gallivm: fix indirect addressing of constant buffer

[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_conv.c

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
index 3f7f2ebde9ce700b99e88e50ae0b8249b43ebb97..77012f1fac62ec524273399bcab1c30fac333761 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
@@ -83,6 +83,9 @@
  *
  * Although the result values can be scaled to an arbitrary bit width specified
  * by dst_width, the actual result type will have the same width.
+ *
+ * Ex: src = { float, float, float, float }
+ * return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
  */
 LLVMValueRef
 lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
@@ -152,6 +155,8 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
 
 /**
  * Inverse of lp_build_clamped_float_to_unsigned_norm above.
+ * Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
+ * return {float, float, float, float} with values in range [0, 1].
  */
 LLVMValueRef
 lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
@@ -219,18 +224,19 @@ lp_build_conv(LLVMBuilderRef builder,
    unsigned num_tmps;
    unsigned i;
 
-   /* Register width must remain constant */
-   assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
-
    /* We must not loose or gain channels. Only precision */
    assert(src_type.length * num_srcs == dst_type.length * num_dsts);
 
    assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
    assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
+   assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
+   assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
 
    tmp_type = src_type;
-   for(i = 0; i < num_srcs; ++i)
+   for(i = 0; i < num_srcs; ++i) {
+      assert(lp_check_value(src_type, src[i]));
       tmp[i] = src[i];
+   }
    num_tmps = num_srcs;
 
    /*
@@ -326,30 +332,25 @@ lp_build_conv(LLVMBuilderRef builder,
 
    /*
     * Truncate or expand bit width
+    *
+    * No data conversion should happen here, although the sign bits are
+    * crucial to avoid bad clamping.
     */
 
-   assert(!tmp_type.floating || tmp_type.width == dst_type.width);
+   {
+      struct lp_type new_type;
 
-   if(tmp_type.width > dst_type.width) {
-      assert(num_dsts == 1);
-      tmp[0] = lp_build_pack(builder, tmp_type, dst_type, TRUE, tmp, num_tmps);
-      tmp_type.width = dst_type.width;
-      tmp_type.length = dst_type.length;
-      num_tmps = 1;
-   }
+      new_type = tmp_type;
+      new_type.sign   = dst_type.sign;
+      new_type.width  = dst_type.width;
+      new_type.length = dst_type.length;
+
+      lp_build_resize(builder, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
 
-   if(tmp_type.width < dst_type.width) {
-      assert(num_tmps == 1);
-      lp_build_unpack(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts);
-      tmp_type.width = dst_type.width;
-      tmp_type.length = dst_type.length;
+      tmp_type = new_type;
       num_tmps = num_dsts;
    }
 
-   assert(tmp_type.width == dst_type.width);
-   assert(tmp_type.length == dst_type.length);
-   assert(num_tmps == num_dsts);
-
    /*
     * Scale to the widest range
     */
@@ -406,8 +407,10 @@ lp_build_conv(LLVMBuilderRef builder,
        }
     }
 
-   for(i = 0; i < num_dsts; ++i)
+   for(i = 0; i < num_dsts; ++i) {
       dst[i] = tmp[i];
+      assert(lp_check_value(dst_type, dst[i]));
+   }
 }