[WebAssembly] LTO: Set POSIX thread model when linking with -shared-memoey

[sbc100]

Summary

This allows LTO to use atomics during codegen.

[dschuff]

I think we should change the logic in the backend to use the atomics target feature rather than the thread model to determine whether to run the LowerAtomics pass. Then this won't be necessary since the functions will have the appropriate attribute.

[sbc100]

Still working on a test case..

In D57800#1387400, @dschuff wrote:

I think we should change the logic in the backend to use the atomics target feature rather than the thread model to determine whether to run the LowerAtomics pass. Then this won't be necessary since the functions will have the appropriate attribute.

I agree in the long term that make sense, but for now I'd like to match clang. Currently the clang wasm toolchain has WebAssembly::getThreadModel() which specific "single" by default. Once we remove that then we can also remove this patch and possible completely from the ThreadModel LTO option.

[aheejin]

I don't think we can remove the threadmodel thing from the backend in near future (because atomics is a subtarget feature), so I suggest landing this because without this LTO wouldn't work.

[dschuff]

Yeah, based on our conversations (e.g. in D57874) we should go ahead with this.

[sbc100]

• add test

[aheejin]

test/wasm/lto/atomics.ll

4

Can this check if the wasm file contains atomic instructions? There's no CHECK lines for that...

5

Is this llc test relevant to this CL?

11

No need for -wasm?

[sbc100]

I believe we can now drop this, and instead remove the setting of C.Options.ThreadModel since the +atomics attribute on the bitcode input should be enough to lower atomics correctly, right @tlively ?

[tlively]

Yes, ThreadModel is no longer used anywhere in the WebAssembly backend.

test/wasm/lto/atomics.ll

 ; RUN: llvm-as %s -o %t.o
 ; RUN: wasm-ld %t.o -o %t.wasm -lto-O0
+; RUN: wasm-ld --shared-memory %t.o -o %t.shared.wasm -lto-O0
+; RUN: llvm-objdump -d %t.shared.wasm
+; RUN: llc %s -mattr=+atomics -o %t.threads.o
+
 ; Atomic operations with fail to compile if the ThreadModel is not
 ; correctly set to Single (i.e. if atomics are not lowered to regular ops).
 
 target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
 target triple = "wasm32-unknown-unknown-wasm"
 
 @foo = hidden global i32 1
 
-define void @_start() {
-  %1 = load atomic i32, i32* @foo unordered, align 4
+define void @_start() #0 {
+  %1 = load atomic i32, i32* @foo seq_cst, align 4
+  store atomic i32 %1, i32* @foo seq_cst, align 4
   ret void
 }
+
+attributes #0 = { noinline optnone "target-features"="+atomics" }

wasm/LTO.cpp

 static std::unique_ptr<lto::LTO> createLTO() {
   lto::Config C;
   C.Options = initTargetOptionsFromCodeGenFlags();
 
   // Always emit a section per function/data with LTO.
   C.Options.FunctionSections = true;
   C.Options.DataSections = true;
 
-  // Wasm currently only supports ThreadModel::Single
-  C.Options.ThreadModel = ThreadModel::Single;
+  // When shared memory is disabled wasm doesn't support atomics operations.
+  // Setting the thead model based on the SharedMemory option seems like
+  // reasonable default.  Idealy the bitcode metadata would set this on
+  // a per function basis instead.
+  C.Options.ThreadModel =
+      Config->SharedMemory ? ThreadModel::POSIX : ThreadModel::Single;
 
   C.DisableVerify = Config->DisableVerify;
   C.DiagHandler = diagnosticHandler;
   C.OptLevel = Config->LTOO;
   C.MAttrs = getMAttrs();
   C.CGOptLevel = args::getCGOptLevel(Config->LTOO);
 
   if (Config->Relocatable)