Skip to content

[DAGCombiner] Optimize more vector element extractions. #80520

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

[DAGCombiner] Optimize more vector element extractions. #80520

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 3 additions & 4 deletions llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22240,9 +22240,8 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
}

// extract_vector_elt (build_vector x, y), 1 -> y
if (((IndexC && VecOp.getOpcode() == ISD::BUILD_VECTOR) ||
VecOp.getOpcode() == ISD::SPLAT_VECTOR) &&
TLI.isTypeLegal(VecVT)) {
if ((IndexC && VecOp.getOpcode() == ISD::BUILD_VECTOR) ||
VecOp.getOpcode() == ISD::SPLAT_VECTOR) {
assert((VecOp.getOpcode() != ISD::BUILD_VECTOR ||
VecVT.isFixedLengthVector()) &&
"BUILD_VECTOR used for scalable vectors");
Expand All @@ -22252,7 +22251,7 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
EVT InEltVT = Elt.getValueType();

if (VecOp.hasOneUse() || TLI.aggressivelyPreferBuildVectorSources(VecVT) ||
isNullConstant(Elt)) {
isIntOrFPConstant(Elt)) {
// Sometimes build_vector's scalar input types do not match result type.
if (ScalarVT == InEltVT)
return Elt;
Expand Down
171 changes: 1 addition & 170 deletions llvm/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,10 +190,6 @@ define <2 x half> @v_test_canonicalize_build_vector_v2f16(half %lo, half %hi) #1
; CI-LABEL: v_test_canonicalize_build_vector_v2f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: s_setpc_b64 s[30:31]
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What has happened here?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As far as I can see, we were previously converting the f16 parameters to f32 and straight back to f16. There is no reason for doing that, so I think it is a good thing that this is getting removed now.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But wasn't it part of the FP canonicalization?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh, I see... It wasn't directly, I think, the fcanonicalize was getting compiled to a no op, but yeah, that may have been because other operations were guaranteed to result in an already-canonical result. I'll take a closer look.

Copy link
Contributor Author

@hvdijk hvdijk Feb 4, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah. DAGCombiner's getCopyFromParts creates a FP_ROUND with a TRUNC value of 1, documented as "TRUNC is a flag, which is always an integer that is zero or one. If TRUNC is 0, this is a normal rounding, if it is 1, this FP_ROUND is known to not change the value of Y. The TRUNC = 1 case is used in cases where we know that the value will not be modified by the node, because Y is not using any of the extra precision of source type." But this looks wrong: even if Y is known not to be using any of the extra precision of the source type, FP_ROUND may still change the value by turning a sNaN into a qNaN. I suspect this is a pre-existing issue which we can also find breakage for without my PR? My PR certainly makes it worse though.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not quieting signaling nans is permitted for non-strict functions. I would just ignore this, I think the forced promotion of half to float in the ABI is nonsense I gave up on fixing

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh, in this context this isn't ignorable. The canonicalize shouldn't be disappearing

