[Beignet] [Patch V2] Fix memcpy and memset bug.
Song, Ruiling
ruiling.song at intel.com
Fri Oct 10 01:08:35 PDT 2014
The patch LGTM
> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of
> Yang Rong
> Sent: Wednesday, October 08, 2014 3:56 PM
> To: beignet at lists.freedesktop.org
> Cc: Yang, Rong R
> Subject: [Beignet] [Patch V2] Fix memcpy and memset bug.
>
> In ocl_memcpy.ll and ocl_memset.ll, index+4 should be less and equal than
> size when use int in memcpy and memset, and need consider alignment.
>
> V2: Add the alignment argument, fix the condition.
> Signed-off-by: Yang Rong <rong.r.yang at intel.com>
> ---
> backend/src/libocl/src/ocl_memcpy.ll | 704
> ++++++++++++++++-----------
> backend/src/libocl/src/ocl_memset.ll | 225 +++++----
> backend/src/llvm/llvm_intrinsic_lowering.cpp | 14 +-
> 3 files changed, 546 insertions(+), 397 deletions(-)
>
> diff --git a/backend/src/libocl/src/ocl_memcpy.ll
> b/backend/src/libocl/src/ocl_memcpy.ll
> index 476033e..64c68bb 100644
> --- a/backend/src/libocl/src/ocl_memcpy.ll
> +++ b/backend/src/libocl/src/ocl_memcpy.ll
> @@ -1,336 +1,446 @@
> ;The memcpy's source code.
> -; INLINE_OVERLOADABLE void __gen_memcpy(uchar* dst, uchar* src,
> size_t size) {
> +; INLINE_OVERLOADABLE void __gen_memcpy(uchar* dst, uchar* src,
> size_t
> +size, uint alignment) {
> ; size_t index = 0;
> -; while((index + 4) >= size) {
> -; *((uint *)(dst + index)) = *((uint *)(src + index));
> -; index += 4;
> -; }
> -; while(index < size) {
> -; dst[index] = src[index];
> -; index++;
> -; }
> +; if(alignment % 4 == 0) {
> +; while((index + 4) <= size) {
> +; *((__global uint *)(dst + index)) = *((__global uint *)(src +
> index));
> +; index += 4;
> +; }
> +; }
> +; while(index < size) {
> +; dst[index] = src[index];
> +; index++;
> +; }
> ; }
>
> -define void @__gen_memcpy_gg(i8 addrspace(1)* %dst, i8
> addrspace(1)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_gg(i8 addrspace(1)* %dst, i8 addrspace(1)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(1)* %add.ptr to i32 addrspace(1)*
> - %1 = load i32 addrspace(1)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(1)* %add.ptr1 to i32 addrspace(1)*
> - store i32 %1, i32 addrspace(1)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.1
> - %3 = load i8 addrspace(1)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(1)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(1)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(1)*
> + %3 = bitcast i8 addrspace(1)* %2 to i32 addrspace(1)*
> + %4 = load i32 addrspace(1)* %3, align 4
> + %5 = ptrtoint i8 addrspace(1)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(1)*
> + %8 = bitcast i8 addrspace(1)* %7 to i32 addrspace(1)*
> + store i32 %4, i32 addrspace(1)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(1)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(1)*
> + %12 = load i8 addrspace(1)* %11, align 1
> + %13 = ptrtoint i8 addrspace(1)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(1)*
> + store i8 %12, i8 addrspace(1)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_gp(i8 addrspace(1)* %dst, i8
> addrspace(0)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_gp(i8 addrspace(1)* %dst, i8 addrspace(0)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(0)* %add.ptr to i32 addrspace(0)*
> - %1 = load i32 addrspace(0)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(1)* %add.ptr1 to i32 addrspace(1)*
> - store i32 %1, i32 addrspace(1)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.1
> - %3 = load i8 addrspace(0)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(1)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(0)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(0)*
> + %3 = bitcast i8 addrspace(0)* %2 to i32 addrspace(0)*
> + %4 = load i32 addrspace(0)* %3, align 4
