[SYCL] Tie BoundArch to Toolchain #4673
Conversation
AidanBeltonS
requested review from
AGindinson,
hchilama and
mdtoguchi
as code owners
There was a problem hiding this comment.
The change looks good conceptually (although @mdtoguchi would have the final say here). The bulk of my comments is style-related - in case any serious concerns arise, I'm hoping to roll all of them out in the next iteration after fully grasping the changes.
| @@ -245,7 +245,7 @@ | |||
| // CHK-PHASE-MULTI-TARG: 3: input, "[[INPUT]]", c++, (device-sycl) | |||
| // CHK-PHASE-MULTI-TARG: 4: preprocessor, {3}, c++-cpp-output, (device-sycl) | |||
| // CHK-PHASE-MULTI-TARG: 5: compiler, {4}, ir, (device-sycl) | |||
| // CHK-PHASE-MULTI-TARG: 6: offload, "host-sycl (x86_64-unknown-linux-gnu)" {2}, "device-sycl (spir64-unknown-unknown)" {5}, c++-cpp-output | |||
| // CHK-PHASE-MULTI-TARG: 6: offload, "host-sycl (x86_64-unknown-linux-gnu)" {2}, "device-sycl (spir64_gen-unknown-unknown)" {5}, c++-cpp-output | |||
There was a problem hiding this comment.
Note to self & other reviewers: this change concerned me at first, as I've doubted the deterministic ordering of the target toolchains after the map introduction, but then realized that usage of MapVector for toolchain/arch mapping ensures that.
clang/test/Driver/sycl-offload.cpp
Outdated
| @@ -68,7 +68,7 @@ | |||
| // RUN: | FileCheck -check-prefix IMPLIED_DEVICE_OBJ %s | |||
| // RUN: %clangxx -### -target x86_64-unknown-linux-gnu -fsycl -fintelfpga %S/Inputs/SYCL/objlin64.o %s 2>&1 \ | |||
| // RUN: | FileCheck -check-prefix IMPLIED_DEVICE_OBJ %s | |||
| // IMPLIED_DEVICE_OBJ: clang-offload-bundler{{.*}} "-type=o"{{.*}} "-targets={{.*}}sycl-spir64-unknown-unknown,sycl-spir64_{{.*}}-unknown-unknown"{{.*}} "-unbundle" | |||
| // IMPLIED_DEVICE_OBJ: clang-offload-bundler{{.*}} "-type=o"{{.*}} "-targets=sycl-spir64_{{.*}}-unknown-unknown,{{.*}}sycl-spir64-unknown-unknown"{{.*}} "-unbundle" | |||
There was a problem hiding this comment.
Looks like the ordering swap isn't taking into account the host triple which is coming in at the front of the -targets
There was a problem hiding this comment.
Thanks! I have added the host triple to the start of the expression. It was only failing in windows as if there are device libraries to link with then there is a clang-offload-bundler which does not have the host triple but does have both device triples.
These changes implement SYCLTargetInfoList which stores both the toolchain and the bound arch. This prevents BoundArchs being incorrectly assigned when actions and dependencies are created. This sets up the possibility of having multiple BoundArchs per triple as a future development.
This fixes an issue where spirv and nvptx swap BoundArch's, causing failure cases. #3631
This patch should also resolve #4662
Some tests were changed in sycl-offload-with-split.c where the output from the device compilation phase was being used as an input by the wrong device. 3 tests are also introduced, two check the ordering behaviour of the triples and the bounded arch. The third checks for behaviour when targetting spirv, nvptx, and amdgcn.