6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-12-29 10:32:47 +00:00
Code Issues Projects Releases Wiki Activity
fd4c349de2
llvm-6502/test/CodeGen/X86/MergeConsecutiveStores.ll

496 lines
18 KiB
LLVM
Raw Normal View History

When merging connsecutive stores, use vectors to store the constant zero. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165267 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-04 22:35:15 +00:00
; RUN: llc -march=x86-64 -mcpu=corei7 -mattr=+avx < %s | FileCheck %s
Add the ability to use GEPs for address sinking in CGP The current memory-instruction optimization logic in CGP, which sinks parts of the address computation that can be adsorbed by the addressing mode, does this by explicitly converting the relevant part of the address computation into IR-level integer operations (making use of ptrtoint and inttoptr). For most targets this is currently not a problem, but for targets wishing to make use of IR-level aliasing analysis during CodeGen, the use of ptrtoint/inttoptr is a problem for two reasons: 1. BasicAA becomes less powerful in the face of the ptrtoint/inttoptr 2. In cases where type-punning was used, and BasicAA was used to override TBAA, BasicAA may no longer do so. (this had forced us to disable all use of TBAA in CodeGen; something which we can now enable again) This (use of GEPs instead of ptrtoint/inttoptr) is not currently enabled by default (except for those targets that use AA during CodeGen), and so aside from some PowerPC subtargets and SystemZ, there should be no change in behavior. We may be able to switch completely away from the ptrtoint/inttoptr sinking on all targets, but further testing is required. I've doubled-up on a number of existing tests that are sensitive to the address sinking behavior (including some store-merging tests that are sensitive to the order of the resulting ADD operations at the SDAG level). git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206092 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-12 00:59:48 +00:00
; RUN: llc -march=x86-64 -mcpu=corei7 -mattr=+avx -addr-sink-using-gep=1 < %s | FileCheck %s
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
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-macosx10.8.0"
%struct.A = type { i8, i8, i8, i8, i8, i8, i8, i8 }
%struct.B = type { i32, i32, i32, i32, i32, i32, i32, i32 }
; CHECK: merge_const_store
; save 1,2,3 ... as one big integer.
; CHECK: movabsq $578437695752307201
; CHECK: ret
define void @merge_const_store(i32 %count, %struct.A* nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph:
%i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
%.01 = phi %struct.A* [ %11, %.lr.ph ], [ %p, %0 ]
%2 = getelementptr inbounds %struct.A* %.01, i64 0, i32 0
store i8 1, i8* %2, align 1
%3 = getelementptr inbounds %struct.A* %.01, i64 0, i32 1
store i8 2, i8* %3, align 1
%4 = getelementptr inbounds %struct.A* %.01, i64 0, i32 2
store i8 3, i8* %4, align 1
%5 = getelementptr inbounds %struct.A* %.01, i64 0, i32 3
store i8 4, i8* %5, align 1
%6 = getelementptr inbounds %struct.A* %.01, i64 0, i32 4
store i8 5, i8* %6, align 1
%7 = getelementptr inbounds %struct.A* %.01, i64 0, i32 5
store i8 6, i8* %7, align 1
%8 = getelementptr inbounds %struct.A* %.01, i64 0, i32 6
store i8 7, i8* %8, align 1
%9 = getelementptr inbounds %struct.A* %.01, i64 0, i32 7
store i8 8, i8* %9, align 1
%10 = add nsw i32 %i.02, 1
%11 = getelementptr inbounds %struct.A* %.01, i64 1
%exitcond = icmp eq i32 %10, %count
br i1 %exitcond, label %._crit_edge, label %.lr.ph
._crit_edge:
ret void
}
Dont merge consecutive loads/stores into vectors when noimplicitfloat is used. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@175190 91177308-0d34-0410-b5e6-96231b3b80d8
2013-02-14 18:28:52 +00:00
; No vectors because we use noimplicitfloat
; CHECK: merge_const_store_no_vec
; CHECK-NOT: vmovups
; CHECK: ret
define void @merge_const_store_no_vec(i32 %count, %struct.B* nocapture %p) noimplicitfloat{
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph:
%i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
%.01 = phi %struct.B* [ %11, %.lr.ph ], [ %p, %0 ]
%2 = getelementptr inbounds %struct.B* %.01, i64 0, i32 0
store i32 0, i32* %2, align 4
%3 = getelementptr inbounds %struct.B* %.01, i64 0, i32 1
store i32 0, i32* %3, align 4
%4 = getelementptr inbounds %struct.B* %.01, i64 0, i32 2
store i32 0, i32* %4, align 4
%5 = getelementptr inbounds %struct.B* %.01, i64 0, i32 3
store i32 0, i32* %5, align 4
%6 = getelementptr inbounds %struct.B* %.01, i64 0, i32 4
store i32 0, i32* %6, align 4
%7 = getelementptr inbounds %struct.B* %.01, i64 0, i32 5
store i32 0, i32* %7, align 4
%8 = getelementptr inbounds %struct.B* %.01, i64 0, i32 6
store i32 0, i32* %8, align 4
%9 = getelementptr inbounds %struct.B* %.01, i64 0, i32 7
store i32 0, i32* %9, align 4
%10 = add nsw i32 %i.02, 1
%11 = getelementptr inbounds %struct.B* %.01, i64 1
%exitcond = icmp eq i32 %10, %count
br i1 %exitcond, label %._crit_edge, label %.lr.ph
._crit_edge:
ret void
}
When merging connsecutive stores, use vectors to store the constant zero. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165267 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-04 22:35:15 +00:00
; Move the constants using a single vector store.