> + %5 = ptrtoint i8 addrspace(1)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(1)*
> + %8 = bitcast i8 addrspace(1)* %7 to i32 addrspace(1)*
> + store i32 %4, i32 addrspace(1)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(0)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(0)*
> + %12 = load i8 addrspace(0)* %11, align 1
> + %13 = ptrtoint i8 addrspace(1)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(1)*
> + store i8 %12, i8 addrspace(1)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_gl(i8 addrspace(1)* %dst, i8
> addrspace(3)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_gl(i8 addrspace(1)* %dst, i8 addrspace(3)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(3)* %add.ptr to i32 addrspace(3)*
> - %1 = load i32 addrspace(3)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(1)* %add.ptr1 to i32 addrspace(1)*
> - store i32 %1, i32 addrspace(1)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.1
> - %3 = load i8 addrspace(3)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(1)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(3)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(3)*
> + %3 = bitcast i8 addrspace(3)* %2 to i32 addrspace(3)*
> + %4 = load i32 addrspace(3)* %3, align 4
> + %5 = ptrtoint i8 addrspace(1)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(1)*
> + %8 = bitcast i8 addrspace(1)* %7 to i32 addrspace(1)*
> + store i32 %4, i32 addrspace(1)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(3)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(3)*
> + %12 = load i8 addrspace(3)* %11, align 1
> + %13 = ptrtoint i8 addrspace(1)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(1)*
> + store i8 %12, i8 addrspace(1)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_pg(i8 addrspace(0)* %dst, i8
> addrspace(1)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_pg(i8 addrspace(0)* %dst, i8 addrspace(1)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(1)* %add.ptr to i32 addrspace(1)*
> - %1 = load i32 addrspace(1)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(0)* %add.ptr1 to i32 addrspace(0)*
> - store i32 %1, i32 addrspace(0)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.1
> - %3 = load i8 addrspace(1)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(0)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(1)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(1)*
> + %3 = bitcast i8 addrspace(1)* %2 to i32 addrspace(1)*
> + %4 = load i32 addrspace(1)* %3, align 4
> + %5 = ptrtoint i8 addrspace(0)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(0)*
> + %8 = bitcast i8 addrspace(0)* %7 to i32 addrspace(0)*
> + store i32 %4, i32 addrspace(0)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(1)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(1)*
> + %12 = load i8 addrspace(1)* %11, align 1
> + %13 = ptrtoint i8 addrspace(0)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(0)*
> + store i8 %12, i8 addrspace(0)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_pp(i8 addrspace(0)* %dst, i8
> addrspace(0)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_pp(i8 addrspace(0)* %dst, i8 addrspace(0)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(0)* %add.ptr to i32 addrspace(0)*
> - %1 = load i32 addrspace(0)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(0)* %add.ptr1 to i32 addrspace(0)*
> - store i32 %1, i32 addrspace(0)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.1
> - %3 = load i8 addrspace(0)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(0)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(0)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(0)*
> + %3 = bitcast i8 addrspace(0)* %2 to i32 addrspace(0)*
> + %4 = load i32 addrspace(0)* %3, align 4
> + %5 = ptrtoint i8 addrspace(0)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(0)*
