[RISCV][TTI] Add checks for invalid cast operations

[arcbbb]

In issue #88802, the LV cost model would query the cost of the TRUNC for source type 2xi1 and destination type 2xi32. This patch adds an early exit check to prevent invalid operations.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-backend-risc-v

Author: Shih-Po Hung (arcbbb)

Changes

In issue #88802, the LV cost model would query the cost of the TRUNC for source type 2xi1 and destination type 2xi32. This patch adds an early exit check to prevent invalid operations.

---

Full diff: https://github.com/llvm/llvm-project/pull/88854.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp (+10)

diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
index 38304ff90252f0..c4f1c275f63b65 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
@@ -956,6 +956,9 @@ InstructionCost RISCVTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
     return getRISCVInstructionCost(Op, DstLT.second, CostKind);
   }
   case ISD::TRUNCATE:
+    // Early return for invalid operation
+    if (Dst->getScalarSizeInBits() >= Src->getScalarSizeInBits())
+      break;
     if (Dst->getScalarSizeInBits() == 1) {
       // We do not use several vncvt to truncate to mask vector. So we could
       // not use PowDiff to calculate it.
@@ -968,6 +971,13 @@ InstructionCost RISCVTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
     [[fallthrough]];
   case ISD::FP_EXTEND:
   case ISD::FP_ROUND: {
+    // Early return for invalid operation
+    if ((ISD == ISD::FP_ROUND) &&
+        Dst->getScalarSizeInBits() >= Src->getScalarSizeInBits())
+      break;
+    if ((ISD == ISD::FP_EXTEND) &&
+        Src->getScalarSizeInBits() >= Dst->getScalarSizeInBits())
+      break;
     // Counts of narrow/widen instructions.
     unsigned SrcEltSize = Src->getScalarSizeInBits();
     unsigned DstEltSize = Dst->getScalarSizeInBits();

[arcbbb]

This fixes a crash in #88802.

[jacquesguan]

LGTM

[preames]

If I'm reading the code and discussion right, the LV is querying a truncate with result wider than source? If so, this does not seem like the right fix. If LV is querying an invalid argument combination, that's a LV bug and should be fixed as such.

[topperc]

Don't repeatedly call getScalarSizeInBits on the same object. Use a variable.

[arcbbb]

Updated. Thanks!

[topperc]

Test case?

[arcbbb]

Test case?

Added in the updates. Thanks!

[arcbbb]

If I'm reading the code and discussion right, the LV is querying a truncate with result wider than source? If so, this does not seem like the right fix. If LV is querying an invalid argument combination, that's a LV bug and should be fixed as such.

Yes, the actual root cause is LV cost model. I think for TTI cost it is better to return invalid cost or delegate it to the base TTI rather than crash on it,

[preames]

If I'm reading the code and discussion right, the LV is querying a truncate with result wider than source? If so, this does not seem like the right fix. If LV is querying an invalid argument combination, that's a LV bug and should be fixed as such.

Yes, the actual root cause is LV cost model. I think for TTI cost it is better to return invalid cost or delegate it to the base TTI rather than crash on it,

TTI should assert sanity of the API, and we should fix the caller. It's very likely there's a higher level semantics bug in the LV causing this, and we need to find and fix that one.

[arcbbb]

TTI should assert sanity of the API, and we should fix the caller. It's very likely there's a higher level semantics bug in the LV causing this, and we need to find and fix that one.

OK, I create #89161 for this.

[arcbbb]

close this since #88802 is no longer RISCV-specific and duplicates #47665.