[Mesa-dev] [PATCH 1/5] ac: properly document a buffer.store LLVM workaround

[Samuel Pitoiset]

2-5 are:

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>

On 10/03/2017 07:30 PM, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
> 
> ---
>   src/amd/common/ac_llvm_build.c | 13 ++++++++-----
>   src/amd/common/ac_llvm_build.h |  2 +-
>   2 files changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
> index 51fb009..0cebaa7 100644
> --- a/src/amd/common/ac_llvm_build.c
> +++ b/src/amd/common/ac_llvm_build.c
> @@ -753,42 +753,45 @@ void
>   ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
>   			    LLVMValueRef rsrc,
>   			    LLVMValueRef vdata,
>   			    unsigned num_channels,
>   			    LLVMValueRef voffset,
>   			    LLVMValueRef soffset,
>   			    unsigned inst_offset,
>   			    bool glc,
>   			    bool slc,
>   			    bool writeonly_memory,
> -			    bool has_add_tid)
> +			    bool swizzle_enable_hint)
>   {
> -	/* TODO: Fix stores with ADD_TID and remove the "has_add_tid" flag. */
> -	if (!has_add_tid) {
> +	/* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
> +	 * (voffset is swizzled, but soffset isn't swizzled).
> +	 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
> +	 */
> +	if (!swizzle_enable_hint) {
>   		/* Split 3 channel stores, becase LLVM doesn't support 3-channel
>   		 * intrinsics. */
>   		if (num_channels == 3) {
>   			LLVMValueRef v[3], v01;
>   
>   			for (int i = 0; i < 3; i++) {
>   				v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
>   						LLVMConstInt(ctx->i32, i, 0), "");
>   			}
>   			v01 = ac_build_gather_values(ctx, v, 2);
>   
>   			ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
>   						    soffset, inst_offset, glc, slc,
> -						    writeonly_memory, has_add_tid);
> +						    writeonly_memory, swizzle_enable_hint);
>   			ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
>   						    soffset, inst_offset + 8,
>   						    glc, slc,
> -						    writeonly_memory, has_add_tid);
> +						    writeonly_memory, swizzle_enable_hint);
>   			return;
>   		}
>   
>   		unsigned func = CLAMP(num_channels, 1, 3) - 1;
>   		static const char *types[] = {"f32", "v2f32", "v4f32"};
>   		char name[256];
>   		LLVMValueRef offset = soffset;
>   
>   		if (inst_offset)
>   			offset = LLVMBuildAdd(ctx->builder, offset,
> diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
> index 1d6dc0a..ac8ea9c 100644
> --- a/src/amd/common/ac_llvm_build.h
> +++ b/src/amd/common/ac_llvm_build.h
> @@ -163,21 +163,21 @@ void
>   ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
>   			    LLVMValueRef rsrc,
>   			    LLVMValueRef vdata,
>   			    unsigned num_channels,
>   			    LLVMValueRef voffset,
>   			    LLVMValueRef soffset,
>   			    unsigned inst_offset,
>   		            bool glc,
>   		            bool slc,
>   			    bool writeonly_memory,
> -			    bool has_add_tid);
> +			    bool swizzle_enable_hint);
>   LLVMValueRef
>   ac_build_buffer_load(struct ac_llvm_context *ctx,
>   		     LLVMValueRef rsrc,
>   		     int num_channels,
>   		     LLVMValueRef vindex,
>   		     LLVMValueRef voffset,
>   		     LLVMValueRef soffset,
>   		     unsigned inst_offset,
>   		     unsigned glc,
>   		     unsigned slc,
>