> + %8 = bitcast i8 addrspace(0)* %7 to i32 addrspace(0)*
> + store i32 %4, i32 addrspace(0)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(0)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(0)*
> + %12 = load i8 addrspace(0)* %11, align 1
> + %13 = ptrtoint i8 addrspace(0)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(0)*
> + store i8 %12, i8 addrspace(0)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_pl(i8 addrspace(0)* %dst, i8
> addrspace(3)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_pl(i8 addrspace(0)* %dst, i8 addrspace(3)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(3)* %add.ptr to i32 addrspace(3)*
> - %1 = load i32 addrspace(3)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(0)* %add.ptr1 to i32 addrspace(0)*
> - store i32 %1, i32 addrspace(0)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.1
> - %3 = load i8 addrspace(3)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(0)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(0)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(3)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(3)*
> + %3 = bitcast i8 addrspace(3)* %2 to i32 addrspace(3)*
> + %4 = load i32 addrspace(3)* %3, align 4
> + %5 = ptrtoint i8 addrspace(0)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(0)*
> + %8 = bitcast i8 addrspace(0)* %7 to i32 addrspace(0)*
> + store i32 %4, i32 addrspace(0)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(3)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(3)*
> + %12 = load i8 addrspace(3)* %11, align 1
> + %13 = ptrtoint i8 addrspace(0)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(0)*
> + store i8 %12, i8 addrspace(0)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_lg(i8 addrspace(3)* %dst, i8
> addrspace(1)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_lg(i8 addrspace(3)* %dst, i8 addrspace(1)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(1)* %add.ptr to i32 addrspace(1)*
> - %1 = load i32 addrspace(1)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(3)* %add.ptr1 to i32 addrspace(3)*
> - store i32 %1, i32 addrspace(3)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(1)* %src, i32 %index.1
> - %3 = load i8 addrspace(1)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(3)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(1)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(1)*
> + %3 = bitcast i8 addrspace(1)* %2 to i32 addrspace(1)*
> + %4 = load i32 addrspace(1)* %3, align 4
> + %5 = ptrtoint i8 addrspace(3)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(3)*
> + %8 = bitcast i8 addrspace(3)* %7 to i32 addrspace(3)*
> + store i32 %4, i32 addrspace(3)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(1)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(1)*
> + %12 = load i8 addrspace(1)* %11, align 1
> + %13 = ptrtoint i8 addrspace(3)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(3)*
> + store i8 %12, i8 addrspace(3)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_lp(i8 addrspace(3)* %dst, i8
> addrspace(0)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_lp(i8 addrspace(3)* %dst, i8 addrspace(0)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(0)* %add.ptr to i32 addrspace(0)*
> - %1 = load i32 addrspace(0)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(3)* %add.ptr1 to i32 addrspace(3)*
> - store i32 %1, i32 addrspace(3)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(0)* %src, i32 %index.1
> - %3 = load i8 addrspace(0)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(3)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(0)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(0)*
> + %3 = bitcast i8 addrspace(0)* %2 to i32 addrspace(0)*
> + %4 = load i32 addrspace(0)* %3, align 4
> + %5 = ptrtoint i8 addrspace(3)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(3)*