; CHECK: merge_const_store_vec
On Sandybridge split unaligned 256bit stores into two xmm-sized stores. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@172894 91177308-0d34-0410-b5e6-96231b3b80d8
2013-01-19 08:38:41 +00:00
; CHECK: vmovups
When merging connsecutive stores, use vectors to store the constant zero. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165267 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-04 22:35:15 +00:00
; CHECK: ret
define void @merge_const_store_vec(i32 %count, %struct.B* nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph:
%i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
%.01 = phi %struct.B* [ %11, %.lr.ph ], [ %p, %0 ]
%2 = getelementptr inbounds %struct.B* %.01, i64 0, i32 0
store i32 0, i32* %2, align 4
%3 = getelementptr inbounds %struct.B* %.01, i64 0, i32 1
store i32 0, i32* %3, align 4
%4 = getelementptr inbounds %struct.B* %.01, i64 0, i32 2
store i32 0, i32* %4, align 4
%5 = getelementptr inbounds %struct.B* %.01, i64 0, i32 3
store i32 0, i32* %5, align 4
%6 = getelementptr inbounds %struct.B* %.01, i64 0, i32 4
store i32 0, i32* %6, align 4
%7 = getelementptr inbounds %struct.B* %.01, i64 0, i32 5
store i32 0, i32* %7, align 4
%8 = getelementptr inbounds %struct.B* %.01, i64 0, i32 6
store i32 0, i32* %8, align 4
%9 = getelementptr inbounds %struct.B* %.01, i64 0, i32 7
store i32 0, i32* %9, align 4
%10 = add nsw i32 %i.02, 1
%11 = getelementptr inbounds %struct.B* %.01, i64 1
%exitcond = icmp eq i32 %10, %count
br i1 %exitcond, label %._crit_edge, label %.lr.ph
._crit_edge:
ret void
}
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
; Move the first 4 constants as a single vector. Move the rest as scalars.
