Unnecessarily large constant created from reordering add and shift

[dzaima]

https://godbolt.org/z/xoKf6bnTb

The code:

#include<stdint.h>
#include<stdbool.h>
bool foo(uint64_t x) {
  uint16_t tag = x>>48;
  return tag>=0b1111111111110010 && tag<=0b1111111111110100;
}

with -O3 as of clang 19 (and still in trunk) compiles to:

foo:
        movabs  rax, 3940649673949184
        add     rax, rdi
        shr     rax, 48
        cmp     eax, 3
        setb    al
        ret

whereas 18.0 did this, which is strictly better (i.e. is the exact same set of instructions, just in a different order and without movabs):

foo:
        shr     rdi, 48
        add     edi, -65522
        cmp     edi, 3
        setb    al
        ret

[RKSimon]

trunk:

define i1 @foo(i64 %x) {
entry:
  %0 = add i64 %x, 3940649673949184
  %1 = icmp ult i64 %0, 844424930131968
  ret i1 %1
}

vs 17.x:

define i1 @foo(i64 %x) {
entry:
  %shr = lshr i64 %x, 48
  %conv = trunc i64 %shr to i32
  %0 = add nsw i32 %conv, -65522
  %1 = icmp ult i32 %0, 3
  ret i1 %1
}

IIRC this has been reported before

[dzaima]

I've previously reported similar (but unrelated to an extent I'd think) #62145 and #111323 in case those might be what you're thinking of; otherwise neither movabs shr cmp nor lshr add icmp trunc turn up anything

[RKSimon]

We lower to:

Optimized legalized selection DAG: %bb.0 'foo:entry'
SelectionDAG has 16 nodes:
  t0: ch,glue = EntryToken
              t2: i64,ch = CopyFromReg t0, Register:i64 %0
            t4: i64 = add t2, Constant:i64<3940649673949184>
          t16: i64 = srl t4, Constant:i8<48>
        t22: i32 = truncate t16
      t24: i32,i32 = X86ISD::SUB t22, Constant:i32<3>
    t26: i8 = X86ISD::SETCC TargetConstant:i8<2>, t24:1
  t11: ch,glue = CopyToReg t0, Register:i8 $al, t26
  t12: ch = X86ISD::RET_GLUE t11, TargetConstant:i32<0>, Register:i8 $al, t11:1

It should be possible to fold some (truncate (srl (add X, C1), C2)) patterns to (add (truncate (srl X, C2)), C1') - probably as a x86 specific fold initially.

• Confirm the alive2 pattern for the constraints on C1 and C2 for the (truncate (srl (add X, C1), C2)) -> (add (truncate (srl X, C2)), C1') fold
• Add x86 test coverage for some i64 srl+add cases
• Add x86 fold - either to combineTruncatedArithmetic (we have a TODO suggesting i64 handling) or it might have to be combineX86SetCC/combineSetCCEFLAGS.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: dzaima (dzaima)

https://godbolt.org/z/xoKf6bnTb

The code:

#include<stdint.h>
#include<stdbool.h>
bool foo(uint64_t x) {
  uint16_t tag = x>>48;
  return tag>=0b1111111111110010 && tag<=0b1111111111110100;
}

with -O3 as of clang 19 (and still in trunk) compiles to:

foo:
        movabs  rax, 3940649673949184
        add     rax, rdi
        shr     rax, 48
        cmp     eax, 3
        setb    al
        ret

whereas 18.0 did this, which is strictly better (i.e. is the exact same set of instructions, just in a different order and without movabs):

foo:
        shr     rdi, 48
        add     edi, -65522
        cmp     edi, 3
        setb    al
        ret

[llvmbot]

Hi!

This issue may be a good introductory issue for people new to working on LLVM. If you would like to work on this issue, your first steps are:

1. Check that no other contributor has already been assigned to this issue. If you believe that no one is actually working on it despite an assignment, ping the person. After one week without a response, the assignee may be changed.
2. In the comments of this issue, request for it to be assigned to you, or just create a pull request after following the steps below. Mention this issue in the description of the pull request.
3. Fix the issue locally.
4. Run the test suite locally. Remember that the subdirectories under test/ create fine-grained testing targets, so you can e.g. use make check-clang-ast to only run Clang's AST tests.
5. Create a Git commit.
6. Run git clang-format HEAD~1 to format your changes.
7. Open a pull request to the upstream repository on GitHub. Detailed instructions can be found in GitHub's documentation. Mention this issue in the description of the pull request.

If you have any further questions about this issue, don't hesitate to ask via a comment in the thread below.

[llvmbot]

@llvm/issue-subscribers-good-first-issue

Author: dzaima (dzaima)

https://godbolt.org/z/xoKf6bnTb

The code:

#include<stdint.h>
#include<stdbool.h>
bool foo(uint64_t x) {
  uint16_t tag = x>>48;
  return tag>=0b1111111111110010 && tag<=0b1111111111110100;
}

with -O3 as of clang 19 (and still in trunk) compiles to:

foo:
        movabs  rax, 3940649673949184
        add     rax, rdi
        shr     rax, 48
        cmp     eax, 3
        setb    al
        ret

whereas 18.0 did this, which is strictly better (i.e. is the exact same set of instructions, just in a different order and without movabs):

foo:
        shr     rdi, 48
        add     edi, -65522
        cmp     edi, 3
        setb    al
        ret

[souvik150]

Hey can I pick this up? Im new to the project and just looking for some beginner friendly issues

[RKSimon]

Sure. Are you familiar with Alive2? The first step is to determine the constraints on the truncation fold.

[souvik150]

Nope. Let me then start by getting familiar with Alive2. I'll spend today looking at that then, thanks!