> + %8 = bitcast i8 addrspace(3)* %7 to i32 addrspace(3)*
> + store i32 %4, i32 addrspace(3)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(0)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(0)*
> + %12 = load i8 addrspace(0)* %11, align 1
> + %13 = ptrtoint i8 addrspace(3)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(3)*
> + store i8 %12, i8 addrspace(3)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
>
> -define void @__gen_memcpy_ll(i8 addrspace(3)* %dst, i8
> addrspace(3)* %src, i32 %size) nounwind alwaysinline {
> +define void @__gen_memcpy_ll(i8 addrspace(3)* %dst, i8 addrspace(3)*
> +%src, i32 %size, i32 %alignment) nounwind alwaysinline {
> entry:
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond3, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.0
> - %0 = bitcast i8 addrspace(3)* %add.ptr to i32 addrspace(3)*
> - %1 = load i32 addrspace(3)* %0, align 4
> - %add.ptr1 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.0
> - %2 = bitcast i8 addrspace(3)* %add.ptr1 to i32 addrspace(3)*
> - store i32 %1, i32 addrspace(3)* %2, align 4
> - br label %while.cond
> -
> -while.cond3: ; preds
> = %while.cond, %while.body5
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body5 ]
> - %cmp4 = icmp ult i32 %index.1, %size
> - br i1 %cmp4, label %while.body5, label %while.end7
> -
> -while.body5: ; preds
> = %while.cond3
> - %arrayidx = getelementptr inbounds i8 addrspace(3)* %src, i32 %index.1
> - %3 = load i8 addrspace(3)* %arrayidx, align 1
> - %arrayidx6 = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.1
> - store i8 %3, i8 addrspace(3)* %arrayidx6, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond3
> -
> -while.end7: ; preds
> = %while.cond3
> + %rem = and i32 %alignment, 3
> + %cmp = icmp ne i32 %rem, 0
> + %cmp113 = icmp ult i32 %size, 4
> + %or.cond = or i1 %cmp, %cmp113
> + br i1 %or.cond, label %while.cond4.preheader, label %while.body
> +
> +while.cond4.preheader: ; preds
> = %entry, %while.body
> + %index.1.ph = phi i32 [ 0, %entry ], [ %add15, %while.body ]
> + %cmp511 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp511, label %while.body6, label %while.end8
> +
> +while.body: ; preds
> = %entry, %while.body
> + %add15 = phi i32 [ %add, %while.body ], [ 4, %entry ]
> + %index.014 = phi i32 [ %add15, %while.body ], [ 0, %entry ]
> + %0 = ptrtoint i8 addrspace(3)* %src to i32
> + %1 = add i32 %0, %index.014
> + %2 = inttoptr i32 %1 to i8 addrspace(3)*
> + %3 = bitcast i8 addrspace(3)* %2 to i32 addrspace(3)*
> + %4 = load i32 addrspace(3)* %3, align 4
> + %5 = ptrtoint i8 addrspace(3)* %dst to i32
> + %6 = add i32 %5, %index.014
> + %7 = inttoptr i32 %6 to i8 addrspace(3)*
> + %8 = bitcast i8 addrspace(3)* %7 to i32 addrspace(3)*
> + store i32 %4, i32 addrspace(3)* %8, align 4
> + %add = add i32 %add15, 4
> + %cmp1 = icmp ugt i32 %add, %size
> + br i1 %cmp1, label %while.cond4.preheader, label %while.body
> +
> +while.body6: ; preds
> = %while.cond4.preheader, %while.body6
> + %index.112 = phi i32 [ %inc, %while.body6 ], [ %index.1.ph,
> +%while.cond4.preheader ]
> + %9 = ptrtoint i8 addrspace(3)* %src to i32
> + %10 = add i32 %9, %index.112
> + %11 = inttoptr i32 %10 to i8 addrspace(3)*
> + %12 = load i8 addrspace(3)* %11, align 1
> + %13 = ptrtoint i8 addrspace(3)* %dst to i32
> + %14 = add i32 %13, %index.112
> + %15 = inttoptr i32 %14 to i8 addrspace(3)*
> + store i8 %12, i8 addrspace(3)* %15, align 1
> + %inc = add i32 %index.112, 1
> + %cmp5 = icmp ult i32 %inc, %size
> + br i1 %cmp5, label %while.body6, label %while.end8
> +
> +while.end8: ; preds
> = %while.body6, %while.cond4.preheader
> ret void
> }
> diff --git a/backend/src/libocl/src/ocl_memset.ll
> b/backend/src/libocl/src/ocl_memset.ll
> index addf9f5..661520d 100644
> --- a/backend/src/libocl/src/ocl_memset.ll