;
; GFX11-LABEL: v_test_canonicalize_build_vector_v2f16:
Expand Down Expand Up @@ -2301,12 +2297,6 @@ define <3 x half> @v_test_canonicalize_var_v3f16(<3 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v3f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v3f16:
Expand Down Expand Up @@ -2341,14 +2331,6 @@ define <4 x half> @v_test_canonicalize_var_v4f16(<4 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v4f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v4f16:
Expand Down Expand Up @@ -2611,9 +2593,7 @@ define <2 x half> @v_test_canonicalize_reg_k_v2f16(half %val) #1 {
; CI-LABEL: v_test_canonicalize_reg_k_v2f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_mov_b32_e32 v1, 2.0
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_reg_k_v2f16:
Expand Down Expand Up @@ -2647,8 +2627,7 @@ define <2 x half> @v_test_canonicalize_k_reg_v2f16(half %val) #1 {
; CI-LABEL: v_test_canonicalize_k_reg_v2f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v0
; CI-NEXT: v_mov_b32_e32 v1, v0
; CI-NEXT: v_mov_b32_e32 v0, 2.0
; CI-NEXT: s_setpc_b64 s[30:31]
;
Expand Down Expand Up @@ -2878,18 +2857,6 @@ define <6 x half> @v_test_canonicalize_var_v6f16(<6 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v6f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
; CI-NEXT: v_cvt_f16_f32_e32 v5, v5
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v4, v4
; CI-NEXT: v_cvt_f32_f16_e32 v5, v5
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v6f16:
Expand Down Expand Up @@ -2933,22 +2900,6 @@ define <8 x half> @v_test_canonicalize_var_v8f16(<8 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v8f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
; CI-NEXT: v_cvt_f16_f32_e32 v5, v5
; CI-NEXT: v_cvt_f16_f32_e32 v6, v6
; CI-NEXT: v_cvt_f16_f32_e32 v7, v7
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v4, v4
; CI-NEXT: v_cvt_f32_f16_e32 v5, v5
; CI-NEXT: v_cvt_f32_f16_e32 v6, v6
; CI-NEXT: v_cvt_f32_f16_e32 v7, v7
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v8f16:
Expand Down Expand Up @@ -3001,30 +2952,6 @@ define <12 x half> @v_test_canonicalize_var_v12f16(<12 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v12f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
; CI-NEXT: v_cvt_f16_f32_e32 v5, v5
; CI-NEXT: v_cvt_f16_f32_e32 v6, v6
; CI-NEXT: v_cvt_f16_f32_e32 v7, v7
; CI-NEXT: v_cvt_f16_f32_e32 v8, v8
; CI-NEXT: v_cvt_f16_f32_e32 v9, v9
; CI-NEXT: v_cvt_f16_f32_e32 v10, v10
; CI-NEXT: v_cvt_f16_f32_e32 v11, v11
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v4, v4
; CI-NEXT: v_cvt_f32_f16_e32 v5, v5
; CI-NEXT: v_cvt_f32_f16_e32 v6, v6
; CI-NEXT: v_cvt_f32_f16_e32 v7, v7
; CI-NEXT: v_cvt_f32_f16_e32 v8, v8
; CI-NEXT: v_cvt_f32_f16_e32 v9, v9
; CI-NEXT: v_cvt_f32_f16_e32 v10, v10
; CI-NEXT: v_cvt_f32_f16_e32 v11, v11
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v12f16:
Expand Down Expand Up @@ -3087,38 +3014,6 @@ define <16 x half> @v_test_canonicalize_var_v16f16(<16 x half> %val) #1 {
; CI-LABEL: v_test_canonicalize_var_v16f16:
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
; CI-NEXT: v_cvt_f16_f32_e32 v5, v5
; CI-NEXT: v_cvt_f16_f32_e32 v6, v6
; CI-NEXT: v_cvt_f16_f32_e32 v7, v7
; CI-NEXT: v_cvt_f16_f32_e32 v8, v8
; CI-NEXT: v_cvt_f16_f32_e32 v9, v9
; CI-NEXT: v_cvt_f16_f32_e32 v10, v10
; CI-NEXT: v_cvt_f16_f32_e32 v11, v11
; CI-NEXT: v_cvt_f16_f32_e32 v12, v12
