ac/nir: fix lsb emission

This makes it match radeonsi. The LLVM backend itself will emit the
correct instruction, but LLVM might do incorrect optimizations since it
thinks the output is undefined when the input is 0, even though it's not
supposed to be. We really need a new intrinsic, or for the backend to
become smarter and recognize this pattern.

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Bas Nieuwenhuizen <basni@google.com>

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 530b581..19b1ef0 100644 (file)
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1187,7 +1187,17 @@ static LLVMValueRef emit_find_lsb(struct ac_llvm_context *ctx,
                 */
                LLVMConstInt(ctx->i1, 1, false),
        };
-       return ac_build_intrinsic(ctx, "llvm.cttz.i32", ctx->i32, params, 2, AC_FUNC_ATTR_READNONE);
+
+       LLVMValueRef lsb = ac_build_intrinsic(ctx, "llvm.cttz.i32", ctx->i32,
+                                             params, 2,
+                                             AC_FUNC_ATTR_READNONE);
+
+       /* TODO: We need an intrinsic to skip this conditional. */
+       /* Check for zero: */
+       return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
+                                                          LLVMIntEQ, src0,
+                                                          ctx->i32_0, ""),
+                              LLVMConstInt(ctx->i32, -1, 0), lsb, "");
 }
 
 static LLVMValueRef emit_ifind_msb(struct ac_llvm_context *ctx,