[llvm-branch-commits] [llvm] [AArch64][TTI] Cost NEON Nots absorbed by And/Or users as free (PR #232788)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Oct 16 07:59:46 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-aarch64
Author: Valeriy Savchenko (SavchenkoValeriy)
<details>
<summary>Changes</summary>
Cost a NEON Not as free when every user is an And/Or in the same block,
since the Not folds into its BIC, ORN or BSL. If a user negates both of
its operands, only one Not can fold, so it still counts.
---
<sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>
---
Full diff: https://github.com/llvm/llvm-project/pull/232788.diff
2 Files Affected:
- (modified) llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp (+12)
- (modified) llvm/test/Analysis/CostModel/AArch64/bsl.ll (+90-9)
``````````diff
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index e4a883d3a960e..5d3be19fe23d0 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -5119,6 +5119,18 @@ InstructionCost AArch64TTIImpl::getArithmeticInstrCost(
if (Opcode == Instruction::And &&
matchBSL(CtxI->user_back(), MainAnd, OtherAnd, A, B))
return 0;
+
+ // A Not is folded into the BIC, ORN or BSL of each And/Or using it, unless
+ // that And/Or negates both of its operands.
+ if (match(CtxI, m_Not(m_Value())) &&
+ all_of(CtxI->users(), [&](const User *U) {
+ const auto *UI = cast<Instruction>(U);
+ return is_contained({Instruction::And, Instruction::Or},
+ UI->getOpcode()) &&
+ UI->getParent() == CtxI->getParent() &&
+ !match(UI, m_BinOp(m_Not(m_Value()), m_Not(m_Value())));
+ }))
+ return 0;
}
// Legalize the type.
diff --git a/llvm/test/Analysis/CostModel/AArch64/bsl.ll b/llvm/test/Analysis/CostModel/AArch64/bsl.ll
index 80a03e7307ac2..39eada0472a88 100644
--- a/llvm/test/Analysis/CostModel/AArch64/bsl.ll
+++ b/llvm/test/Analysis/CostModel/AArch64/bsl.ll
@@ -4,10 +4,9 @@
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
-; TODO: The Not is folded into BSL and should be free.
define <8 x i8> @bsl(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b) {
; COST-LABEL: 'bsl'
-; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <8 x i8> %m, splat (i8 -1)
; COST-NEXT: Cost Model: Found costs of 0 for: %x = and <8 x i8> %m, %a
; COST-NEXT: Cost Model: Found costs of 0 for: %y = and <8 x i8> %n, %b
; COST-NEXT: Cost Model: Found costs of 1 for: %r = or <8 x i8> %x, %y
@@ -24,10 +23,9 @@ define <8 x i8> @bsl(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b) {
ret <8 x i8> %r
}
-; TODO: The Not should be free here as well
define void @bsl_shared_mask(<16 x i8> %m, <16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> %d, ptr %p, ptr %q) {
; COST-LABEL: 'bsl_shared_mask'
-; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <16 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <16 x i8> %m, splat (i8 -1)
; COST-NEXT: Cost Model: Found costs of 0 for: %x0 = and <16 x i8> %m, %a
; COST-NEXT: Cost Model: Found costs of 0 for: %y0 = and <16 x i8> %n, %b
; COST-NEXT: Cost Model: Found costs of 1 for: %r0 = or <16 x i8> %x0, %y0
@@ -81,10 +79,93 @@ define <8 x i8> @bsl_not_multiuse(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b, ptr %p)
ret <8 x i8> %r
}
-; TODO: BIT is still formed, so %y and the Not should be free.
