[AMDGPU] Generate s_absdiff_i32 #164835

LU-JOHN · 2025-10-23T15:33:16Z

Generate s_absdiff_i32. Tested in absdiff.ll. Also update s_cmp_0.ll to test that s_absdiff_i32 is foldable with a s_cmp_lg_u32 sX, 0.

Signed-off-by: John Lu <John.Lu@amd.com>

llvmbot · 2025-10-23T15:34:05Z

@llvm/pr-subscribers-backend-amdgpu

Author: None (LU-JOHN)

Changes

Generate s_absdiff_i32. Tested in absdiff.ll. Also update s_cmp_0.ll to test that s_absdiff_i32 is foldable with a s_cmp_lg_u32 sX, 0.

Full diff: https://github.com/llvm/llvm-project/pull/164835.diff

3 Files Affected:

(modified) llvm/lib/Target/AMDGPU/SOPInstructions.td (+5-3)
(added) llvm/test/CodeGen/AMDGPU/absdiff.ll (+12)
(modified) llvm/test/CodeGen/AMDGPU/s_cmp_0.ll (+20-5)

diff --git a/llvm/lib/Target/AMDGPU/SOPInstructions.td b/llvm/lib/Target/AMDGPU/SOPInstructions.td
index 84287b621fe78..220318cb6bd40 100644
--- a/llvm/lib/Target/AMDGPU/SOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SOPInstructions.td
@@ -838,9 +838,11 @@ def S_CBRANCH_G_FORK : SOP2_Pseudo <
   let SubtargetPredicate = isGFX6GFX7GFX8GFX9;
 }
 
-let Defs = [SCC] in {
-def S_ABSDIFF_I32 : SOP2_32 <"s_absdiff_i32">;
-} // End Defs = [SCC]
+let isCommutable = 1, Defs = [SCC] in {
+def S_ABSDIFF_I32 : SOP2_32 <"s_absdiff_i32",
+  [(set i32:$sdst, (UniformUnaryFrag<abs> (sub_oneuse i32:$src0, i32:$src1)))]
+>;
+} // End isCommutable = 1, Defs = [SCC]
 
 let SubtargetPredicate = isGFX8GFX9 in {
   def S_RFE_RESTORE_B64 : SOP2_Pseudo <
diff --git a/llvm/test/CodeGen/AMDGPU/absdiff.ll b/llvm/test/CodeGen/AMDGPU/absdiff.ll
new file mode 100644
index 0000000000000..3ce58f55ff0b9
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/absdiff.ll
@@ -0,0 +1,12 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s
+
+define amdgpu_ps i32 @absdiff_v1(i32 inreg %arg, i32 inreg %arg2) {
+; CHECK-LABEL: absdiff_v1:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_absdiff_i32 s0, s0, s1
+; CHECK-NEXT:    ; return to shader part epilog
+  %diff = sub i32 %arg, %arg2
+  %res = call i32 @llvm.abs.i32(i32 %diff, i1 false)
+  ret i32 %res
+}
diff --git a/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll b/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll
index dd5f838b4a206..0166d7ac7ddc2 100644
--- a/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll
+++ b/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll
@@ -110,6 +110,21 @@ define amdgpu_ps i32 @abs32(i32 inreg %val0) {
   ret i32 %zext
 }
 
+define amdgpu_ps i32 @absdiff32(i32 inreg %val0, i32 inreg %val1) {
+; CHECK-LABEL: absdiff32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_absdiff_i32 s0, s0, s1
+; CHECK-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; CHECK-NEXT:    v_readfirstlane_b32 s0, v0
+; CHECK-NEXT:    ; return to shader part epilog
+  %diff = sub i32 %val0, %val1
+  %result = call i32 @llvm.abs.i32(i32 %diff, i1 false)
+  %cmp = icmp ne i32 %result, 0
+  %zext = zext i1 %cmp to i32
+  ret i32 %zext
+}
+
 define amdgpu_ps i32 @and32(i32 inreg %val0, i32 inreg %val1) {
 ; CHECK-LABEL: and32:
 ; CHECK:       ; %bb.0:
@@ -608,14 +623,14 @@ define amdgpu_ps i32 @si_pc_add_rel_offset_must_not_optimize() {
 ; CHECK-NEXT:    s_add_u32 s0, s0, __unnamed_1@rel32@lo+4
 ; CHECK-NEXT:    s_addc_u32 s1, s1, __unnamed_1@rel32@hi+12
 ; CHECK-NEXT:    s_cmp_lg_u64 s[0:1], 0
-; CHECK-NEXT:    s_cbranch_scc0 .LBB35_2
+; CHECK-NEXT:    s_cbranch_scc0 .LBB36_2
 ; CHECK-NEXT:  ; %bb.1: ; %endif
 ; CHECK-NEXT:    s_mov_b32 s0, 1
-; CHECK-NEXT:    s_branch .LBB35_3
-; CHECK-NEXT:  .LBB35_2: ; %if
+; CHECK-NEXT:    s_branch .LBB36_3
+; CHECK-NEXT:  .LBB36_2: ; %if
 ; CHECK-NEXT:    s_mov_b32 s0, 0
-; CHECK-NEXT:    s_branch .LBB35_3
-; CHECK-NEXT:  .LBB35_3:
+; CHECK-NEXT:    s_branch .LBB36_3
+; CHECK-NEXT:  .LBB36_3:
   %cmp = icmp ne ptr addrspace(4) @1, null
   br i1 %cmp, label %endif, label %if

jayfoad

LGTM

llvm/lib/Target/AMDGPU/SOPInstructions.td

jayfoad · 2025-10-23T15:49:17Z

llvm/test/CodeGen/AMDGPU/s_cmp_0.ll

+; CHECK-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; CHECK-NEXT:    v_readfirstlane_b32 s0, v0


Codegen is terrible here, but that's not your fault. Should be:

Suggested change

; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0

; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1]

; CHECK-NEXT: v_readfirstlane_b32 s0, v0

; CHECK-NEXT: s_cselect_b32 s0, 1, 0

Signed-off-by: John Lu <John.Lu@amd.com>

arsenm · 2025-10-24T03:21:33Z

llvm/test/CodeGen/AMDGPU/absdiff.ll

+  %diff = sub i32 %arg, %arg2
+  %res = call i32 @llvm.abs.i32(i32 %diff, i1 false)
+  ret i32 %res
+}


Negative test for the multi use case?

Added negative test for multi-use case.

arsenm · 2025-10-24T03:21:44Z

llvm/test/CodeGen/AMDGPU/absdiff.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s
+
+define amdgpu_ps i32 @absdiff_v1(i32 inreg %arg, i32 inreg %arg2) {


Test vectors

Tested <2 x i32> and <4 x i32>

arsenm · 2025-10-24T03:21:53Z

llvm/test/CodeGen/AMDGPU/absdiff.ll

+; CHECK-NEXT:    s_absdiff_i32 s0, s0, s1
+; CHECK-NEXT:    ; return to shader part epilog
+  %diff = sub i32 %arg, %arg2
+  %res = call i32 @llvm.abs.i32(i32 %diff, i1 false)


Also test with true?

Added test with true variant.

arsenm · 2025-10-24T03:22:07Z

llvm/test/CodeGen/AMDGPU/absdiff.ll

@@ -0,0 +1,12 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s
+


Also test 16-bit promoted case

Added 16-bit test case.

Signed-off-by: John Lu <John.Lu@amd.com>

jayfoad

LGTM

Generate s_absdiff_i32. Tested in absdiff.ll. Also update s_cmp_0.ll to test that s_absdiff_i32 is foldable with a s_cmp_lg_u32 sX, 0. --------- Signed-off-by: John Lu <John.Lu@amd.com>

Generate s_absdiff_i32

a371ba3

Signed-off-by: John Lu <John.Lu@amd.com>

llvmbot added the backend:AMDGPU label Oct 23, 2025

LU-JOHN mentioned this pull request Oct 23, 2025

[AMDGPU] 32-bit ABS is a legal DAG node #163907

Merged

jayfoad reviewed Oct 23, 2025

View reviewed changes

llvm/lib/Target/AMDGPU/SOPInstructions.td Outdated Show resolved Hide resolved

jayfoad reviewed Oct 23, 2025

View reviewed changes

Remove unnecessary braces

81b0703

Signed-off-by: John Lu <John.Lu@amd.com>

LU-JOHN force-pushed the gen_absdiff branch from bf7da2d to 81b0703 Compare October 23, 2025 16:15

LU-JOHN requested a review from jayfoad October 23, 2025 17:28

arsenm reviewed Oct 24, 2025

View reviewed changes

Add absdiff test variations

6720b3d

Signed-off-by: John Lu <John.Lu@amd.com>

LU-JOHN requested a review from arsenm October 24, 2025 22:11

jayfoad approved these changes Oct 27, 2025

View reviewed changes

LU-JOHN merged commit 7d14733 into llvm:main Oct 27, 2025
10 checks passed

		@@ -0,0 +1,12 @@
		; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
		; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s \| FileCheck %s

[AMDGPU] Generate s_absdiff_i32 #164835

[AMDGPU] Generate s_absdiff_i32 #164835

Uh oh!

Conversation

LU-JOHN commented Oct 23, 2025

Uh oh!

llvmbot commented Oct 23, 2025

Uh oh!

jayfoad left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LU-JOHN Oct 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jayfoad left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

LU-JOHN Oct 24, 2025 •

edited

Loading