[Mesa-dev] [PATCH 2/5] ac: add helpers for fast integer division by a constant

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

---
 src/amd/common/ac_llvm_build.c | 44 ++++++++++++++++++++++++++++++++++++++++++
 src/amd/common/ac_llvm_build.h | 13 +++++++++++++
 2 files changed, 57 insertions(+)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index ab0ba09..793ff81 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -571,20 +571,64 @@ ac_build_fdiv(struct ac_llvm_context *ctx,
 	 */
 	LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, ctx->f32_1, den, "");
 	LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
 
 	/* Use v_rcp_f32 instead of precise division. */
 	if (!LLVMIsConstant(ret))
 		LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
 	return ret;
 }
 
+/* See fast_idiv_by_const.h. */
+/* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
+				LLVMValueRef num,
+				LLVMValueRef multiplier,
+				LLVMValueRef pre_shift,
+				LLVMValueRef post_shift,
+				LLVMValueRef increment)
+{
+	LLVMBuilderRef builder = ctx->builder;
+
+	num = LLVMBuildLShr(builder, num, pre_shift, "");
+	num = LLVMBuildMul(builder,
+			   LLVMBuildZExt(builder, num, ctx->i64, ""),
+			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+	num = LLVMBuildAdd(builder, num,
+			   LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
+	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+	return LLVMBuildLShr(builder, num, post_shift, "");
+}
+
+/* See fast_idiv_by_const.h. */
+/* If num != UINT_MAX, this more efficient version can be used. */
+/* Set: increment = util_fast_udiv_info::increment; */
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
+				    LLVMValueRef num,
+				    LLVMValueRef multiplier,
+				    LLVMValueRef pre_shift,
+				    LLVMValueRef post_shift,
+				    LLVMValueRef increment)
+{
+	LLVMBuilderRef builder = ctx->builder;
+
+	num = LLVMBuildLShr(builder, num, pre_shift, "");
+	num = LLVMBuildNUWAdd(builder, num, increment, "");
+	num = LLVMBuildMul(builder,
+			   LLVMBuildZExt(builder, num, ctx->i64, ""),
+			   LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
+	num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
+	num = LLVMBuildTrunc(builder, num, ctx->i32, "");
+	return LLVMBuildLShr(builder, num, post_shift, "");
+}
+
 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
  * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
  * already multiplied by two. id is the cube face number.
  */
 struct cube_selection_coords {
 	LLVMValueRef stc[2];
 	LLVMValueRef ma;
 	LLVMValueRef id;
 };
 
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 0df9234..722faad 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -172,20 +172,33 @@ ac_build_gather_values(struct ac_llvm_context *ctx,
 		       unsigned value_count);
 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 				     LLVMValueRef value,
 				     unsigned num_channels);
 
 LLVMValueRef
 ac_build_fdiv(struct ac_llvm_context *ctx,
 	      LLVMValueRef num,
 	      LLVMValueRef den);
 
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
+				LLVMValueRef num,
+				LLVMValueRef multiplier,
+				LLVMValueRef pre_shift,
+				LLVMValueRef post_shift,
+				LLVMValueRef increment);
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
+				    LLVMValueRef num,
+				    LLVMValueRef multiplier,
+				    LLVMValueRef pre_shift,
+				    LLVMValueRef post_shift,
+				    LLVMValueRef increment);
+
 void
 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 		       bool is_deriv, bool is_array, bool is_lod,
 		       LLVMValueRef *coords_arg,
 		       LLVMValueRef *derivs_arg);
 
 
 LLVMValueRef
 ac_build_fs_interp(struct ac_llvm_context *ctx,
 		   LLVMValueRef llvm_chan,
-- 
2.7.4