LLVM Branch Weight Metadata

Introduction

Branch Weight Metadata represents branch weights as its likeliness to be taken (see LLVM Block Frequency Terminology). Metadata is assigned to an Instruction that is a terminator as a MDNode of the MD_prof kind. The first operator is always a MDString node with the string “branch_weights”. Number of operators depends on the terminator type.

Branch weights might be fetch from the profiling file, or generated based on __builtin_expect and __builtin_expect_with_probability instruction.

All weights are represented as an unsigned 32-bit values, where higher value indicates greater chance to be taken.

Supported Instructions

BranchInst

Metadata is only assigned to the conditional branches. There are two extra operands for the true and the false branch. We optionally track if the metadata was added by __builtin_expect or __builtin_expect_with_probability with an optional field !"expected".

!0 = !{
  !"branch_weights",
  [ !"expected", ]
  i32 <TRUE_BRANCH_WEIGHT>,
  i32 <FALSE_BRANCH_WEIGHT>
}

SwitchInst

Branch weights are assigned to every case (including the default case which is always case #0).

!0 = !{
  !"branch_weights",
  [ !"expected", ]
  i32 <DEFAULT_BRANCH_WEIGHT>
  [ , i32 <CASE_BRANCH_WEIGHT> ... ]
}

IndirectBrInst

Branch weights are assigned to every destination.

!0 = !{
  !"branch_weights",
  [ !"expected", ]
  i32 <LABEL_BRANCH_WEIGHT>
  [ , i32 <LABEL_BRANCH_WEIGHT> ... ]
}

CallInst

Calls may have branch weight metadata, containing the execution count of the call. It is currently used in SamplePGO mode only, to augment the block and entry counts which may not be accurate with sampling.

!0 = !{
  !"branch_weights",
  [ !"expected", ]
  i32 <CALL_BRANCH_WEIGHT>
}

InvokeInst

Invoke instruction may have branch weight metadata with one or two weights. The second weight is optional and corresponds to the unwind branch. If only one weight is set then it contains the execution count of the call and used in SamplePGO mode only as described for the call instruction. If both weights are specified then the second weight contains count of unwind branch taken and the first weights contains the execution count of the call minus the count of unwind branch taken. Both weights specified are used to calculate BranchProbability as for BranchInst and for SamplePGO the sum of both weights is used.

!0 = !{
  !"branch_weights",
  [ !"expected", ]
  i32 <INVOKE_NORMAL_WEIGHT>
  [ , i32 <INVOKE_UNWIND_WEIGHT> ]
}

Other

Other terminator instructions are not allowed to contain Branch Weight Metadata.

Built-in expect Instructions

__builtin_expect(long exp, long c) instruction provides branch prediction information. The return value is the value of exp.

It is especially useful in conditional statements. Currently Clang supports two conditional statements:

if statement

The exp parameter is the condition. The c parameter is the expected comparison value. If it is equal to 1 (true), the condition is likely to be true, in other case condition is likely to be false. For example:

if (__builtin_expect(x > 0, 1)) {
  // This block is likely to be taken.
}

switch statement

The exp parameter is the value. The c parameter is the expected value. If the expected value doesn’t show on the cases list, the default case is assumed to be likely taken.

switch (__builtin_expect(x, 5)) {
default: break;
case 0:  // ...
case 3:  // ...
case 5:  // This case is likely to be taken.
}

Built-in expect.with.probability Instruction

__builtin_expect_with_probability(long exp, long c, double probability) has the same semantics as __builtin_expect, but the caller provides the probability that exp == c. The last argument probability must be constant floating-point expression and be in the range [0.0, 1.0] inclusive. The usage is also similar as __builtin_expect, for example:

if statement

If the expect comparison value c is equal to 1(true), and probability value probability is set to 0.8, that means the probability of condition to be true is 80% while that of false is 20%.

if (__builtin_expect_with_probability(x > 0, 1, 0.8)) {
  // This block is likely to be taken with probability 80%.
}

switch statement

This is basically the same as switch statement in __builtin_expect. The probability that exp is equal to the expect value is given in the third argument probability, while the probability of other value is the average of remaining probability(1.0 - probability). For example:

switch (__builtin_expect_with_probability(x, 5, 0.7)) {
default: break;  // Take this case with probability 10%
case 0:  break;  // Take this case with probability 10%
case 3:  break;  // Take this case with probability 10%
case 5:  break;  // This case is likely to be taken with probability 70%
}

CFG Modifications

Branch Weight Metatada is not proof against CFG changes. If terminator operands’ are changed some action should be taken. In other case some misoptimizations may occur due to incorrect branch prediction information.

Function Entry Counts

To allow comparing different functions during inter-procedural analysis and optimization, MD_prof nodes can also be assigned to a function definition. The first operand is a string indicating the name of the associated counter.

Currently, one counter is supported: “function_entry_count”. The second operand is a 64-bit counter that indicates the number of times that this function was invoked (in the case of instrumentation-based profiles). In the case of sampling-based profiles, this operand is an approximation of how many times the function was invoked.

For example, in the code below, the instrumentation for function foo() indicates that it was called 2,590 times at runtime.

define i32 @foo() !prof !1 {
  ret i32 0
}
!1 = !{!"function_entry_count", i64 2590}

If “function_entry_count” has more than 2 operands, the later operands are the GUID of the functions that needs to be imported by ThinLTO. This is only set by sampling based profile. It is needed because the sampling based profile was collected on a binary that had already imported and inlined these functions, and we need to ensure the IR matches in the ThinLTO backends for profile annotation. The reason why we cannot annotate this on the callsite is that it can only goes down 1 level in the call chain. For the cases where foo_in_a_cc()->bar_in_b_cc()->baz_in_c_cc(), we will need to go down 2 levels in the call chain to import both bar_in_b_cc and baz_in_c_cc.