From: Nicolai Hähnle <nicolai.haeh...@amd.com>
---
src/gallium/drivers/radeon/radeon_llvm.h | 4 +++
.../drivers/radeon/radeon_setup_tgsi_llvm.c | 29 ++++++++++++++++++++++
src/gallium/drivers/radeonsi/si_shader.c | 19 +-------------
3 files changed, 34 insertions(+), 18 deletions(-)
diff --git a/src/gallium/drivers/radeon/radeon_llvm.h
b/src/gallium/drivers/radeon/radeon_llvm.h
index 0276ef3..da5b7f5 100644
--- a/src/gallium/drivers/radeon/radeon_llvm.h
+++ b/src/gallium/drivers/radeon/radeon_llvm.h
@@ -109,20 +109,24 @@ struct radeon_llvm_context {
struct gallivm_state gallivm;
};
LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type);
LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type, LLVMValueRef value);
+LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
+ LLVMValueRef index,
+ unsigned num);
+
void radeon_llvm_emit_prepare_cube_coords(struct lp_build_tgsi_context
*bld_base,
struct lp_build_emit_data *emit_data,
LLVMValueRef *coords_arg,
LLVMValueRef *derivs_arg);
void radeon_llvm_context_init(struct radeon_llvm_context *ctx,
const char *triple,
const struct tgsi_shader_info *info,
const struct tgsi_token *tokens);
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index dd7d60b..7cdf228 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -66,20 +66,49 @@ LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
if (dst_type)
return LLVMBuildBitCast(builder, value, dst_type, "");
else
return value;
}
+/**
+ * Return a value that is equal to the given i32 \p index if it lies in [0,num)
+ * or an undefined value in the same interval otherwise.
+ */
+LLVMValueRef radeon_llvm_bound_index(struct radeon_llvm_context *ctx,
+ LLVMValueRef index,
+ unsigned num)
+{
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef c_max = lp_build_const_int32(gallivm, num - 1);
+ LLVMValueRef cc;
+
+ if (util_is_power_of_two(num)) {
+ index = LLVMBuildAnd(builder, index, c_max, "");
+ } else {
+ /* In theory, this MAX pattern should result in code that is
+ * as good as the bit-wise AND above.
+ *
+ * In practice, LLVM generates worse code (at the time of
+ * writing), because its value tracking is not strong enough.
+ */
+ cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
+ index = LLVMBuildSelect(builder, cc, index, c_max, "");
+ }
+
+ return index;
+}
+
static struct radeon_llvm_loop *get_current_loop(struct radeon_llvm_context
*ctx)
{
return ctx->loop_depth > 0 ? ctx->loop + (ctx->loop_depth - 1) : NULL;
}
static struct radeon_llvm_branch *get_current_branch(struct
radeon_llvm_context *ctx)
{
return ctx->branch_depth > 0 ?
ctx->branch + (ctx->branch_depth - 1) : NULL;
}
diff --git a/src/gallium/drivers/radeonsi/si_shader.c
b/src/gallium/drivers/radeonsi/si_shader.c
index 06b5c9c..a5b566e 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -558,47 +558,30 @@ static LLVMValueRef get_indirect_index(struct
si_shader_context *ctx,
}
/**
* Like get_indirect_index, but restricts the return value to a (possibly
* undefined) value inside [0..num).
*/
static LLVMValueRef get_bounded_indirect_index(struct si_shader_context *ctx,
const struct tgsi_ind_register
*ind,
int rel_index, unsigned num)
{
- struct gallivm_state *gallivm = &ctx->radeon_bld.gallivm;
- LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef result = get_indirect_index(ctx, ind, rel_index);
- LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
- LLVMValueRef cc;
/* LLVM 3.8: If indirect resource indexing is used:
* - SI & CIK hang
* - VI crashes
*/
if (HAVE_LLVM <= 0x0308)
return LLVMGetUndef(ctx->i32);
- if (util_is_power_of_two(num)) {
- result = LLVMBuildAnd(builder, result, c_max, "");
- } else {
- /* In theory, this MAX pattern should result in code that is
- * as good as the bit-wise AND above.
- *
- * In practice, LLVM generates worse code (at the time of
- * writing), because its value tracking is not strong enough.
- */
- cc = LLVMBuildICmp(builder, LLVMIntULE, result, c_max, "");
- result = LLVMBuildSelect(builder, cc, result, c_max, "");
- }
-
- return result;
+ return radeon_llvm_bound_index(&ctx->radeon_bld, result, num);
}
/**
* Calculate a dword address given an input or output register and a stride.
*/
static LLVMValueRef get_dw_address(struct si_shader_context *ctx,
const struct tgsi_full_dst_register *dst,
const struct tgsi_full_src_register *src,
LLVMValueRef vertex_dw_stride,
--
2.7.4
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