> +++ b/backend/src/libocl/src/ocl_memset.ll
> @@ -1,127 +1,160 @@
> ;The memset's source code.
> -; INLINE_OVERLOADABLE void __gen_memset(uchar* dst, uchar val, size_t
> size) {
> +; INLINE_OVERLOADABLE void __gen_memset(__global uchar* dst, uchar
> val,
> +size_t size, uint alignment) {
> ; size_t index = 0;
> -; uint v = (val << 24) | (val << 16) | (val << 8) | val;
> -; while((index + 4) >= size) {
> -; *((uint *)(dst + index)) = v;
> -; index += 4;
> -; }
> +; uint v;
> +; if(alignment % 4 == 0) {
> +; v = (val << 24) | (val << 16) | (val << 8) | val;
> +; while((index + 4) <= size) {
> +; *((__global uint *)(dst + index)) = v;
> +; index += 4;
> +; }
> +; }
> ; while(index < size) {
> ; dst[index] = val;
> ; index++;
> ; }
> ; }
>
> -define void @__gen_memset_p(i8* %dst, i8 zeroext %val, i32 %size)
> nounwind alwaysinline {
> +define void @__gen_memset_p(i8 addrspace(0)* %dst, i8 zeroext %val, i32
> +%size, i32 %alignment) nounwind alwaysinline {
> entry:
> + %rem = and i32 %alignment, 3
> + %cmp = icmp eq i32 %rem, 0
> + br i1 %cmp, label %if.then, label %while.cond11.preheader
> +
> +if.then: ; preds = %entry
> %conv = zext i8 %val to i32
> %shl = shl nuw i32 %conv, 24
> %shl2 = shl nuw nsw i32 %conv, 16
> - %or = or i32 %shl, %shl2
> %shl4 = shl nuw nsw i32 %conv, 8
> - %or5 = or i32 %or, %shl4
> - %or7 = or i32 %or5, %conv
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond10, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8* %dst, i32 %index.0
> - %0 = bitcast i8* %add.ptr to i32*
> - store i32 %or7, i32* %0, align 4
> - br label %while.cond
> -
> -while.cond10: ; preds
> = %while.cond, %while.body13
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body13 ]
> - %cmp11 = icmp ult i32 %index.1, %size
> - br i1 %cmp11, label %while.body13, label %while.end14
> -
> -while.body13: ; preds
> = %while.cond10
> - %arrayidx = getelementptr inbounds i8* %dst, i32 %index.1
> - store i8 %val, i8* %arrayidx, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond10
> -
> -while.end14: ; preds
> = %while.cond10
> + %or = or i32 %shl2, %conv
> + %or5 = or i32 %or, %shl
> + %or7 = or i32 %or5, %shl4
> + %cmp814 = icmp ult i32 %size, 4
> + br i1 %cmp814, label %while.cond11.preheader, label %while.body
> +
> +while.cond11.preheader: ; preds
> = %if.then, %while.body, %entry
> + %index.1.ph = phi i32 [ 0, %entry ], [ 0, %if.then ], [ %add16,
> +%while.body ]
> + %cmp1212 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp1212, label %while.body14, label %while.end15
> +
> +while.body: ; preds
> = %if.then, %while.body
> + %add16 = phi i32 [ %add, %while.body ], [ 4, %if.then ]
> + %index.015 = phi i32 [ %add16, %while.body ], [ 0, %if.then ]
> + %0 = ptrtoint i8 addrspace(0)* %dst to i32
> + %1 = add i32 %0, %index.015
> + %2 = inttoptr i32 %1 to i8 addrspace(0)*
> + %3 = bitcast i8 addrspace(0)* %2 to i32 addrspace(0)*
> + store i32 %or7, i32 addrspace(0)* %3, align 4
> + %add = add i32 %add16, 4
> + %cmp8 = icmp ugt i32 %add, %size
> + br i1 %cmp8, label %while.cond11.preheader, label %while.body
> +
> +while.body14: ; preds
> = %while.cond11.preheader, %while.body14
> + %index.113 = phi i32 [ %inc, %while.body14 ], [ %index.1.ph,
> +%while.cond11.preheader ]
> + %4 = ptrtoint i8 addrspace(0)* %dst to i32
> + %5 = add i32 %4, %index.113
> + %6 = inttoptr i32 %5 to i8 addrspace(0)*
> + store i8 %val, i8 addrspace(0)* %6, align 1
> + %inc = add i32 %index.113, 1
> + %cmp12 = icmp ult i32 %inc, %size
> + br i1 %cmp12, label %while.body14, label %while.end15
> +
> +while.end15: ; preds
> = %while.body14, %while.cond11.preheader
> ret void
> }
>
> -define void @__gen_memset_g(i8 addrspace(1)* %dst, i8 zeroext %val,
> i32 %size) nounwind alwaysinline {
> +define void @__gen_memset_g(i8 addrspace(1)* %dst, i8 zeroext %val, i32
> +%size, i32 %alignment) nounwind alwaysinline {
