PNaCl dynamic linking: Fix handling of local relocations with addends

PNaCl dynamic linking: Fix handling of local relocations with addends

While testing ConvertToPSO on newlib's libc.a, I found that the pass
didn't work on relocations with addends that refer to module-local
variables.

The reason is that we're running FlattenGlobals twice (once in
"-convert-to-pso" and once in "-pnacl-abi-simplify-postopt"), but
FlattenGlobals isn't idempotent for this case.  FlattenGlobals can
handle a getelementptr as input, but it converts it to ptrtoint+add,
which it doesn't handle.

Fix this by changing FlattenGlobals to handle "add" ConstantExprs.

Testing:

 * convert-to-pso.ll: Test this end-to-end.

 * flatten-globals.ll: Check this specific case, and check idempotence
   in general.

   The latter requires that we use getSExtValue() (sign-extending)
   rather than getZExtValue() (zero-extending), otherwise the
   negative-addend test fails with "FlattenGlobals: Offset does not
   fit into 32 bits".

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4351
TEST=test/Transforms/NaCl/*.ll
R=phosek@chromium.org

Committed: https://chromium.googlesource.com/native_client/pnacl-llvm/+/aa09f2c4ce3ced974ef25e004106345daaaebd04

[Mark Seaborn, 2016-02-12 19:18:52]

Description was changed from

==========
PNaCl: Make the FlattenGlobals pass idempotent
==========

to

==========
PNaCl dynamic linking: Fix handling of local relocations with addends

While testing ConvertToPSO on newlib's libc.a, I found that the pass
didn't work on relocations with addends that refer to module-local
variables.

The reason is that we're running FlattenGlobals twice (once in
"-convert-to-pso" and once in "-pnacl-abi-simplify-postopt"), but
FlattenGlobals isn't idempotent for this case.  FlattenGlobals can
handle a getelementptr as input, but it converts it to ptrtoint+add,
which it doesn't handle.

Fix this by changing FlattenGlobals to handle "add" ConstantExprs.

Testing:

 * convert-to-pso.ll: Test this end-to-end.

 * flatten-globals.ll: Check this specific case, and check idempotence
   in general.

   The latter requires that we use getSExtValue() (sign-extending)
   rather than getZExtValue() (zero-extending), otherwise the
   negative-addend test fails with "FlattenGlobals: Offset does not
   fit into 32 bits".

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4351
TEST=test/Transforms/NaCl/*.ll
==========

[Mark Seaborn, 2016-02-12 19:18:52]

mseaborn@chromium.org changed reviewers:
+ phosek@chromium.org

[Petr Hosek, 2016-02-12 19:55:48]

lgtm

[Mark Seaborn, 2016-02-12 22:17:38]

Description was changed from

==========
PNaCl dynamic linking: Fix handling of local relocations with addends

While testing ConvertToPSO on newlib's libc.a, I found that the pass
didn't work on relocations with addends that refer to module-local
variables.

The reason is that we're running FlattenGlobals twice (once in
"-convert-to-pso" and once in "-pnacl-abi-simplify-postopt"), but
FlattenGlobals isn't idempotent for this case.  FlattenGlobals can
handle a getelementptr as input, but it converts it to ptrtoint+add,
which it doesn't handle.

Fix this by changing FlattenGlobals to handle "add" ConstantExprs.

Testing:

 * convert-to-pso.ll: Test this end-to-end.

 * flatten-globals.ll: Check this specific case, and check idempotence
   in general.

   The latter requires that we use getSExtValue() (sign-extending)
   rather than getZExtValue() (zero-extending), otherwise the
   negative-addend test fails with "FlattenGlobals: Offset does not
   fit into 32 bits".

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4351
TEST=test/Transforms/NaCl/*.ll
==========

to

==========
PNaCl dynamic linking: Fix handling of local relocations with addends

While testing ConvertToPSO on newlib's libc.a, I found that the pass
didn't work on relocations with addends that refer to module-local
variables.

The reason is that we're running FlattenGlobals twice (once in
"-convert-to-pso" and once in "-pnacl-abi-simplify-postopt"), but
FlattenGlobals isn't idempotent for this case.  FlattenGlobals can
handle a getelementptr as input, but it converts it to ptrtoint+add,
which it doesn't handle.

Fix this by changing FlattenGlobals to handle "add" ConstantExprs.

Testing:

 * convert-to-pso.ll: Test this end-to-end.

 * flatten-globals.ll: Check this specific case, and check idempotence
   in general.

   The latter requires that we use getSExtValue() (sign-extending)
   rather than getZExtValue() (zero-extending), otherwise the
   negative-addend test fails with "FlattenGlobals: Offset does not
   fit into 32 bits".

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4351
TEST=test/Transforms/NaCl/*.ll
R=phosek@chromium.org

Committed:
https://chromium.googlesource.com/native_client/pnacl-llvm/+/aa09f2c4ce3ced97...
==========

[Mark Seaborn, 2016-02-12 22:17:39]

Committed patchset #1 (id:1) manually as
aa09f2c4ce3ced974ef25e004106345daaaebd04 (presubmit successful).