6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-11-02 07:17:36 +00:00
Code Issues Projects Releases Wiki Activity

SLP Vectorizer: Disable the vectorization of non power of two chains, such as <3 x float>, because we dont have a good cost model for these types.

Browse Source 
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187265 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Nadav Rotem
2013-07-26 22:53:11 +00:00
parent c1bb2d4325
commit a629c3a4f0
3 changed files with 39 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/Transforms/Vectorize/SLPVectorizer.cpp
View File

@@ -1664,23 +1664,7 @@ bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
} }
} }
if (Changed || ChainLen > VF)
return Changed; return Changed;
// Handle short chains. This helps us catch types such as <3 x float> that
// are smaller than vector size.
R.buildTree(Chain);
int Cost = R.getTreeCost();
if (Cost < CostThreshold) {
DEBUG(dbgs() << "SLP: Found store chain cost = " << Cost
<< " for size = " << ChainLen << "\n");
R.vectorizeTree();
return true;
}
return false;
} }
bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores, bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,

60
test/Transforms/SLPVectorizer/X86/cycle_dup.ll
View File

@@ -3,55 +3,61 @@
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.9.0" target triple = "x86_64-apple-macosx10.9.0"
; int foo(int *A) { ; int foo(int *A) {
; int r = A[0], g = A[1], b = A[2]; ; int r = A[0], g = A[1], b = A[2], a = A[3];
; for (int i=0; i < A[13]; i++) ; for (int i=0; i < A[13]; i++) {
; r*=18; g*=19; b*=12; ; r*=18; g*=19; b*=12; a *=9;
; A[0] = r; A[1] = g; A[2] = b; ; }
; A[0] = r; A[1] = g; A[2] = b; A[3] = a;
; } ; }
;CHECK-LABEL: @foo ;CHECK-LABEL: @foo
;CHECK: bitcast i32* %A to <3 x i32>* ;CHECK: bitcast i32* %A to <4 x i32>*
;CHECK-NEXT: load <3 x i32> ;CHECK-NEXT: load <4 x i32>
;CHECK: phi <3 x i32> ;CHECK: phi <4 x i32>
;CHECK-NEXT: mul <3 x i32> ;CHECK-NEXT: mul <4 x i32>
;CHECK-NOT: mul ;CHECK-NOT: mul
;CHECK: phi <3 x i32> ;CHECK: phi <4 x i32>
;CHECK: bitcast i32* %A to <3 x i32>* ;CHECK: bitcast i32* %A to <4 x i32>*
;CHECK-NEXT: store <3 x i32> ;CHECK-NEXT: store <4 x i32>
;CHECK-NEXT:ret i32 undef ;CHECK-NEXT:ret i32 undef
define i32 @foo(i32* nocapture %A) { define i32 @foo(i32* nocapture %A) #0 {
entry: entry:
%0 = load i32* %A, align 4 %0 = load i32* %A, align 4
%arrayidx1 = getelementptr inbounds i32* %A, i64 1 %arrayidx1 = getelementptr inbounds i32* %A, i64 1
%1 = load i32* %arrayidx1, align 4 %1 = load i32* %arrayidx1, align 4
%arrayidx2 = getelementptr inbounds i32* %A, i64 2 %arrayidx2 = getelementptr inbounds i32* %A, i64 2
%2 = load i32* %arrayidx2, align 4 %2 = load i32* %arrayidx2, align 4
%arrayidx3 = getelementptr inbounds i32* %A, i64 13 %arrayidx3 = getelementptr inbounds i32* %A, i64 3