; CHECK: merge_nonconst_store
; CHECK: movl $67305985
; CHECK: movb
; CHECK: movb
; CHECK: movb
; CHECK: movb
; CHECK: ret
define void @merge_nonconst_store(i32 %count, i8 %zz, %struct.A* nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph:
%i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
%.01 = phi %struct.A* [ %11, %.lr.ph ], [ %p, %0 ]
%2 = getelementptr inbounds %struct.A* %.01, i64 0, i32 0
store i8 1, i8* %2, align 1
%3 = getelementptr inbounds %struct.A* %.01, i64 0, i32 1
store i8 2, i8* %3, align 1
%4 = getelementptr inbounds %struct.A* %.01, i64 0, i32 2
store i8 3, i8* %4, align 1
%5 = getelementptr inbounds %struct.A* %.01, i64 0, i32 3
store i8 4, i8* %5, align 1
%6 = getelementptr inbounds %struct.A* %.01, i64 0, i32 4
store i8 %zz, i8* %6, align 1 ; <----------- Not a const;
%7 = getelementptr inbounds %struct.A* %.01, i64 0, i32 5
store i8 6, i8* %7, align 1
%8 = getelementptr inbounds %struct.A* %.01, i64 0, i32 6
store i8 7, i8* %8, align 1
%9 = getelementptr inbounds %struct.A* %.01, i64 0, i32 7
store i8 8, i8* %9, align 1
%10 = add nsw i32 %i.02, 1
%11 = getelementptr inbounds %struct.A* %.01, i64 1
%exitcond = icmp eq i32 %10, %count
br i1 %exitcond, label %._crit_edge, label %.lr.ph
._crit_edge:
ret void
}
fix typo, add spaces; NFC git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224384 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-16 22:48:42 +00:00
; CHECK-LABEL: merge_loads_i16:
; load:
; CHECK: movw
; store:
; CHECK: movw
; CHECK: ret
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
define void @merge_loads_i16(i32 %count, %struct.A* noalias nocapture %q, %struct.A* noalias nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph: ; preds = %0
%2 = getelementptr inbounds %struct.A* %q, i64 0, i32 0
%3 = getelementptr inbounds %struct.A* %q, i64 0, i32 1
br label %4
; <label>:4 ; preds = %4, %.lr.ph
%i.02 = phi i32 [ 0, %.lr.ph ], [ %9, %4 ]
%.01 = phi %struct.A* [ %p, %.lr.ph ], [ %10, %4 ]
%5 = load i8* %2, align 1
%6 = load i8* %3, align 1
%7 = getelementptr inbounds %struct.A* %.01, i64 0, i32 0
store i8 %5, i8* %7, align 1
%8 = getelementptr inbounds %struct.A* %.01, i64 0, i32 1
store i8 %6, i8* %8, align 1
%9 = add nsw i32 %i.02, 1
%10 = getelementptr inbounds %struct.A* %.01, i64 1
%exitcond = icmp eq i32 %9, %count
br i1 %exitcond, label %._crit_edge, label %4
._crit_edge: ; preds = %4, %0
ret void
}
fix typo, add spaces; NFC git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224384 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-16 22:48:42 +00:00
; The loads and the stores are interleaved. Can't merge them.
; CHECK-LABEL: no_merge_loads:
; CHECK: movb
; CHECK: movb
; CHECK: movb
; CHECK: movb
; CHECK: ret
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
define void @no_merge_loads(i32 %count, %struct.A* noalias nocapture %q, %struct.A* noalias nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph: ; preds = %0
%2 = getelementptr inbounds %struct.A* %q, i64 0, i32 0
%3 = getelementptr inbounds %struct.A* %q, i64 0, i32 1
br label %a4
a4: ; preds = %4, %.lr.ph
%i.02 = phi i32 [ 0, %.lr.ph ], [ %a9, %a4 ]
%.01 = phi %struct.A* [ %p, %.lr.ph ], [ %a10, %a4 ]
%a5 = load i8* %2, align 1
%a7 = getelementptr inbounds %struct.A* %.01, i64 0, i32 0
store i8 %a5, i8* %a7, align 1
%a8 = getelementptr inbounds %struct.A* %.01, i64 0, i32 1
%a6 = load i8* %3, align 1
store i8 %a6, i8* %a8, align 1
%a9 = add nsw i32 %i.02, 1
%a10 = getelementptr inbounds %struct.A* %.01, i64 1
%exitcond = icmp eq i32 %a9, %count
br i1 %exitcond, label %._crit_edge, label %a4
._crit_edge: ; preds = %4, %0
ret void
}
fix typo, add spaces; NFC git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224384 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-16 22:48:42 +00:00
; CHECK-LABEL: merge_loads_integer:
; load:
; CHECK: movq
; store:
; CHECK: movq
; CHECK: ret
