amd/llvm: fix handling of unsupported vec3 loads on gfx6

VMEM loads promoted from vec3 to vec4 didn't trim the vector, thus
returning vec4 on gfx6 and vec3 on later generations, which callers
don't expect.

SMEM loads were adding an extra component on gfx6, causing same issues.

Fixes: 82919e2d - amd: lower subdword UBO loads in NIR
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/8693

Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22117>

src/amd/llvm/ac_llvm_build.c

@@ -1330,8 +1330,11 @@ static LLVMValueRef ac_build_buffer_load_common(struct ac_llvm_context *ctx, LLV
       snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s", indexing_kind, type_name);
    }
 
-   return ac_build_intrinsic(ctx, name, type, args, idx,
-                             can_speculate ? AC_ATTR_INVARIANT_LOAD : 0);
+   LLVMValueRef result = ac_build_intrinsic(ctx, name, type, args, idx,
+                                            can_speculate ? AC_ATTR_INVARIANT_LOAD : 0);
+   if (func > num_channels)
+      result = ac_trim_vector(ctx, result, num_channels);
+   return result;
 }
 
 LLVMValueRef ac_build_buffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc, int num_channels,
@@ -1369,8 +1372,6 @@ LLVMValueRef ac_build_buffer_load(struct ac_llvm_context *ctx, LLVMValueRef rsrc
       if (num_channels == 1)
          return result[0];
 
-      if (num_channels == 3 && !ac_has_vec3_support(ctx->gfx_level, false))
-         result[num_channels++] = LLVMGetUndef(channel_type);
       return ac_build_gather_values(ctx, result, num_channels);
    }