> entry:
> + %rem = and i32 %alignment, 3
> + %cmp = icmp eq i32 %rem, 0
> + br i1 %cmp, label %if.then, label %while.cond11.preheader
> +
> +if.then: ; preds = %entry
> %conv = zext i8 %val to i32
> %shl = shl nuw i32 %conv, 24
> %shl2 = shl nuw nsw i32 %conv, 16
> - %or = or i32 %shl, %shl2
> %shl4 = shl nuw nsw i32 %conv, 8
> - %or5 = or i32 %or, %shl4
> - %or7 = or i32 %or5, %conv
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond10, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.0
> - %0 = bitcast i8 addrspace(1)* %add.ptr to i32 addrspace(1)*
> - store i32 %or7, i32 addrspace(1)* %0, align 4
> - br label %while.cond
> -
> -while.cond10: ; preds
> = %while.cond, %while.body13
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body13 ]
> - %cmp11 = icmp ult i32 %index.1, %size
> - br i1 %cmp11, label %while.body13, label %while.end14
> -
> -while.body13: ; preds
> = %while.cond10
> - %arrayidx = getelementptr inbounds i8 addrspace(1)* %dst, i32 %index.1
> - store i8 %val, i8 addrspace(1)* %arrayidx, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond10
> -
> -while.end14: ; preds
> = %while.cond10
> + %or = or i32 %shl2, %conv
> + %or5 = or i32 %or, %shl
> + %or7 = or i32 %or5, %shl4
> + %cmp814 = icmp ult i32 %size, 4
> + br i1 %cmp814, label %while.cond11.preheader, label %while.body
> +
> +while.cond11.preheader: ; preds
> = %if.then, %while.body, %entry
> + %index.1.ph = phi i32 [ 0, %entry ], [ 0, %if.then ], [ %add16,
> +%while.body ]
> + %cmp1212 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp1212, label %while.body14, label %while.end15
> +
> +while.body: ; preds
> = %if.then, %while.body
> + %add16 = phi i32 [ %add, %while.body ], [ 4, %if.then ]
> + %index.015 = phi i32 [ %add16, %while.body ], [ 0, %if.then ]
> + %0 = ptrtoint i8 addrspace(1)* %dst to i32
> + %1 = add i32 %0, %index.015
> + %2 = inttoptr i32 %1 to i8 addrspace(1)*
> + %3 = bitcast i8 addrspace(1)* %2 to i32 addrspace(1)*
> + store i32 %or7, i32 addrspace(1)* %3, align 4
> + %add = add i32 %add16, 4
> + %cmp8 = icmp ugt i32 %add, %size
> + br i1 %cmp8, label %while.cond11.preheader, label %while.body
> +
> +while.body14: ; preds
> = %while.cond11.preheader, %while.body14
> + %index.113 = phi i32 [ %inc, %while.body14 ], [ %index.1.ph,
> +%while.cond11.preheader ]
> + %4 = ptrtoint i8 addrspace(1)* %dst to i32
> + %5 = add i32 %4, %index.113
> + %6 = inttoptr i32 %5 to i8 addrspace(1)*
> + store i8 %val, i8 addrspace(1)* %6, align 1
> + %inc = add i32 %index.113, 1
> + %cmp12 = icmp ult i32 %inc, %size
> + br i1 %cmp12, label %while.body14, label %while.end15
> +
> +while.end15: ; preds
> = %while.body14, %while.cond11.preheader
> ret void
> }
>
> -define void @__gen_memset_l(i8 addrspace(3)* %dst, i8 zeroext %val,
> i32 %size) nounwind alwaysinline {
> +define void @__gen_memset_l(i8 addrspace(3)* %dst, i8 zeroext %val, i32
> +%size, i32 %alignment) nounwind alwaysinline {
> entry:
> + %rem = and i32 %alignment, 3
> + %cmp = icmp eq i32 %rem, 0
> + br i1 %cmp, label %if.then, label %while.cond11.preheader
> +
> +if.then: ; preds = %entry
> %conv = zext i8 %val to i32
> %shl = shl nuw i32 %conv, 24
> %shl2 = shl nuw nsw i32 %conv, 16
> - %or = or i32 %shl, %shl2
> %shl4 = shl nuw nsw i32 %conv, 8
> - %or5 = or i32 %or, %shl4
> - %or7 = or i32 %or5, %conv
> - br label %while.cond
> -
> -while.cond: ; preds
> = %while.body, %entry
> - %index.0 = phi i32 [ 0, %entry ], [ %add, %while.body ]
> - %add = add i32 %index.0, 4
> - %cmp = icmp ult i32 %add, %size
> - br i1 %cmp, label %while.cond10, label %while.body
> -
> -while.body: ; preds
> = %while.cond
> - %add.ptr = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.0
> - %0 = bitcast i8 addrspace(3)* %add.ptr to i32 addrspace(3)*
> - store i32 %or7, i32 addrspace(3)* %0, align 4
> - br label %while.cond
> -
