[CostModel] Mark ssa_copy as free

[davemgreen]

These are intrinsics are only used ephemerally and be should be given a zero cost.

[llvmbot]

@llvm/pr-subscribers-llvm-analysis

Author: David Green (davemgreen)

Changes

These are intrinsics are only used ephemerally and be should be given a zero cost.

---

Full diff: https://github.com/llvm/llvm-project/pull/75294.diff

2 Files Affected:

• (modified) llvm/include/llvm/CodeGen/BasicTTIImpl.h (+2)
• (modified) llvm/test/Analysis/CostModel/ARM/intrinsic-cost-kinds.ll (+12-12)

diff --git a/llvm/include/llvm/CodeGen/BasicTTIImpl.h b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
index e05ce2890a08c8..33db350a12102a 100644
--- a/llvm/include/llvm/CodeGen/BasicTTIImpl.h
+++ b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -1510,6 +1510,8 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     Intrinsic::ID IID = ICA.getID();
     if (Function::isTargetIntrinsic(IID))
       return TargetTransformInfo::TCC_Basic;
+    if (IID == Intrinsic::ssa_copy)
+      return TargetTransformInfo::TCC_Free;
 
     if (ICA.isTypeBasedOnly())
       return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
diff --git a/llvm/test/Analysis/CostModel/ARM/intrinsic-cost-kinds.ll b/llvm/test/Analysis/CostModel/ARM/intrinsic-cost-kinds.ll
index 9a9193b36de029..d58f3348e41fed 100644
--- a/llvm/test/Analysis/CostModel/ARM/intrinsic-cost-kinds.ll
+++ b/llvm/test/Analysis/CostModel/ARM/intrinsic-cost-kinds.ll
@@ -341,27 +341,27 @@ define void @memcpy(ptr %a, ptr %b, i32 %c) {
 define void @ssa_copy() {
   ; CHECK: %{{.*}} = llvm.intr.ssa.copy %{{.*}} : f32
 ; THRU-LABEL: 'ssa_copy'
-; THRU-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
-; THRU-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
-; THRU-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
+; THRU-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
+; THRU-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
+; THRU-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
 ; THRU-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; LATE-LABEL: 'ssa_copy'
-; LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
-; LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
-; LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
+; LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
+; LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
+; LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
 ; LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
 ;
 ; SIZE-LABEL: 'ssa_copy'
-; SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
-; SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
-; SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
+; SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
+; SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
+; SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
 ; SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
 ;
 ; SIZE_LATE-LABEL: 'ssa_copy'
-; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
-; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
-; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
+; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %i = call i32 @llvm.ssa.copy.i32(i32 undef)
+; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %f = call float @llvm.ssa.copy.f32(float undef)
+; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %p = call ptr @llvm.ssa.copy.p0(ptr undef)
 ; SIZE_LATE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
 ;
   %i = call i32 @llvm.ssa.copy.i32(i32 undef)

[fhahn]

I am curious if you encountered cases where those intrinsics are still around after the middle-end optimization pipeline. I think all users of those intrinsics should clean them up, but there might be some places that got missed.

[davemgreen]

No the costing is done during a pass that adds and removes them. This could be handled specially in that pass, but it is more general fix to give a zero cost to everywhere.

[nikic]

This should be in TargetTransformInfoImpl.h.

[davemgreen]

Ah I see. That is better, thanks.

[nikic]

LGTM