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
define void @merge_loads_integer(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
%1 = icmp sgt i32 %count, 0
br i1 %1, label %.lr.ph, label %._crit_edge
.lr.ph: ; preds = %0
%2 = getelementptr inbounds %struct.B* %q, i64 0, i32 0
%3 = getelementptr inbounds %struct.B* %q, i64 0, i32 1
br label %4
; <label>:4 ; preds = %4, %.lr.ph
%i.02 = phi i32 [ 0, %.lr.ph ], [ %9, %4 ]
%.01 = phi %struct.B* [ %p, %.lr.ph ], [ %10, %4 ]
%5 = load i32* %2
%6 = load i32* %3
%7 = getelementptr inbounds %struct.B* %.01, i64 0, i32 0
store i32 %5, i32* %7
%8 = getelementptr inbounds %struct.B* %.01, i64 0, i32 1
store i32 %6, i32* %8
%9 = add nsw i32 %i.02, 1
%10 = getelementptr inbounds %struct.B* %.01, i64 1
%exitcond = icmp eq i32 %9, %count
br i1 %exitcond, label %._crit_edge, label %4
._crit_edge: ; preds = %4, %0
ret void
}
fix typo, add spaces; NFC git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224384 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-16 22:48:42 +00:00
; CHECK-LABEL: merge_loads_vector:
; load:
; CHECK: movups
; store:
; CHECK: movups
; CHECK: ret
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
define void @merge_loads_vector(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
%a1 = icmp sgt i32 %count, 0
br i1 %a1, label %.lr.ph, label %._crit_edge
.lr.ph: ; preds = %0
%a2 = getelementptr inbounds %struct.B* %q, i64 0, i32 0
%a3 = getelementptr inbounds %struct.B* %q, i64 0, i32 1
%a4 = getelementptr inbounds %struct.B* %q, i64 0, i32 2
%a5 = getelementptr inbounds %struct.B* %q, i64 0, i32 3
br label %block4
block4: ; preds = %4, %.lr.ph
%i.02 = phi i32 [ 0, %.lr.ph ], [ %c9, %block4 ]
%.01 = phi %struct.B* [ %p, %.lr.ph ], [ %c10, %block4 ]
%a7 = getelementptr inbounds %struct.B* %.01, i64 0, i32 0
%a8 = getelementptr inbounds %struct.B* %.01, i64 0, i32 1
%a9 = getelementptr inbounds %struct.B* %.01, i64 0, i32 2
%a10 = getelementptr inbounds %struct.B* %.01, i64 0, i32 3
%b1 = load i32* %a2
%b2 = load i32* %a3
%b3 = load i32* %a4
%b4 = load i32* %a5
store i32 %b1, i32* %a7
store i32 %b2, i32* %a8
store i32 %b3, i32* %a9
store i32 %b4, i32* %a10
%c9 = add nsw i32 %i.02, 1
%c10 = getelementptr inbounds %struct.B* %.01, i64 1
%exitcond = icmp eq i32 %c9, %count
br i1 %exitcond, label %._crit_edge, label %block4
._crit_edge: ; preds = %4, %0
ret void
}
fix typo, add spaces; NFC git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224384 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-16 22:48:42 +00:00
; CHECK-LABEL: merge_loads_no_align:
; load:
; CHECK: movl
; CHECK: movl
; CHECK: movl
; CHECK: movl
; store:
; CHECK: movl
; CHECK: movl
; CHECK: movl
; CHECK: movl
; CHECK: ret
A DAGCombine optimization for mergeing consecutive stores to memory. The optimization is not profitable in many cases because modern processors perform multiple stores in parallel and merging stores prior to merging requires extra work. We handle two main cases: 1. Store of multiple consecutive constants: q->a = 3; q->4 = 5; In this case we store a single legal wide integer. 2. Store of multiple consecutive loads: int a = p->a; int b = p->b; q->a = a; q->b = b; In this case we load/store either ilegal vector registers or legal wide integer registers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@165125 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-03 16:11:15 +00:00
define void @merge_loads_no_align(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
%a1 = icmp sgt i32 %count, 0
br i1 %a1, label %.lr.ph, label %._crit_edge
.lr.ph: ; preds = %0
%a2 = getelementptr inbounds %struct.B* %q, i64 0, i32 0
%a3 = getelementptr inbounds %struct.B* %q, i64 0, i32 1
%a4 = getelementptr inbounds %struct.B* %q, i64 0, i32 2
%a5 = getelementptr inbounds %struct.B* %q, i64 0, i32 3
br label %block4
block4: ; preds = %4, %.lr.ph