> -while.cond10: ; preds
> = %while.cond, %while.body13
> - %index.1 = phi i32 [ %index.0, %while.cond ], [ %inc, %while.body13 ]
> - %cmp11 = icmp ult i32 %index.1, %size
> - br i1 %cmp11, label %while.body13, label %while.end14
> -
> -while.body13: ; preds
> = %while.cond10
> - %arrayidx = getelementptr inbounds i8 addrspace(3)* %dst, i32 %index.1
> - store i8 %val, i8 addrspace(3)* %arrayidx, align 1
> - %inc = add i32 %index.1, 1
> - br label %while.cond10
> -
> -while.end14: ; preds
> = %while.cond10
> + %or = or i32 %shl2, %conv
> + %or5 = or i32 %or, %shl
> + %or7 = or i32 %or5, %shl4
> + %cmp814 = icmp ult i32 %size, 4
> + br i1 %cmp814, label %while.cond11.preheader, label %while.body
> +
> +while.cond11.preheader: ; preds
> = %if.then, %while.body, %entry
> + %index.1.ph = phi i32 [ 0, %entry ], [ 0, %if.then ], [ %add16,
> +%while.body ]
> + %cmp1212 = icmp ult i32 %index.1.ph, %size
> + br i1 %cmp1212, label %while.body14, label %while.end15
> +
> +while.body: ; preds
> = %if.then, %while.body
> + %add16 = phi i32 [ %add, %while.body ], [ 4, %if.then ]
> + %index.015 = phi i32 [ %add16, %while.body ], [ 0, %if.then ]
> + %0 = ptrtoint i8 addrspace(3)* %dst to i32
> + %1 = add i32 %0, %index.015
> + %2 = inttoptr i32 %1 to i8 addrspace(3)*
> + %3 = bitcast i8 addrspace(3)* %2 to i32 addrspace(3)*
> + store i32 %or7, i32 addrspace(3)* %3, align 4
> + %add = add i32 %add16, 4
> + %cmp8 = icmp ugt i32 %add, %size
> + br i1 %cmp8, label %while.cond11.preheader, label %while.body
> +
> +while.body14: ; preds
> = %while.cond11.preheader, %while.body14
> + %index.113 = phi i32 [ %inc, %while.body14 ], [ %index.1.ph,
> +%while.cond11.preheader ]
> + %4 = ptrtoint i8 addrspace(3)* %dst to i32
> + %5 = add i32 %4, %index.113
> + %6 = inttoptr i32 %5 to i8 addrspace(3)*
> + store i8 %val, i8 addrspace(3)* %6, align 1
> + %inc = add i32 %index.113, 1
> + %cmp12 = icmp ult i32 %inc, %size
> + br i1 %cmp12, label %while.body14, label %while.end15
> +
> +while.end15: ; preds
> = %while.body14, %while.cond11.preheader
> ret void
> }
> diff --git a/backend/src/llvm/llvm_intrinsic_lowering.cpp
> b/backend/src/llvm/llvm_intrinsic_lowering.cpp
> index 7d04318..1466de0 100644
> --- a/backend/src/llvm/llvm_intrinsic_lowering.cpp
> +++ b/backend/src/llvm/llvm_intrinsic_lowering.cpp
> @@ -126,14 +126,17 @@ namespace gbe {
> Type *IntPtr = TD.getIntPtrType(Context);
> Value *Size =
> Builder.CreateIntCast(CI->getArgOperand(2), IntPtr,
> /* isSigned
> */ false);
> - Value *Ops[3];
> + Value *alignment =
> Builder.CreateIntCast(CI->getArgOperand(3), IntPtr,
> + /*
> isSigned */ false);
> + Value *Ops[4];
> Ops[0] = CI->getArgOperand(0);
> Ops[1] = CI->getArgOperand(1);
> Ops[2] = Size;
> + Ops[3] = alignment;
> char name[16] = "__gen_memcpy_xx";
> name[13] = convertSpaceToName(Ops[0]);
> name[14] = convertSpaceToName(Ops[1]);
> - replaceCallWith(name, CI, Ops, Ops+3,
> Type::getVoidTy(Context));
> + replaceCallWith(name, CI, Ops, Ops+4,
> + Type::getVoidTy(Context));
> break;
> }
> case Intrinsic::memset: { @@ -143,14 +146,17 @@
> namespace gbe {
> Type *IntPtr = TD.getIntPtrType(Op0->getType());
> Value *Size =
> Builder.CreateIntCast(CI->getArgOperand(2), IntPtr,
> /* isSigned
> */ false);
> - Value *Ops[3];
> + Value *alignment =
> Builder.CreateIntCast(CI->getArgOperand(3), IntPtr,
> + /*
> isSigned */ false);
> + Value *Ops[4];
> Ops[0] = Op0;
> // Extend the amount to i32.
> Ops[1] = val;
> Ops[2] = Size;
> + Ops[3] = alignment;
> char name[16] = "__gen_memset_x";
> name[13] = convertSpaceToName(Ops[0]);
> - replaceCallWith(name, CI, Ops, Ops+3,
> Type::getVoidTy(Context));
> + replaceCallWith(name, CI, Ops, Ops+4,
> + Type::getVoidTy(Context));
> break;
> }
> default:
> --
> 1.8.3.2
>
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet
More information about the Beignet
mailing list