; CI-NEXT: v_cvt_f16_f32_e32 v13, v13
; CI-NEXT: v_cvt_f16_f32_e32 v14, v14
; CI-NEXT: v_cvt_f16_f32_e32 v15, v15
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v4, v4
; CI-NEXT: v_cvt_f32_f16_e32 v5, v5
; CI-NEXT: v_cvt_f32_f16_e32 v6, v6
; CI-NEXT: v_cvt_f32_f16_e32 v7, v7
; CI-NEXT: v_cvt_f32_f16_e32 v8, v8
; CI-NEXT: v_cvt_f32_f16_e32 v9, v9
; CI-NEXT: v_cvt_f32_f16_e32 v10, v10
; CI-NEXT: v_cvt_f32_f16_e32 v11, v11
; CI-NEXT: v_cvt_f32_f16_e32 v12, v12
; CI-NEXT: v_cvt_f32_f16_e32 v13, v13
; CI-NEXT: v_cvt_f32_f16_e32 v14, v14
; CI-NEXT: v_cvt_f32_f16_e32 v15, v15
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v16f16:
Expand Down Expand Up @@ -3216,71 +3111,7 @@ define <32 x half> @v_test_canonicalize_var_v32f16(<32 x half> %val) #1 {
; CI: ; %bb.0:
; CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CI-NEXT: buffer_load_dword v31, off, s[0:3], s32
; CI-NEXT: v_cvt_f16_f32_e32 v0, v0
; CI-NEXT: v_cvt_f16_f32_e32 v1, v1
; CI-NEXT: v_cvt_f16_f32_e32 v2, v2
; CI-NEXT: v_cvt_f16_f32_e32 v3, v3
; CI-NEXT: v_cvt_f16_f32_e32 v4, v4
; CI-NEXT: v_cvt_f16_f32_e32 v5, v5
; CI-NEXT: v_cvt_f16_f32_e32 v6, v6
; CI-NEXT: v_cvt_f16_f32_e32 v7, v7
; CI-NEXT: v_cvt_f16_f32_e32 v8, v8
; CI-NEXT: v_cvt_f16_f32_e32 v9, v9
; CI-NEXT: v_cvt_f16_f32_e32 v10, v10
; CI-NEXT: v_cvt_f16_f32_e32 v11, v11
; CI-NEXT: v_cvt_f16_f32_e32 v12, v12
; CI-NEXT: v_cvt_f16_f32_e32 v13, v13
; CI-NEXT: v_cvt_f16_f32_e32 v14, v14
; CI-NEXT: v_cvt_f16_f32_e32 v15, v15
; CI-NEXT: v_cvt_f16_f32_e32 v16, v16
; CI-NEXT: v_cvt_f16_f32_e32 v17, v17
; CI-NEXT: v_cvt_f16_f32_e32 v18, v18
; CI-NEXT: v_cvt_f16_f32_e32 v19, v19
; CI-NEXT: v_cvt_f16_f32_e32 v20, v20
; CI-NEXT: v_cvt_f16_f32_e32 v21, v21
; CI-NEXT: v_cvt_f16_f32_e32 v22, v22
; CI-NEXT: v_cvt_f16_f32_e32 v23, v23
; CI-NEXT: v_cvt_f16_f32_e32 v24, v24
; CI-NEXT: v_cvt_f16_f32_e32 v25, v25
; CI-NEXT: v_cvt_f16_f32_e32 v26, v26
; CI-NEXT: v_cvt_f16_f32_e32 v27, v27
; CI-NEXT: v_cvt_f16_f32_e32 v28, v28
; CI-NEXT: v_cvt_f16_f32_e32 v29, v29
; CI-NEXT: v_cvt_f16_f32_e32 v30, v30
; CI-NEXT: v_cvt_f32_f16_e32 v0, v0
; CI-NEXT: v_cvt_f32_f16_e32 v1, v1
; CI-NEXT: v_cvt_f32_f16_e32 v2, v2
; CI-NEXT: v_cvt_f32_f16_e32 v3, v3
; CI-NEXT: v_cvt_f32_f16_e32 v4, v4
; CI-NEXT: v_cvt_f32_f16_e32 v5, v5
; CI-NEXT: v_cvt_f32_f16_e32 v6, v6
; CI-NEXT: v_cvt_f32_f16_e32 v7, v7
; CI-NEXT: v_cvt_f32_f16_e32 v8, v8
; CI-NEXT: v_cvt_f32_f16_e32 v9, v9
; CI-NEXT: v_cvt_f32_f16_e32 v10, v10
; CI-NEXT: v_cvt_f32_f16_e32 v11, v11
; CI-NEXT: v_cvt_f32_f16_e32 v12, v12
; CI-NEXT: v_cvt_f32_f16_e32 v13, v13
; CI-NEXT: v_cvt_f32_f16_e32 v14, v14
; CI-NEXT: v_cvt_f32_f16_e32 v15, v15
; CI-NEXT: v_cvt_f32_f16_e32 v16, v16
; CI-NEXT: v_cvt_f32_f16_e32 v17, v17
; CI-NEXT: v_cvt_f32_f16_e32 v18, v18
; CI-NEXT: v_cvt_f32_f16_e32 v19, v19
; CI-NEXT: v_cvt_f32_f16_e32 v20, v20
; CI-NEXT: v_cvt_f32_f16_e32 v21, v21
; CI-NEXT: v_cvt_f32_f16_e32 v22, v22
; CI-NEXT: v_cvt_f32_f16_e32 v23, v23
; CI-NEXT: v_cvt_f32_f16_e32 v24, v24
; CI-NEXT: v_cvt_f32_f16_e32 v25, v25
; CI-NEXT: v_cvt_f32_f16_e32 v26, v26
; CI-NEXT: v_cvt_f32_f16_e32 v27, v27
; CI-NEXT: v_cvt_f32_f16_e32 v28, v28
; CI-NEXT: v_cvt_f32_f16_e32 v29, v29
; CI-NEXT: v_cvt_f32_f16_e32 v30, v30
; CI-NEXT: s_waitcnt vmcnt(0)
; CI-NEXT: v_cvt_f16_f32_e32 v31, v31
; CI-NEXT: v_cvt_f32_f16_e32 v31, v31
; CI-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_test_canonicalize_var_v32f16:
Expand Down
Loading