[X86] Add support for X86ISD float2int constant folding

[RKSimon]

Bugzilla Link	37798
Version	trunk
OS	Windows NT
Depends On	#37099
CC	@adibiagio,@topperc,@filcab,@gregbedwell,@LebedevRI,@rotateright

Extended Description

Split off from [Bug #​37751].

Now that we've standardised to using X86ISD opcodes for float2int conversions:

  // Vector float/double to signed/unsigned integer.
  CVTP2SI, CVTP2UI, CVTP2SI_RND, CVTP2UI_RND,
  // Scalar float/double to signed/unsigned integer.
  CVTS2SI_RND, CVTS2UI_RND,

  // Vector float/double to signed/unsigned integer with truncation.
  CVTTP2SI, CVTTP2UI, CVTTP2SI_RND, CVTTP2UI_RND,
  // Scalar float/double to signed/unsigned integer with truncation.
  CVTTS2SI_RND, CVTTS2UI_RND,

We should investigate adding safe constant folding support to these. We could at least support exact conversions (these should be rounding mode independent for non-truncation instructions?).

For truncation instructions we can probably accept inexact conversions as well, as long as no underflow/overflow occurs - this would need to be tested to confirm.

[rotateright]

Any preference about whether this happens in IR or DAG?

There's a substantial chunk of target-specific intrinsic folding in InstCombineCalls, and that's probably easier to code with the help of match().

A quick look over X86ISelLowering seems like there's not much chance of creating the DAG nodes with constants (we should have already constant folded if they're being created from generic casting nodes.

[RKSimon]

InstCombineCalls probably makes more sense - I don't think DAGCombine generates X86ISD nodes from anything that wouldn't already have constant folded if it could?

[rotateright]

InstCombineCalls probably makes more sense - I don't think DAGCombine
generates X86ISD nodes from anything that wouldn't already have constant
folded if it could?

Agreed - another blob of instcombine.

IIRC, there were complaints about this a while back. We really should move all of the target-specific intrinsic folding to its own pass.

[llvmbot]

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

[Samhitha1212]

Hi, my team would like to work on this

[RKSimon]

Hi, my team would like to work on this

Cheers - are you interested in handling this for constexpr cases and/or in the middle IR?

[Samhitha1212]

First we will work on middle IR.

…

On Sat, Nov 1, 2025 at 5:49 PM Simon Pilgrim ***@***.***> wrote: Assigned #37146 <#37146> to @Samhitha1212 <https://github.com/Samhitha1212>. — Reply to this email directly, view it on GitHub <#37146 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BCTKVSVEGVGNZFAEFU7DSBL32SQLTAVCNFSM6AAAAACK25ZLX2VHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMRQGY2TCOBXG4YTCMI> . You are receiving this because you were assigned.Message ID: ***@***.***>

-- Disclaimer:- This footer text is to convey that this email is sent by one of the users of IITH. So, do not mark it as SPAM.