RFC: SROA for method argument


I am working to improve SROA to generate better code when a method has a struct in its arguments. I would appreciate it if I could have any suggestions or comments on how I can best proceed with this optimization.

  • Problem *
    I observed that LLVM often generates redundant instructions around glibc’s istreambuf_iterator. The problem comes from the scalar replacement (SROA) for methods with an aggregate as an argument. Here is a simplified example in C.

struct record {
long long a;
int b;
int c;

int func(struct record r) {
for (int i = 0; i < r.c; i++)
return r.b;

When updating r.b (or r.c as well), SROA generates redundant instructions on some platforms (such as x86_64 and ppc64); here, r.b and r.c are packed into one 64-bit GPR when the struct is passed as a method argument. The problem is caused when the same memory location is accessed by load/store instructions of different types.
For this example, CLANG generates following IRs to initialize the struct for ppc64 and x86_64. For both platforms, the 64-bit value is stored into memory allocated by alloca first. Later, the same memory location is accessed as 32-bit integer values (r.b and r.c).

for ppc64
%struct.record = type { i64, i32, i32 }

define signext i32 @ppc64le_func([2 x i64] %r.coerce) #0 {
%r = alloca %struct.record, align 8
%0 = bitcast %struct.record* %r to [2 x i64]*
store [2 x i64] %r.coerce, [2 x i64]* %0, align 8

for x86_64
define i32 @x86_64_func(i64 %r.coerce0, i64 %r.coerce1) #0 {
%r = alloca %struct.record, align 8
%0 = bitcast %struct.record* %r to { i64, i64 }*
%1 = getelementptr inbounds { i64, i64 }, { i64, i64 }* %0, i32 0, i32 0
store i64 %r.coerce0, i64* %1, align 8
%2 = getelementptr inbounds { i64, i64 }, { i64, i64 }* %0, i32 0, i32 1
store i64 %r.coerce1, i64* %2, align 8

For such code sequence, the current SROA generates instructions to update only upper (or lower) half of the 64-bit value when storing r.b (or r.c). SROA can split an i64 value into two i32 values under some conditions (e.g. when the struct contains only int b and int c in this example), but it is not capable of splitting complex cases.

  • Approach *
    In SROA pass, AggLoadStoreRewriter splits a load or store instructions for an aggregate into multiple load or store instructions for simple values. In above ppc64 case, store [2 x i64] is splitted into two store for i64.
    I am extending AggLoadStoreRewriter to split a store of an aggregate that comes from a method argument based on the format of the aggregate (Here, { i64, i32, i32 } instead of [2 x i64]). I have submitted a work-in-progress patch in Phabricator ( https://reviews.llvm.org/D32998 ). This optimization depends on the ABI, so I enabled this only for ppc64 with ELFv2 ABI so far.

I truly appreciate any advices and comments.
Best regards,

When there are accesses of mixed type to an alloca, SROA just treats the whole alloca as a big integer, and generates PHI nodes appropriately. In many cases, instcombine would then slice up the generated PHI nodes to use more appropriate types, but that doesn't work out here. (See InstCombiner::SliceUpIllegalIntegerPHI.) Probably the right solution is to make instcombine more aggressive here; it's hard to come up with a generally useful transform in SROA without reasoning about control flow.


Thank you for the advice.
InstCombiner::SliceUpIllegalIntegerPHI splits only when the output of the PHI node is only used by trunc or lshr+trunc (and the integer size is not legal).
So it cannot not handle update to the field. If we extend SliceUpIllegalIntegerPHI to support more generic cases, I feel it may become similar to the split functionality in SROA.

Extending the split functionality (in SROA or InstCombiner) will give higher coverage but it requires complicated (and potentially costly) analysis.
My approach of splitting arguments in SROA preprocessing aim to cover a common case without such complicated analysis.
Which approach is more preferable/acceptable for the community?

Here, I attach the IRs generated by SROA for the example in my first mail.
The output of PHI (%r.sroa.2.0) is used by the and instruction.

define signext i32 @_Z4func6record([2 x i64] %r.coerce) #0 {
%r.coerce.fca.0.extract = extractvalue [2 x i64] %r.coerce, 0
%r.coerce.fca.1.extract = extractvalue [2 x i64] %r.coerce, 1
br label %for.cond

for.cond: ; preds = %for.body, %entry
%r.sroa.2.0 = phi i64 [ %r.coerce.fca.1.extract, %entry ], [ %r.sroa.2.8.insert.insert, %for.body ]
%i.0 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]
%conv = sext i32 %i.0 to i64
%cmp = icmp slt i64 %conv, %r.coerce.fca.0.extract
br i1 %cmp, label %for.body, label %for.cond.cleanup

for.cond.cleanup: ; preds = %for.cond
%r.sroa.2.8.extract.trunc6 = trunc i64 %r.sroa.2.0 to i32
ret i32 %r.sroa.2.8.extract.trunc6

for.body: ; preds = %for.cond
%r.sroa.2.8.extract.trunc = trunc i64 %r.sroa.2.0 to i32
%inc = add nsw i32 %r.sroa.2.8.extract.trunc, 1
%r.sroa.2.8.insert.ext = zext i32 %inc to i64
%r.sroa.2.8.insert.mask = and i64 %r.sroa.2.0, -4294967296
%r.sroa.2.8.insert.insert = or i64 %r.sroa.2.8.insert.mask, %r.sroa.2.8.insert.ext
%inc1 = add nsw i32 %i.0, 1
br label %for.cond

I'll propose a different heuristic. SROA should ignore stores of arguments
into allocas in the entry block when deciding what slices to form. Such
stores happen exactly once, and are usually coercions that we have to do
for ABI reasons. SROA should generate code like this before promoting
allocas to SSA form:

define i32 @func(i64 %r.coerce.0, i64 %r.coerce.1) {
  %r.slice.0 = alloca i64
  %r.slice.1 = alloca i32
  %r.slice.2 = alloca i32
  store i64 %r.coerce.0, i64* %r.slice.0
  %r.1.shr = lshr i64 %r.coerce.1, 32
  %r.1 = trunc i64 %r.1.shr
  %r.2 = trunc i64 %r.coerce.1
  store i32 %r.1, i32* %r.slice.1
  store i32 %r.2, i32* %r.slice.2

This is basically "reasoning about the CFG" without actually looking at
loop info. Stores of arguments in the entry block can't be in a loop. Even
if they end up in one after inlining, instcombine should be able to
simplify the {i32,i32}->i64->{i32,i32} code.

I agree with Reid’s suggestion. We can ignore the first stores of arguments into allocas in the SROA analysis.

Orthogonal to Reid’s suggestion, I implemented another approach to resolve the problem.
This approach is more generic (not limited to arguments), simpler and ABI independent.

Currently, SROA splits loads and stores only when they are accessing the whole alloca.
I relax this limitation to split a load/store if all other loads and stores to the alloca are
disjoint to or fully included in the current load/store.
The whole-alloca loads and stores meet this new condition and so they are still splittable.
Does this approach make sense?

I updated the patch in Phabricator. https://reviews.llvm.org/D32998

Since the new approach is ABI independent, I confirmed that it now works on both x86 and ppc.
With this SROA optimization, the sample program is compiled into only few instructions without loop.
Without the optimization, unnecessary loop-carried dependency is introduced by SROA and
the loop cannot be eliminated by the later optimizers.