Increase number of icmps needed for converting to a switch statement

[hiraditya]

In RISC-V the minimum number of case needed for code-generating a jump table is 5, while on AArch64 it is 13. Ideally we should query the backend for their preferable jump table size before deciding to convert a sequence of compares to a switch statement.

Also make this a parameter for anyone to tune.

Fixes: #48188

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: AdityaK (hiraditya)

Changes

In RISC-V the minimum number of case needed for code-generating a jump table is 5, while on AArch64 it is 13. Ideally we should query the backend for their preferable jump table size before deciding to convert a sequence of compares to a switch statement.

Also make this a parameter for anyone to tune.

Fixes: #48188

---

Full diff: https://github.com/llvm/llvm-project/pull/99269.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Utils/SimplifyCFG.cpp (+6-2)

diff --git a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
index 8f717cb43bcb4..e85fd5de3d4f7 100644
--- a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
@@ -180,6 +180,10 @@ static cl::opt<unsigned> MaxSwitchCasesPerResult(
     "max-switch-cases-per-result", cl::Hidden, cl::init(16),
     cl::desc("Limit cases to analyze when converting a switch to select"));
 
+static cl::opt<unsigned> MinICmpsToSwitch(
+    "min-icmps-to-switch", cl::Hidden, cl::init(9),
+    cl::desc("Minimum number of icmp sequences to convert to switch"));
+
 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
 STATISTIC(NumLinearMaps,
           "Number of switch instructions turned into linear mapping");
@@ -4893,8 +4897,8 @@ bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
   if (!CompVal)
     return false;
 
-  // Avoid turning single icmps into a switch.
-  if (UsedICmps <= 1)
+  // Avoid turning few icmps into a switch.
+  if (UsedICmps < MinICmpsToSwitch)
     return false;
 
   bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));

[nikic]

Converting icmp sequences to switch is part of the canonical form, and many other optimizations depend on it.

[hiraditya]

As vectorization gets blocked in the cases noted in #48188, how should we go about addressing that issue?

[hiraditya]

#99808 is a better approach.