[WebAssembly] Added test for inline assembly roundtrip.

[aardappel]

Summary

Due to previous work to make WebAssembly MC by default stack-only
inline assembly now "just works" (previously it didn't since it had
no way to know types of registers), so no further work required.

So far we only have tests (in inline-asm.ll) which test with
non-existing instructions, so this adds a test that roundtrips
both the inline assembly and its surrounding code thru the assembler.

[sbc100]

Sweet!

I kind of prefer the test to be as minimal as possible even if derived from c source, but perhaps thats overly pedantic.

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

21

drop this line

23

Drop the "-wasm": here and in the command line, like the other tests in this dir

38

Hmm.. is main compiled by clang as hidden? In any case, you can drop that for the test

54

Can you trip all this attribute stuff to keep the test nice and small?

[aheejin]

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

38

Maybe local_unnamed_addr too

54

I usually delete all the attributes unless they are necessary for the current test, but maybe it's because I'm pedantic too

[aheejin]

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

54

Oh, these are not attributes but debug info.. but anyway, maybe we need neither of them here

[aardappel]

Addressing code review comments: simplified test.

[aardappel]

Simplified test..

[aheejin]

If we are using {{$}} on every line, we can probably replace that with FileCheck's [[ https://llvm.org/docs/CommandGuide/FileCheck.html#cmdoption-match-full-lines | --match-full-lines option ]].

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

16

Indent these five lines to match the top (?

[aardappel]

More code review fixes.

[aardappel]

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

16

I had originally indented them such that the instructions line up, since the first doesn't have a \t. But that's probably unconventional :)

[aheejin]

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

35

Here the compiler happens to place src and dst in the same local, but I guess that can't be guaranteed. How about changing them to variables, like

; CHECK-NEXT:  i32.const  1
; CHECK-NEXT:  set_local  [[SRC:[0-9]+]]
; CHECK-NEXT:  i32.const  2
; CHECK-NEXT:  get_local  [[SRC]]
; CHECK-NEXT:  i32.add
; CHECK-NEXT:  set_local  [[DST:[0-9]+]]
; CHECK-NEXT:  get_local  [[DST]]
...

[aardappel]

Replaced hard-coded local indices by variables.

[aheejin]

Thanks!

test/CodeGen/WebAssembly/inline-asm-roundtrip.ll

+; RUN: llc < %s | llvm-mc -triple=wasm32-unknown-unknown | FileCheck %s
+
+; Test basic inline assembly can actually be assembled by the assembler.
+
+; .ll code below is the result of this code run thru
+; clang -target wasm32-unknown-unknown-wasm -O2 -S -emit-llvm test.c
+
+; int main(int argc, const char *argv[]) {
+;   int src = 1;
+;   int dst;
+;   asm ("i32.const\t2\n"
+;      "\tget_local\t%1\n"
+;      "\ti32.add\n"
+;      "\tset_local\t%0"
+;       : "=r" (dst)
+;       : "r" (src));
+;   return dst != 3;
+; }
+
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-unknown"
+
+; CHECK-LABEL: main:
+; CHECK-NEXT: .param    i32, i32{{$}}
+; CHECK-NEXT: .local    i32{{$}}
+; CHECK-NEXT: i32.const 1{{$}}
+; CHECK-NEXT: set_local 2{{$}}
+; CHECK-NEXT: i32.const 2{{$}}
+; CHECK-NEXT: get_local 2{{$}}
+; CHECK-NEXT: i32.add {{$}}
+; CHECK-NEXT: set_local 2{{$}}
+; CHECK-NEXT: get_local 2{{$}}
+; CHECK-NEXT: i32.const 3{{$}}
+; CHECK-NEXT: i32.ne {{$}}
+
+define i32 @main(i32 %argc, i8** nocapture readnone %argv) #0 {
+entry:
+  %0 = tail call i32 asm "i32.const\092\0A\09get_local\09$1\0A\09i32.add\0A\09set_local\09$0", "=r,r"(i32 1) #1
+  %cmp = icmp ne i32 %0, 3
+  %conv = zext i1 %cmp to i32
+  ret i32 %conv
+}
+