%i.02 = phi i32 [ 0, %.lr.ph ], [ %c9, %block4 ]
%.01 = phi %struct.B* [ %p, %.lr.ph ], [ %c10, %block4 ]
%a7 = getelementptr inbounds %struct.B* %.01, i64 0, i32 0
%a8 = getelementptr inbounds %struct.B* %.01, i64 0, i32 1
%a9 = getelementptr inbounds %struct.B* %.01, i64 0, i32 2
%a10 = getelementptr inbounds %struct.B* %.01, i64 0, i32 3
%b1 = load i32* %a2, align 1
%b2 = load i32* %a3, align 1
%b3 = load i32* %a4, align 1
%b4 = load i32* %a5, align 1
store i32 %b1, i32* %a7, align 1
store i32 %b2, i32* %a8, align 1
store i32 %b3, i32* %a9, align 1
store i32 %b4, i32* %a10, align 1
%c9 = add nsw i32 %i.02, 1
%c10 = getelementptr inbounds %struct.B* %.01, i64 1
%exitcond = icmp eq i32 %c9, %count
br i1 %exitcond, label %._crit_edge, label %block4
._crit_edge: ; preds = %4, %0
ret void
}
Merge load/store sequences with adresses: base + index + offset We would also like to merge sequences that involve a variable index like in the example below. int index = *idx++ int i0 = c[index+0]; int i1 = c[index+1]; b[0] = i0; b[1] = i1; By extending the parsing of the base pointer to handle dags that contain a base, index, and offset we can handle examples like the one above. The dag for the code above will look something like: (load (i64 add (i64 copyfromreg %c) (i64 signextend (i8 load %index)))) (load (i64 add (i64 copyfromreg %c) (i64 signextend (i32 add (i32 signextend (i8 load %index)) (i32 1))))) The code that parses the tree ignores the intermediate sign extensions. However, if there is a sign extension it needs to be on all indexes. (load (i64 add (i64 copyfromreg %c) (i64 signextend (add (i8 load %index) (i8 1)))) vs (load (i64 add (i64 copyfromreg %c) (i64 signextend (i32 add (i32 signextend (i8 load %index)) (i32 1))))) radar://13536387 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178483 91177308-0d34-0410-b5e6-96231b3b80d8
2013-04-01 18:12:58 +00:00
; Make sure that we merge the consecutive load/store sequence below and use a
; word (16 bit) instead of a byte copy.
; CHECK: MergeLoadStoreBaseIndexOffset
; CHECK: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
; CHECK: movw [[REG]], (%{{.*}})
define void @MergeLoadStoreBaseIndexOffset(i64* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %11, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %10, %1 ]
%.0 = phi i64* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i64* %.0, i64 1
%3 = load i64* %.0, align 1
%4 = getelementptr inbounds i8* %c, i64 %3
%5 = load i8* %4, align 1
%6 = add i64 %3, 1
%7 = getelementptr inbounds i8* %c, i64 %6
%8 = load i8* %7, align 1
store i8 %5, i8* %.08, align 1
%9 = getelementptr inbounds i8* %.08, i64 1
store i8 %8, i8* %9, align 1
%10 = getelementptr inbounds i8* %.08, i64 2
%11 = add nsw i32 %.09, -1
%12 = icmp eq i32 %11, 0
br i1 %12, label %13, label %1
; <label>:13
ret void
}
; Make sure that we merge the consecutive load/store sequence below and use a
; word (16 bit) instead of a byte copy even if there are intermediate sign
; extensions.
; CHECK: MergeLoadStoreBaseIndexOffsetSext
; CHECK: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
; CHECK: movw [[REG]], (%{{.*}})
define void @MergeLoadStoreBaseIndexOffsetSext(i8* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
%.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i8* %.0, i64 1
%3 = load i8* %.0, align 1
%4 = sext i8 %3 to i64
%5 = getelementptr inbounds i8* %c, i64 %4
%6 = load i8* %5, align 1
%7 = add i64 %4, 1
%8 = getelementptr inbounds i8* %c, i64 %7
%9 = load i8* %8, align 1
store i8 %6, i8* %.08, align 1
%10 = getelementptr inbounds i8* %.08, i64 1
store i8 %9, i8* %10, align 1
%11 = getelementptr inbounds i8* %.08, i64 2
%12 = add nsw i32 %.09, -1
%13 = icmp eq i32 %12, 0
br i1 %13, label %14, label %1
; <label>:14
ret void
}
; However, we can only merge ignore sign extensions when they are on all memory
; computations;
; CHECK: loadStoreBaseIndexOffsetSextNoSex
