summaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/radeon
diff options
context:
space:
mode:
authorNicolai Hähnle <[email protected]>2016-08-08 22:15:04 +0200
committerNicolai Hähnle <[email protected]>2016-08-17 12:11:22 +0200
commitfa84296a5a7c707e87299c921acde3bd1533833f (patch)
tree379db7ead80d13538576f5e9cadf757c2ecd86cd /src/gallium/drivers/radeon
parent92b66b38c94bbcdf9633756f92c196fcfea20ac1 (diff)
gallium/radeon: clarify the comment on the array alloca heuristic
Reviewed-by: Marek Olšák <[email protected]>
Diffstat (limited to 'src/gallium/drivers/radeon')
-rw-r--r--src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c29
1 files changed, 19 insertions, 10 deletions
diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index 22ff18e9bcd..e4bfa74537f 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -444,19 +444,28 @@ static void emit_declaration(struct lp_build_tgsi_context *bld_base,
ctx->arrays[id].range = decl->Range;
- /* If the array is more than 16 elements (each element
- * is 32-bits), then store it in a vector. Storing the
- * array in a vector will causes the compiler to store
- * the array in registers and access it using indirect
- * addressing. 16 is number of vector elements that
- * LLVM will store in a register.
- * FIXME: We shouldn't need to do this. LLVM should be
- * smart enough to promote allocas int registers when
- * profitable.
+ /* If the array has more than 16 elements, store it
+ * in memory using an alloca that spans the entire
+ * array.
+ *
+ * Otherwise, store each array element individually.
+ * We will then generate vectors (per-channel, up to
+ * <4 x float>) for indirect addressing.
+ *
+ * Note that 16 is the number of vector elements that
+ * LLVM will store in a register, so theoretically an
+ * array with up to 4 * 16 = 64 elements could be
+ * handled this way, but whether that's a good idea
+ * depends on VGPR register pressure elsewhere.
+ *
+ * FIXME: We shouldn't need to have the non-alloca
+ * code path for arrays. LLVM should be smart enough to
+ * promote allocas into registers when profitable.
*/
if (decl_size > 16) {
array_alloca = LLVMBuildAlloca(builder,
- LLVMArrayType(bld_base->base.vec_type, decl_size),"array");
+ LLVMArrayType(bld_base->base.vec_type,
+ decl_size), "array");
ctx->arrays[id].alloca = array_alloca;
}
}