+define <8 x i8> @bic(<8 x i8> %a, <8 x i8> %m) {
+; COST-LABEL: 'bic'
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %r = and <8 x i8> %a, %n
+; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret <8 x i8> %r
+;
+; CODE-LABEL: bic:
+; CODE: // %bb.0:
+; CODE-NEXT: bic v0.8b, v0.8b, v1.8b
+; CODE-NEXT: ret
+ %n = xor <8 x i8> %m, splat (i8 -1)
+ %r = and <8 x i8> %a, %n
+ ret <8 x i8> %r
+}
+
+define <8 x i8> @orn(<8 x i8> %a, <8 x i8> %m) {
+; COST-LABEL: 'orn'
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %r = or <8 x i8> %a, %n
+; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret <8 x i8> %r
+;
+; CODE-LABEL: orn:
+; CODE: // %bb.0:
+; CODE-NEXT: orn v0.8b, v0.8b, v1.8b
+; CODE-NEXT: ret
+ %n = xor <8 x i8> %m, splat (i8 -1)
+ %r = or <8 x i8> %a, %n
+ ret <8 x i8> %r
+}
+
+define <8 x i8> @neg_not_xor(<8 x i8> %a, <8 x i8> %m) {
+; COST-LABEL: 'neg_not_xor'
+; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %r = xor <8 x i8> %a, %n
+; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret <8 x i8> %r
+;
+; CODE-LABEL: neg_not_xor:
+; CODE: // %bb.0:
+; CODE-NEXT: eor v0.8b, v1.8b, v0.8b
+; CODE-NEXT: mvn v0.8b, v0.8b
+; CODE-NEXT: ret
+ %n = xor <8 x i8> %m, splat (i8 -1)
+ %r = xor <8 x i8> %a, %n
+ ret <8 x i8> %r
+}
+
+define <8 x i8> @neg_not_two_nots(<8 x i8> %a, <8 x i8> %b) {
+; COST-LABEL: 'neg_not_two_nots'
+; COST-NEXT: Cost Model: Found costs of 1 for: %na = xor <8 x i8> %a, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %nb = xor <8 x i8> %b, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %r = and <8 x i8> %na, %nb
+; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret <8 x i8> %r
+;
+; CODE-LABEL: neg_not_two_nots:
+; CODE: // %bb.0:
+; CODE-NEXT: mvn v0.8b, v0.8b
+; CODE-NEXT: bic v0.8b, v0.8b, v1.8b
+; CODE-NEXT: ret
+ %na = xor <8 x i8> %a, splat (i8 -1)
+ %nb = xor <8 x i8> %b, splat (i8 -1)
+ %r = and <8 x i8> %na, %nb
+ ret <8 x i8> %r
+}
+
+define <8 x i8> @neg_not_bic_and_add(<8 x i8> %a, <8 x i8> %m) {
+; COST-LABEL: 'neg_not_bic_and_add'
+; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 1 for: %r = and <8 x i8> %a, %n
+; COST-NEXT: Cost Model: Found costs of 1 for: %s = add <8 x i8> %r, %n
+; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: ret <8 x i8> %s
+;
+; CODE-LABEL: neg_not_bic_and_add:
+; CODE: // %bb.0:
+; CODE-NEXT: mvn v2.8b, v1.8b
+; CODE-NEXT: bic v0.8b, v0.8b, v1.8b
+; CODE-NEXT: add v0.8b, v0.8b, v2.8b
+; CODE-NEXT: ret
+ %n = xor <8 x i8> %m, splat (i8 -1)
+ %r = and <8 x i8> %a, %n
+ %s = add <8 x i8> %r, %n
+ ret <8 x i8> %s
+}
+
+; TODO: BIT is still formed, so %y should be free.
define <8 x i8> @neg_bsl_and_multiuse(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b, ptr %p) {
; COST-LABEL: 'neg_bsl_and_multiuse'
-; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <8 x i8> %m, splat (i8 -1)
; COST-NEXT: Cost Model: Found costs of 1 for: %x = and <8 x i8> %m, %a
; COST-NEXT: Cost Model: Found costs of 1 for: %y = and <8 x i8> %n, %b
; COST-NEXT: Cost Model: Found costs of 1 for: %r = or <8 x i8> %x, %y
@@ -108,7 +189,7 @@ define <8 x i8> @neg_bsl_and_multiuse(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b, ptr
define <8 x i8> @neg_bsl_other_block(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b, i1 %f, ptr %p) {
; COST-LABEL: 'neg_bsl_other_block'
-; COST-NEXT: Cost Model: Found costs of 1 for: %n = xor <8 x i8> %m, splat (i8 -1)
+; COST-NEXT: Cost Model: Found costs of 0 for: %n = xor <8 x i8> %m, splat (i8 -1)
; COST-NEXT: Cost Model: Found costs of 1 for: %x = and <8 x i8> %m, %a
; COST-NEXT: Cost Model: Found costs of 1 for: %y = and <8 x i8> %n, %b
; COST-NEXT: Cost Model: Found costs of RThru:0 CodeSize:1 Lat:1 SizeLat:1 for: br i1 %f, label %t, label %e
@@ -119,13 +200,13 @@ define <8 x i8> @neg_bsl_other_block(<8 x i8> %m, <8 x i8> %a, <8 x i8> %b, i1 %
;
; CODE-LABEL: neg_bsl_other_block:
; CODE: // %bb.0: // %entry
-; CODE-NEXT: tbz w0, #0, .LBB4_2
+; CODE-NEXT: tbz w0, #0, .LBB9_2
; CODE-NEXT: // %bb.1: // %t
; CODE-NEXT: and v1.8b, v0.8b, v1.8b
; CODE-NEXT: bic v0.8b, v2.8b, v0.8b
; CODE-NEXT: orr v1.8b, v1.8b, v0.8b
; CODE-NEXT: str d1, [x1]
-; CODE-NEXT: .LBB4_2: // %common.ret
+; CODE-NEXT: .LBB9_2: // %common.ret
; CODE-NEXT: fmov d0, d1
; CODE-NEXT: ret
entry:
``````````
</details>
https://github.com/llvm/llvm-project/pull/232788
More information about the llvm-branch-commits
mailing list