%3 = load i32* %arrayidx3, align 4 %3 = load i32* %arrayidx3, align 4
%cmp18 = icmp sgt i32 %3, 0 %arrayidx4 = getelementptr inbounds i32* %A, i64 13
br i1 %cmp18, label %for.body, label %for.end %4 = load i32* %arrayidx4, align 4
%cmp24 = icmp sgt i32 %4, 0
br i1 %cmp24, label %for.body, label %for.end
for.body: ; preds = %entry, %for.body for.body: ; preds = %entry, %for.body
%i.022 = phi i32 [ %inc, %for.body ], [ 0, %entry ] %i.029 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
%b.021 = phi i32 [ %mul5, %for.body ], [ %2, %entry ] %a.028 = phi i32 [ %mul7, %for.body ], [ %3, %entry ]
%g.020 = phi i32 [ %mul4, %for.body ], [ %1, %entry ] %b.027 = phi i32 [ %mul6, %for.body ], [ %2, %entry ]
%r.019 = phi i32 [ %mul, %for.body ], [ %0, %entry ] %g.026 = phi i32 [ %mul5, %for.body ], [ %1, %entry ]
%mul = mul nsw i32 %r.019, 18 %r.025 = phi i32 [ %mul, %for.body ], [ %0, %entry ]
%mul4 = mul nsw i32 %g.020, 19 %mul = mul nsw i32 %r.025, 18
%mul5 = mul nsw i32 %b.021, 12 %mul5 = mul nsw i32 %g.026, 19
%inc = add nsw i32 %i.022, 1 %mul6 = mul nsw i32 %b.027, 12
%cmp = icmp slt i32 %inc, %3 %mul7 = mul nsw i32 %a.028, 9
%inc = add nsw i32 %i.029, 1
%cmp = icmp slt i32 %inc, %4
br i1 %cmp, label %for.body, label %for.end br i1 %cmp, label %for.body, label %for.end
for.end: ; preds = %for.body, %entry for.end: ; preds = %for.body, %entry
%b.0.lcssa = phi i32 [ %2, %entry ], [ %mul5, %for.body ] %a.0.lcssa = phi i32 [ %3, %entry ], [ %mul7, %for.body ]
%g.0.lcssa = phi i32 [ %1, %entry ], [ %mul4, %for.body ] %b.0.lcssa = phi i32 [ %2, %entry ], [ %mul6, %for.body ]
%g.0.lcssa = phi i32 [ %1, %entry ], [ %mul5, %for.body ]
%r.0.lcssa = phi i32 [ %0, %entry ], [ %mul, %for.body ] %r.0.lcssa = phi i32 [ %0, %entry ], [ %mul, %for.body ]
store i32 %r.0.lcssa, i32* %A, align 4 store i32 %r.0.lcssa, i32* %A, align 4
store i32 %g.0.lcssa, i32* %arrayidx1, align 4 store i32 %g.0.lcssa, i32* %arrayidx1, align 4
store i32 %b.0.lcssa, i32* %arrayidx2, align 4 store i32 %b.0.lcssa, i32* %arrayidx2, align 4
store i32 %a.0.lcssa, i32* %arrayidx3, align 4
ret i32 undef ret i32 undef
} }

12
test/Transforms/SLPVectorizer/X86/odd_store.ll
View File

@@ -10,12 +10,12 @@ target triple = "x86_64-apple-macosx10.8.0"
;} ;}
;CHECK-LABEL: @foo( ;CHECK-LABEL: @foo(
;CHECK: load <3 x float> ;CHECK-NOT: load <3 x float>
;CHECK: fmul <3 x float> ;CHECK-NOT: fmul <3 x float>
;CHECK: fpext <3 x float> ;CHECK-NOT: fpext <3 x float>
;CHECK: fadd <3 x double> ;CHECK-NOT: fadd <3 x double>
;CHECK: fptosi <3 x double> ;CHECK-NOT: fptosi <3 x double>
;CHECK: store <3 x i8> ;CHECK-NOT: store <3 x i8>
;CHECK: ret ;CHECK: ret
define i32 @foo(i8* noalias nocapture %A, float* noalias nocapture %B, float %T) { define i32 @foo(i8* noalias nocapture %A, float* noalias nocapture %B, float %T) {
%1 = getelementptr inbounds float* %B, i64 10 %1 = getelementptr inbounds float* %B, i64 10