summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorMarek Olšák <[email protected]>2018-09-22 21:17:52 -0400
committerMarek Olšák <[email protected]>2018-10-16 17:23:25 -0400
commitbfc795670ec82af5767cf360806e32322664604e (patch)
treeb6d2672e2e4cf42bfa1d7e10ca05959a68f92d30
parentea039f789d9b54e1bd1d644b6a29863ca3500314 (diff)
ac: add helpers for fast integer division by a constant
-rw-r--r--src/amd/common/ac_llvm_build.c61
-rw-r--r--src/amd/common/ac_llvm_build.h17
2 files changed, 78 insertions, 0 deletions
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index d0bcfe2dfb3..2d78ca1b52a 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -594,6 +594,67 @@ ac_build_fdiv(struct ac_llvm_context *ctx,
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, "");
+}
+
+/* See fast_idiv_by_const.h. */
+/* Both operands must fit in 31 bits and the divisor must not be 1. */
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
+ LLVMValueRef num,
+ LLVMValueRef multiplier,
+ LLVMValueRef post_shift)
+{
+ LLVMBuilderRef builder = ctx->builder;
+
+ 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.
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 5c3e221b2cc..f68efbc49ff 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -182,6 +182,23 @@ 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);
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
+ LLVMValueRef num,
+ LLVMValueRef multiplier,
+ LLVMValueRef post_shift);
+
void
ac_prepare_cube_coords(struct ac_llvm_context *ctx,
bool is_deriv, bool is_array, bool is_lod,