ac/llvm: add support for texturing with clamped LOD

[mesa.git] / src / amd / llvm / ac_llvm_build.c

diff --git a/src/amd/llvm/ac_llvm_build.c b/src/amd/llvm/ac_llvm_build.c
index eab1b7fb721a1b746db688238560c224f2764bb1..ebcb91bd4b3374c02d1bb60e19773d15c4822cdd 100644 (file)
--- a/src/amd/llvm/ac_llvm_build.c
+++ b/src/amd/llvm/ac_llvm_build.c
@@ -144,7 +144,7 @@ int
 ac_get_llvm_num_components(LLVMValueRef value)
 {
        LLVMTypeRef type = LLVMTypeOf(value);
-       unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
+       unsigned num_components = LLVMGetTypeKind(type) == LLVMFixedVectorTypeKind
                                      ? LLVMGetVectorSize(type)
                                      : 1;
        return num_components;
@@ -155,7 +155,7 @@ ac_llvm_extract_elem(struct ac_llvm_context *ac,
                     LLVMValueRef value,
                     int index)
 {
-       if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
+       if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMFixedVectorTypeKind) {
                assert(index == 0);
                return value;
        }
@@ -167,7 +167,7 @@ ac_llvm_extract_elem(struct ac_llvm_context *ac,
 int
 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
 {
-       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
+       if (LLVMGetTypeKind(type) == LLVMFixedVectorTypeKind)
                type = LLVMGetElementType(type);
 
        if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
@@ -206,7 +206,7 @@ ac_get_type_size(LLVMTypeRef type)
                if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
                        return 4;
                return 8;
-       case LLVMVectorTypeKind:
+       case LLVMFixedVectorTypeKind:
                return LLVMGetVectorSize(type) *
                       ac_get_type_size(LLVMGetElementType(type));
        case LLVMArrayTypeKind:
@@ -235,7 +235,7 @@ static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeR
 LLVMTypeRef
 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
 {
-       if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
+       if (LLVMGetTypeKind(t) == LLVMFixedVectorTypeKind) {
                LLVMTypeRef elem_type = LLVMGetElementType(t);
                return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
                                      LLVMGetVectorSize(t));
@@ -290,7 +290,7 @@ static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef
 LLVMTypeRef
 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
 {
-       if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
+       if (LLVMGetTypeKind(t) == LLVMFixedVectorTypeKind) {
                LLVMTypeRef elem_type = LLVMGetElementType(t);
                return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
                                      LLVMGetVectorSize(t));
@@ -352,7 +352,7 @@ void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
 
        assert(bufsize >= 8);
 
-       if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
+       if (LLVMGetTypeKind(type) == LLVMFixedVectorTypeKind) {
                int ret = snprintf(buf, bufsize, "v%u",
                                        LLVMGetVectorSize(type));
                if (ret < 0) {
@@ -627,7 +627,7 @@ ac_build_expand(struct ac_llvm_context *ctx,
        LLVMTypeRef elemtype;
        LLVMValueRef chan[dst_channels];
 
-       if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
+       if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFixedVectorTypeKind) {
                unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
 
                if (src_channels == dst_channels && vec_size == dst_channels)
@@ -2373,6 +2373,9 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
               (a->lod ? 1 : 0) +
               (a->level_zero ? 1 : 0) +
               (a->derivs[0] ? 1 : 0) <= 1);
+       assert((a->min_lod ? 1 : 0) +
+              (a->lod ? 1 : 0) +
+              (a->level_zero ? 1 : 0) <= 1);
 
        if (a->opcode == ac_image_get_lod) {
                switch (dim) {
@@ -2428,6 +2431,9 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
                args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
        if (a->lod)
                args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
+       if (a->min_lod)
+               args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
+
        overload[num_overloads++] = sample ? ".f32" : ".i32";
 
        args[num_args++] = a->resource;
@@ -2481,7 +2487,7 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
        char intr_name[96];
        snprintf(intr_name, sizeof(intr_name),
                 "llvm.amdgcn.image.%s%s" /* base name */
-                "%s%s%s" /* sample/gather modifiers */
+                "%s%s%s%s" /* sample/gather modifiers */
                 ".%s.%s%s%s%s", /* dimension and type overloads */
                 name, atomic_subop,
                 a->compare ? ".c" : "",
@@ -2489,6 +2495,7 @@ LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
                 lod_suffix ? ".l" :
                 a->derivs[0] ? ".d" :
                 a->level_zero ? ".lz" : "",
+                a->min_lod ? ".cl" : "",
                 a->offset ? ".o" : "",
                 dimname,
                 atomic ? "i32" : "v4f32",
@@ -3125,9 +3132,6 @@ void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
 
                        target -= V_008DFC_SQ_EXP_PARAM;
 
-                       if ((1u << target) & skip_output_mask)
-                               continue;
-
                        /* Parse the instruction. */
                        memset(&exp, 0, sizeof(exp));
                        exp.offset = target;
@@ -3151,12 +3155,13 @@ void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
                        }
 
                        /* Eliminate constant and duplicated PARAM exports. */
-                       if (ac_eliminate_const_output(vs_output_param_offset,
-                                                     num_outputs, &exp) ||
-                           ac_eliminate_duplicated_output(ctx,
-                                                          vs_output_param_offset,
-                                                          num_outputs, &exports,
-                                                          &exp)) {
+                       if (!((1u << target) & skip_output_mask) &&
+                            (ac_eliminate_const_output(vs_output_param_offset,
+                                                      num_outputs, &exp) ||
+                            ac_eliminate_duplicated_output(ctx,
+                                                           vs_output_param_offset,
+                                                           num_outputs, &exports,
+                                                           &exp))) {
                                removed_any = true;
                        } else {
                                exports.exp[exports.num++] = exp;