; CHECK-NOT: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
; CHECK-NOT: movw [[REG]], (%{{.*}})
define void @loadStoreBaseIndexOffsetSextNoSex(i8* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
%.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i8* %.0, i64 1
%3 = load i8* %.0, align 1
%4 = sext i8 %3 to i64
%5 = getelementptr inbounds i8* %c, i64 %4
%6 = load i8* %5, align 1
%7 = add i8 %3, 1
%wrap.4 = sext i8 %7 to i64
%8 = getelementptr inbounds i8* %c, i64 %wrap.4
%9 = load i8* %8, align 1
store i8 %6, i8* %.08, align 1
%10 = getelementptr inbounds i8* %.08, i64 1
store i8 %9, i8* %10, align 1
%11 = getelementptr inbounds i8* %.08, i64 2
%12 = add nsw i32 %.09, -1
%13 = icmp eq i32 %12, 0
br i1 %13, label %14, label %1
; <label>:14
ret void
}
merge consecutive stores of extracted vector elements (PR21711) This is a 2nd try at the same optimization as http://reviews.llvm.org/D6698. That patch was checked in at r224611, but reverted at r225031 because it caused a failure outside of the regression tests. The cause of the crash was not recognizing consecutive stores that have mixed source values (loads and vector element extracts), so this patch adds a check to bail out if any store value is not coming from a vector element extract. This patch also refactors the shared logic of the constant source and vector extracted elements source cases into a helper function. Differential Revision: http://reviews.llvm.org/D6850 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@226845 91177308-0d34-0410-b5e6-96231b3b80d8
2015-01-22 18:21:26 +00:00
; PR21711 ( http://llvm.org/bugs/show_bug.cgi?id=21711 )
define void @merge_vec_element_store(<8 x float> %v, float* %ptr) {
%vecext0 = extractelement <8 x float> %v, i32 0
%vecext1 = extractelement <8 x float> %v, i32 1
%vecext2 = extractelement <8 x float> %v, i32 2
%vecext3 = extractelement <8 x float> %v, i32 3
%vecext4 = extractelement <8 x float> %v, i32 4
%vecext5 = extractelement <8 x float> %v, i32 5
%vecext6 = extractelement <8 x float> %v, i32 6
%vecext7 = extractelement <8 x float> %v, i32 7
%arrayidx1 = getelementptr inbounds float* %ptr, i64 1
%arrayidx2 = getelementptr inbounds float* %ptr, i64 2
%arrayidx3 = getelementptr inbounds float* %ptr, i64 3
%arrayidx4 = getelementptr inbounds float* %ptr, i64 4
%arrayidx5 = getelementptr inbounds float* %ptr, i64 5
%arrayidx6 = getelementptr inbounds float* %ptr, i64 6
%arrayidx7 = getelementptr inbounds float* %ptr, i64 7
store float %vecext0, float* %ptr, align 4
store float %vecext1, float* %arrayidx1, align 4
store float %vecext2, float* %arrayidx2, align 4
store float %vecext3, float* %arrayidx3, align 4
store float %vecext4, float* %arrayidx4, align 4
store float %vecext5, float* %arrayidx5, align 4
store float %vecext6, float* %arrayidx6, align 4
store float %vecext7, float* %arrayidx7, align 4
ret void
; CHECK-LABEL: merge_vec_element_store
; CHECK: vmovups
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
}
; This is a minimized test based on real code that was failing.
; We could merge stores (and loads) like this...
define void @merge_vec_element_and_scalar_load([6 x i64]* %array) {
%idx0 = getelementptr inbounds [6 x i64]* %array, i64 0, i64 0
%idx1 = getelementptr inbounds [6 x i64]* %array, i64 0, i64 1
%idx4 = getelementptr inbounds [6 x i64]* %array, i64 0, i64 4
%idx5 = getelementptr inbounds [6 x i64]* %array, i64 0, i64 5
%a0 = load i64* %idx0, align 8
store i64 %a0, i64* %idx4, align 8
%b = bitcast i64* %idx1 to <2 x i64>*
%v = load <2 x i64>* %b, align 8
%a1 = extractelement <2 x i64> %v, i32 0
store i64 %a1, i64* %idx5, align 8
ret void
; CHECK-LABEL: merge_vec_element_and_scalar_load
; CHECK: movq (%rdi), %rax
; CHECK-NEXT: movq %rax, 32(%rdi)
; CHECK-NEXT: movq 8(%rdi), %rax
; CHECK-NEXT: movq %rax, 40(%rdi)
; CHECK-NEXT: retq
}
Reference in New Issue Copy Permalink