AMDGPU: Avoid introducing illegal fminnum_ieee/fmaxnum_ieee#177418
Merged
AMDGPU: Avoid introducing illegal fminnum_ieee/fmaxnum_ieee#177418
Conversation
Keep bf16/f16 values encoded as the low half of a 32-bit register, instead of promoting to float. This avoids unwanted FP effects from the fpext/fptrunc which should not be implied by just passing an argument. This also fixes ABI divergence between SelectionDAG and GlobalISel. I've wanted to make this change for ages, and failed the last few times. The main complication was the hack to return shader integer types in SGPRs, which now needs to inspect the underlying IR type.
This prevents a regression in a future change.
Avoid introducing fminnum_ieee/fmaxnum_ieee on f16 if f16 is not legal. This avoids regressing minimum/maximum cases in a future commit.
This was referenced Jan 22, 2026
Contributor
Author
This was referenced Jan 22, 2026
Member
|
@llvm/pr-subscribers-backend-amdgpu @llvm/pr-subscribers-llvm-globalisel Author: Matt Arsenault (arsenm) ChangesAvoid introducing fminnum_ieee/fmaxnum_ieee on f16 if f16 Full diff: https://github.com/llvm/llvm-project/pull/177418.diff 1 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 0ba66af9bd41c..69d73975bf9ef 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15273,8 +15273,9 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
// for some types, but at a higher cost since it's implemented with a 3
// operand form.
const SDNodeFlags Flags = N->getFlags();
- if ((Opc == ISD::FMINIMUM || Opc == ISD::FMAXIMUM) &&
- !Subtarget->hasIEEEMinimumMaximumInsts() && Flags.hasNoNaNs()) {
+ if ((Opc == ISD::FMINIMUM || Opc == ISD::FMAXIMUM) && Flags.hasNoNaNs() &&
+ !Subtarget->hasIEEEMinimumMaximumInsts() &&
+ isOperationLegal(ISD::FMINNUM_IEEE, VT.getScalarType())) {
unsigned NewOpc =
Opc == ISD::FMINIMUM ? ISD::FMINNUM_IEEE : ISD::FMAXNUM_IEEE;
return DAG.getNode(NewOpc, SDLoc(N), VT, Op0, Op1, Flags);
|
Base automatically changed from
users/arsenm/amdgpu/expand-strict-fp16-to-fp
to
main
January 22, 2026 19:16
rampitec
reviewed
Jan 22, 2026
| class CCIfNotInReg<CCAction A> : CCIf<"!ArgFlags.isInReg()", A> {} | ||
| class CCIfExtend<CCAction A> | ||
| : CCIf<"ArgFlags.isSExt() || ArgFlags.isZExt()", A>; | ||
| class CCIfOrigTypeShaderCCIsSGPR<CCAction A> |
Collaborator
There was a problem hiding this comment.
It does not look like a part of this change.
Contributor
Author
There was a problem hiding this comment.
If you don't rebase stacked PRs, it looks like all the previous patches from the already merged branches are here. you have to look at the top commit only
Collaborator
There was a problem hiding this comment.
I see, but top commit does not seem to have tests.
Contributor
Author
There was a problem hiding this comment.
It triggers in the existing minimum tests after #177419, I'm not sure how to trick the legalization steps into getting here until then
rampitec
approved these changes
Jan 22, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Avoid introducing fminnum_ieee/fmaxnum_ieee on f16 if f16
is not legal. This avoids regressing minimum/maximum cases
in a future commit.