PPC should use byval for large byval arguments

[isanbard]

Bugzilla Link	2068
Resolution	FIXED
Resolved on	Mar 10, 2008 18:08
Version	unspecified
OS	All
Attachments	What my machine actually generates

Extended Description

2008-01-25-ByValReadNone.c is failing on my machine (PPC Mac G4 running Tiger):

FAIL: /Users/wendling/llvm/llvm.src/test/CFrontend/2008-01-25-
ByValReadNone.c
Failed with exit(1) at line 1
while running: /Users/wendling/llvm/llvm-gcc-4.2.install/bin/llvm-gcc
-emit-llvm -O3 -S -o - -emit-llvm /Users/wendling/llvm/llvm.src/test/
CFrontend/2008-01-25-ByValReadNone.c | not grep readonly
%tmp1505 = call i32 @​g( i64 %x.0.0, i64 %x.0.1, i64 %x.0.2,
i64 %x.0.3, i64 %x.0.4, i64 %x.0.5, i64 %x.0.6, i64 %x.0.7, i64 %x.
0.8, i64 %x.0.9, i64 %x.0.10, i64 %x.0.11, i64 %x.0.12, i64 %x.0.13,
i64 %x.0.14, i64 %x.0.15, i64 %x.0.16, i64 %x.0.17, i64 %x.0.18, i64 %
x.0.19, i64 %x.0.20, i64 %x.0.21, i64 %x.0.22, i64 %x.0.23, i64 %x.
0.24, i64 %x.0.25, i64 %x.0.26, i64 %x.0.27, i64 %x.0.28, i64 %x.
0.29, i64 %x.0.30, i64 %x.0.31, i64 %x.0.32, i64 %x.0.33, i64 %x.
0.34, i64 %x.0.35, i64 %x.0.36, i64 %x.0.37, i64 %x.0.38, i64 %x.
...
0.487, i64 %x.0.488, i64 %x.0.489, i64 %x.0.490, i64 %x.0.491, i64 %x.
0.492, i64 %x.0.493, i64 %x.0.494, i64 %x.0.495, i64 %x.0.496, i64 %x.
0.497, i64 %x.0.498, i64 %x.0.499 ) nounwind readonly ;
[#uses=1]
...

It appears to be passing the humongous array into the functions by individual values instead of by value.

[lattner]

This is a known issue, I think Dale is working on this, but he might be willing to let you take it from him :)

[lattner]

rdar://5718480

[llvmbot]

ppc calling convention conformance is on my list, yes. This is not a correctness issue, but is something that could reasonably be fixed at the same time.

[isanbard]

Is this the correct way to xfail a test for all but x86 platforms?

Index: 2008-01-25-ByValReadNone.c

--- 2008-01-25-ByValReadNone.c (revision 47335)
+++ 2008-01-25-ByValReadNone.c (working copy)
@@ -1,7 +1,7 @@
// RUN: %llvmgcc -O3 -S -o - -emit-llvm %s | not grep readonly
// RUN: %llvmgcc -O3 -S -o - -emit-llvm %s | not grep readnone
+// XFAIL: alpha|ia64|arm|amd64|ppc|sparc

// The struct being passed byval means that we cannot mark the
// function readnone. Readnone would allow stores to the arg to
// be deleted in the caller. We also don't allow readonly since

[llvmbot]

Could we please investigate why this is not showing up in the nightly test runs first? I think that is important and it may get dropped on the floor if we disable the test.

[isanbard]

Will do.

[llvmbot]

Working on it. I wrote above that this is not a correctness issue, but this is not exactly right, there are some cases where structs are currently passed incorrectly that will be fixed by implementing byval. They could also be fixed other ways but byval is the right way to go.

[llvmbot]

This is fixed. Several patches are involved but this is the one that turns on the machinery:
http://llvm.org/viewvc/llvm-project?view=rev&revision=48059

[lattner]

Nice! resolution -> fixed?

[llvmbot]

Yep.