[llvm] r253902 - [WinEH] Fix problem where CodeGenPrepare incorrectly sinks a bitcast into an EH pad.

[Kaylor, Andrew via llvm-commits]

The Windows exception handling mechanism needs to have well-defined entry and exit points for exception handling code.  The new EH pad instructions have been defined in the language reference in such a way that the EH pad instruction is a terminator but must also be the first non-PHI instruction in its block as a way of preventing optimizations from placing instructions into locations that are ambiguous with regard to whether they are meant to be inside or outside of the exception handling code. 

There's a bit more discussion here:

http://llvm.org/docs/ExceptionHandling.html#exception-handling-using-the-windows-runtime

-Andy

-----Original Message-----
From: Quentin Colombet [mailto:qcolombet at apple.com] 
Sent: Monday, November 30, 2015 3:47 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com>
Cc: llvm-commits at lists.llvm.org
Subject: Re: [llvm] r253902 - [WinEH] Fix problem where CodeGenPrepare incorrectly sinks a bitcast into an EH pad.

Hi Andrew,

Out of curiosity, what is the problem of having non-phi instructions in those blocks?

Cheers,
-Quentin

> On Nov 23, 2015, at 11:16 AM, Andrew Kaylor via llvm-commits <llvm-commits at lists.llvm.org> wrote:
> 
> Author: akaylor
> Date: Mon Nov 23 13:16:15 2015
> New Revision: 253902
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=253902&view=rev
> Log:
> [WinEH] Fix problem where CodeGenPrepare incorrectly sinks a bitcast into an EH pad.
> 
> Differential Revision: http://reviews.llvm.org/D14842
> 
> 
> Added:
>    llvm/trunk/test/Transforms/CodeGenPrepare/catchpad-phi-cast.ll
> Modified:
>    llvm/trunk/lib/CodeGen/CodeGenPrepare.cpp
> 
> Modified: llvm/trunk/lib/CodeGen/CodeGenPrepare.cpp
> URL: 
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/CodeGenPrep
> are.cpp?rev=253902&r1=253901&r2=253902&view=diff
> ======================================================================
> ========
> --- llvm/trunk/lib/CodeGen/CodeGenPrepare.cpp (original)
> +++ llvm/trunk/lib/CodeGen/CodeGenPrepare.cpp Mon Nov 23 13:16:15 2015
> @@ -730,6 +730,12 @@ static bool SinkCast(CastInst *CI) {
>     // Preincrement use iterator so we don't invalidate it.
>     ++UI;
> 
> +    // If the block selected to receive the cast is an EH pad that does not
> +    // allow non-PHI instructions before the terminator, we can't sink the
> +    // cast.
> +    if (UserBB->getTerminator()->isEHPad())
> +      continue;
> +
>     // If this user is in the same block as the cast, don't change the cast.
>     if (UserBB == DefBB) continue;
> 
> 
> Added: llvm/trunk/test/Transforms/CodeGenPrepare/catchpad-phi-cast.ll
> URL: 
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/CodeGen
> Prepare/catchpad-phi-cast.ll?rev=253902&view=auto
> ======================================================================
> ========
> --- llvm/trunk/test/Transforms/CodeGenPrepare/catchpad-phi-cast.ll 
> (added)
> +++ llvm/trunk/test/Transforms/CodeGenPrepare/catchpad-phi-cast.ll Mon 
> +++ Nov 23 13:16:15 2015
> @@ -0,0 +1,59 @@
> +; RUN: opt -codegenprepare -S < %s | FileCheck %s
> +
> +; The following target lines are needed for the test to exercise what it should.
> +; Without these lines, CodeGenPrepare does not try to sink the bitcasts.
> +target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
> +target triple = "x86_64-pc-windows-msvc"
> +
> +declare i32 @__CxxFrameHandler3(...)
> +
> +declare void @f()
> +
> +declare void @g(i8*)
> +
> +; CodeGenPrepare will want to sink these bitcasts, but it selects the 
> +catchpad ; blocks as the place to which the bitcast should be sunk.  
> +Since catchpads ; do not allow non-phi instructions before the terminator, this isn't possible.
> +
> +; CHECK-LABEL: @test(
> +define void @test(i32* %addr) personality i32 (...)* 
> + at __CxxFrameHandler3 { ; CHECK: entry:
> +; CHECK-NEXT: %x = getelementptr i32, i32* %addr, i32 1 ; CHECK-NEXT: 
> +%p1 = bitcast i32* %x to i8*
> +entry:
> +  %x = getelementptr i32, i32* %addr, i32 1
> +  %p1 = bitcast i32* %x to i8*
> +  invoke void @f()
> +          to label %invoke.cont unwind label %catch1
> +
> +; CHECK: invoke.cont:
> +; CHECK-NEXT: %y = getelementptr i32, i32* %addr, i32 2 ; CHECK-NEXT: 
> +%p2 = bitcast i32* %y to i8*
> +invoke.cont:
> +  %y = getelementptr i32, i32* %addr, i32 2
> +  %p2 = bitcast i32* %y to i8*
> +  invoke void @f()
> +          to label %done unwind label %catch2
> +
> +done:
> +  ret void
> +
> +catch1:
> +  %cp1 = catchpad [] to label %catch.dispatch unwind label %catchend1
> +
> +catch2:
> +  %cp2 = catchpad [] to label %catch.dispatch unwind label %catchend2
> +
> +; CHECK: catch.dispatch:
> +; CHECK-NEXT: %p = phi i8* [ %p1, %catch1 ], [ %p2, %catch2 ]
> +catch.dispatch:
> +  %p = phi i8* [ %p1, %catch1 ], [ %p2, %catch2 ]
> +  call void @g(i8* %p)
> +  unreachable
> +
> +catchend1:
> +  catchendpad unwind to caller
> +
> +catchend2:
> +  catchendpad unwind to caller
> +}
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits