[Mesa-dev] [PATCH] radeonsi: add a workaround for weird s_buffer_load_dword behavior on SI

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

See my LLVM patch which fixes the root cause.

Users have to apply this patch and then they have 2 choices:
- Downgrade to LLVM 5.0
- Update to LLVM git after my LLVM patch is pushed.

It won't be possible to use current and earlier development version
of LLVM 6.0.
---
 src/gallium/drivers/radeonsi/si_shader.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 1320c6f..a248cea 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -2008,28 +2008,35 @@ static LLVMValueRef fetch_constant(
 	if (sel->info.const_buffers_declared == 1 &&
 	    sel->info.shader_buffers_declared == 0) {
 		LLVMValueRef ptr =
 			LLVMGetParam(ctx->main_fn, ctx->param_const_and_shader_buffers);
 
 		/* This enables use of s_load_dword and flat_load_dword for const buffer 0
 		 * loads, and up to x4 load opcode merging. However, it leads to horrible
 		 * code reducing SIMD wave occupancy from 8 to 2 in many cases.
 		 *
 		 * Using s_buffer_load_dword (x1) seems to be the best option right now.
+		 *
+		 * LLVM 5.0 on SI doesn't insert a required s_nop between SALU setting
+		 * a descriptor and s_buffer_load_dword using it, so we can't expand
+		 * the pointer into a full descriptor like below. We have to use
+		 * s_load_dword instead. The only case when LLVM 5.0 would select
+		 * s_buffer_load_dword (that we have to prevent) is when we use use
+		 * a literal offset where we don't need bounds checking.
 		 */
-#if 0 /* keep this codepath disabled */
-		if (!reg->Register.Indirect) {
+		if (ctx->screen->b.chip_class == SI &&
+                    HAVE_LLVM < 0x0600 &&
+                    !reg->Register.Indirect) {
 			addr = LLVMBuildLShr(ctx->ac.builder, addr, LLVMConstInt(ctx->i32, 2, 0), "");
 			LLVMValueRef result = ac_build_load_invariant(&ctx->ac, ptr, addr);
 			return bitcast(bld_base, type, result);
 		}
-#endif
 
 		/* Do the bounds checking with a descriptor, because
 		 * doing computation and manual bounds checking of 64-bit
 		 * addresses generates horrible VALU code with very high
 		 * VGPR usage and very low SIMD occupancy.
 		 */
 		ptr = LLVMBuildPtrToInt(ctx->ac.builder, ptr, ctx->i64, "");
 		ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->v2i32, "");
 
 		LLVMValueRef desc_elems[] = {
-- 
2.7.4