if (Entry.CostExcludingGraphEntryPoints > LargeClusterThreshold) {
         // Check if the amount of code in common makes it worth it.
         assert(SimilarDepsCost && Entry.CostExcludingGraphEntryPoints);
-        const double Ratio =
-            SimilarDepsCost / Entry.CostExcludingGraphEntryPoints;
+        const double Ratio = static_cast<double>(SimilarDepsCost) /
+                             Entry.CostExcludingGraphEntryPoints;
         assert(Ratio >= 0.0 && Ratio <= 1.0);
-        if (LargeFnOverlapForMerge > Ratio) {
+        if (Ratio > LargeFnOverlapForMerge) {
           // For debug, just print "L", so we'll see "L3=P3" for instance, which
           // will mean we reached max depth and chose P3 based on this
           // heuristic.