[clang][RISCV] Extend intrinsic size check variable from 16 -> 32 bits. NFC

[4vtomat]

We currently have over 67000 intrinsics, uint16_t will overflow.

[llvmbot]

@llvm/pr-subscribers-clang

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

Author: Brandon Wu (4vtomat)

Changes

We currently have over 67000 intrinsics, uint16_t will overflow.

---

Full diff: https://github.com/llvm/llvm-project/pull/111481.diff

1 Files Affected:

• (modified) clang/lib/Sema/SemaRISCV.cpp (+1-1)

diff --git a/clang/lib/Sema/SemaRISCV.cpp b/clang/lib/Sema/SemaRISCV.cpp
index 3da4b515b1b114..f49c2088ef4776 100644
--- a/clang/lib/Sema/SemaRISCV.cpp
+++ b/clang/lib/Sema/SemaRISCV.cpp
@@ -399,7 +399,7 @@ void RISCVIntrinsicManagerImpl::InitRVVIntrinsic(
                                      Record.HasFRMRoundModeOp);
 
   // Put into IntrinsicList.
-  uint16_t Index = IntrinsicList.size();
+  uint32_t Index = IntrinsicList.size();
   assert(IntrinsicList.size() == (size_t)Index &&
          "Intrinsics indices overflow.");
   IntrinsicList.push_back({BuiltinName, Signature});

[topperc]

Why doesn't any test catch it?

[jerryzj]

Why doesn't any test catch it?

I suppose the CI of intrinsic doc should catch this, but it is not reporting errors as expected.
I happen to notice the failures and working on a fix now.

[wangpc-pp]

LGTM.

[topperc]

Why don’t any of our lit tests that use every intrinsic catch it?

[topperc]

What is the type of Intrinsics on this line further down

Intrinsics.insert({Name, Index});

is it also uint16_t?

[wangpc-pp]

Why don’t any of our lit tests that use every intrinsic catch it?

We don't see any error, is it because your downstream added some intrinsics?

[topperc]

Does the number intrinsics vary with the number of extensions enabled? Maybe we don't have a test with all extensions?

[4vtomat]

Why don’t any of our lit tests that use every intrinsic catch it?

Somehow if we add more "target-feature" the intrinsics increase, I'm not sure why.
For example:

test.c

#include "riscv_vector.h"
  vfloat16mf4x7_t test_vlseg7e16ff_v_f16mf4x7(const _Float16 *base, size_t *new_vl, size_t vl) {
  return __riscv_vlseg7e16ff_v_f16mf4x7(base, new_vl, vl);
}

Compile with this command:

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-abi lp64d test.c

generate 61674 intrinsics.
However if we add -target-feature +zvfbfmin

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-feature +zvfbfmin -target-abi lp64d test.c

generate 64568 intrinsics

it's weird.

[topperc]

Why don’t any of our lit tests that use every intrinsic catch it?

Somehow if we add more "target-feature" the intrinsics increase, I'm not sure why. For example:

test.c

#include "riscv_vector.h"
  vfloat16mf4x7_t test_vlseg7e16ff_v_f16mf4x7(const _Float16 *base, size_t *new_vl, size_t vl) {
  return __riscv_vlseg7e16ff_v_f16mf4x7(base, new_vl, vl);
}

Compile with this command:

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-abi lp64d test.c

generate 61674 intrinsics. However if we add -target-feature +zvfbfmin

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-feature +zvfbfmin -target-abi lp64d test.c

generate 64568 intrinsics

it's weird.

Why is it weird? Some intrinsics require specific features. If that feature isn't enabled, the intrinsic is skipped.

[4vtomat]

Why don’t any of our lit tests that use every intrinsic catch it?

Somehow if we add more "target-feature" the intrinsics increase, I'm not sure why. For example:

test.c

#include "riscv_vector.h"
  vfloat16mf4x7_t test_vlseg7e16ff_v_f16mf4x7(const _Float16 *base, size_t *new_vl, size_t vl) {
  return __riscv_vlseg7e16ff_v_f16mf4x7(base, new_vl, vl);
}

Compile with this command:

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-abi lp64d test.c

generate 61674 intrinsics. However if we add -target-feature +zvfbfmin

clang -cc1 -triple riscv64-- -S -target-feature +v -target-feature +zvfh -target-feature +zvfbfmin -target-abi lp64d test.c

generate 64568 intrinsics

it's weird.

Why is it weird? Some intrinsics require specific features. If that feature isn't enabled, the intrinsic is skipped.

Oh I see, but if we extend the size, the memory usage will grow.

[topperc]

I think we need to change

struct RVVOverloadIntrinsicDef {
  // Indexes of RISCVIntrinsicManagerImpl::IntrinsicList.
  SmallVector<uint16_t, 8> Indexes;
};

and

  // Mapping function name to index of IntrinsicList.
  StringMap<uint16_t> Intrinsics;

Those places were both changed by e0092ea in addition to the place you've currently changed.

[4vtomat]

I think we need to change

struct RVVOverloadIntrinsicDef {
  // Indexes of RISCVIntrinsicManagerImpl::IntrinsicList.
  SmallVector<uint16_t, 8> Indexes;
};

and

  // Mapping function name to index of IntrinsicList.
  StringMap<uint16_t> Intrinsics;

Those places were both changed by e0092ea in addition to the place you've currently changed.

Oh, nice catch! I didn't even remember I've changed this before lol~

[topperc]

LGTM

[goldsteinn]

Can this get merged?