Skip to content

S
swiftshader
Switch branch/tag
  1. 04 Sep, 2014 4 commits
  3. 03 Sep, 2014 2 commits
  5. 02 Sep, 2014 3 commits
  7. 29 Aug, 2014 1 commit
    • Convert lit tests to check disassembled assembly. · dddc3069
      Jan Voung authored 
      Then when we have an integrated assembler, we can check
its disassembly and the result should be the same.
This only touches the tests that invoke llvm-mc currently.
There are other tests which check for .s file output.

There are quite a bit of quirks with llvm-objdump,
which is unfortunate:

(*) The symbolizer doesn't pick up non-section-local
function calls. Some externals were converted to be
local functions. Workaround: where it counts, I just
left a check via .s files and a new --check-prefix.
It's a little better in 3.6.

(*) The symbolizer doesn't pick up global variable names.
I just checked for the relocation addend instead.
Didn't check if it was better in 3.6, but maybe.

(*) We have a bug in bundling lock + instructions.
See
BUG=https://code.google.com/p/nativeclient/issues/detail?id=3929

(*) There's no disassembly for branch lables.
Checks of jump instructions were converted to check
for positive or negative values, depending on whether
it is a forward or backward branch.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/509233002
  9. 28 Aug, 2014 2 commits
  11. 27 Aug, 2014 4 commits
  13. 26 Aug, 2014 4 commits
  15. 18 Aug, 2014 1 commit
    • Subzero: Fix the simple register allocation for -Om1. · 4d79fe5b
      Jim Stichnoth authored 
      Background: After lowering each high-level ICE instruction, Om1 calls
postLower() to do simple register allocation.  It only assigns
registers where absolutely necessary, specifically for infinite-weight
variables, while honoring pre-coloring decisions.  The original Om1
register allocation never tried to reuse registers within a lowered
sequence, which was generally OK except for very long lowering
sequences, such as call instructions or some intrinsics.  In these
cases, when it ran out of physical registers, it would just reset the
free list and hope for the best, but with no guarantee of correctness.

The fix involves keeping track of which instruction in the lowered
sequence holds the last use of each variable, and releasing each
register back to the free list after its last use.  This makes much
better use of registers.  It's not necessarily optimal, at least with
respect to pre-colored variables, since those registers are
black-listed even if they don't interfere with an infinite-weight
variable.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/483453002
  17. 15 Aug, 2014 2 commits
  19. 14 Aug, 2014 2 commits
    • Subzero: Align spill locations to natural alignment. · d4799f47
      Matt Wala authored 
      This requires sorting the spilled variables based on alignment and
introducing additional padding around the spill location areas.

These changes allow vector instructions to accept memory operands.

Old stack frame layout:  New stack frame layout:
+---------------------+  +---------------------+
| return address      |  | return address      |
+---------------------+  +---------------------+
| preserved registers |  | preserved registers |
+---------------------+  +---------------------+
| global spill area   |  | padding             |
+---------------------+  +---------------------+
| local spill area    |  | global spill area   |
+---------------------+  +---------------------+
| padding             |  | padding             |
+---------------------+  +---------------------+
| local variables     |  | local spill area    |
+---------------------+  +---------------------+
                         | padding             |
                         +---------------------+
                         | local variables     |
                         +---------------------+

BUG=none
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/465413003
    • Emit .local before .comm for bss to make llvm-mc happy. · f820da5e
      Jan Voung authored 
      Otherwise llvm-mc asserts. This is also the order that llc emits the directives.
Change a couple of RUIN -> RUN in lit tests.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/469973002
  21. 13 Aug, 2014 1 commit
    • Convert lit test llvm-mc -arch arguments to full -triple. · c8e87812
      Jan Voung authored 
      Mostly to make them a bit more portable across OSes.
Otherwise the OS assumed by llvm-mc is the build/host OS. So,
on Mac llvm-mc will assume it's targeting darwin and only accepts macho
assembler directives. Assembler directives like .rodata.cst8 are not accepted
(I'm guessing it uses .cstring, .literal4, etc. instead?).

Force an OS (NaCl) so that ELF-related assembler macros make sense.

Also remove a now unused function typeIdentString to make clang happy.

Example errors:
Command 5 Stderr:
<stdin>:5:2: error: unknown directive
        .type   fixed_400,@function
        ^
<stdin>:23:2: error: unknown directive
        .type   variable_n,@function
        ^
<stdin>:40:11: error: mach-o section specifier uses an unknown section type
        .section        .rodata.cst4,"aM",@progbits,4
                        ^
<stdin>:42:11: error: mach-o section specifier uses an unknown section type
        .section        .rodata.cst8,"aM",@progbits,8
                        ^

BUG=none
R=stichnot@chromium.org, wala@chromium.org

Review URL: https://codereview.chromium.org/467103004
  23. 12 Aug, 2014 4 commits
  25. 08 Aug, 2014 3 commits
  27. 07 Aug, 2014 2 commits
    • Subzero: Use scalar arithmetic when no vector instruction exists. · afeaee41
      Matt Wala authored 
      Implement scalarizeArithmetic() which extracts the components of the
input vectors, performs the operation with scalar instructions, and
builds the output vector component by component.

Fix the lowering of sdiv and srem.  These were previously emitting a
wrong instruction (cdq) for i8 and i16 inputs (needing cbw, cwd).

In the test_arith crosstest, mask the inputs to vector shift
operations to ensure that the shifts are in range.  Otherwise the
Subzero output is not identical to the llc output in some (undefined)
cases.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/443203003
    • Subzero: A few fixes toward running larger programs. · 206833c6
      Jim Stichnoth authored 
      1. Add 'llvm2ice -disable-globals' to disable Subzero translation of
global initializers, since full support isn't yet implemented.

2. Change the names of intra-block branch target labels to avoid
collisions with basic block labels.

3. Fix lowering of "br i1 <constant>, label ...", which was producing
invalid instructions like "cmp 1, 0".

4. Fix the "make format-diff" operation, which was diffing against the wrong target.

BUG= none
R=wala@chromium.org

Review URL: https://codereview.chromium.org/449093002
  29. 05 Aug, 2014 1 commit
    • Subzero: Fix and clean up some cross tests. · 7da431b5
      Jim Stichnoth authored 
      1. It turns out that the crosstest scripts mix different versions of
clang - build_pnacl_ir.py uses pnacl-clang from the NaCl SDK for the
tests, while crosstest.py uses clang/clang++ from LLVM_BIN_PATH for
the driver.  The SDK has been updated to use a different version of
the standard library, and now there is a mismatch as to whether int8_t
is typedef'd to 'char' or 'signed char', leading to name mangling
mismatches.  (char, signed char, and unsigned char are distinct
types.)  We deal with this by using myint8_t which is explicitly
defined as signed char.

2. Some ugly function pointer casting in test_arith_main.cpp is fixed/removed.

3. std::endl is replaced with "\n".

4. License text is added to tests that were touched by the above items.

BUG= none
R=wala@chromium.org

Review URL: https://codereview.chromium.org/435353002
  31. 31 Jul, 2014 1 commit
    • Subzero: Fix some issues related to legalization and undef handling. · e377767c
      Matt Wala authored 
      1. Much of the lowering code for vector operations was not properly
checking that the input operand was in a register or memory. This
problem could be exhibited by passing undef values as inputs.

=> Change the vector legalization code to legalize input operands to
register or memory before producing instructions that use the
operands. Also, append a suffix to the variable names in the vector
legalization code to clarify the legalization status of the values.

2. Undef values should never be emitted directly. Rather, they should
have been appropriately legalized to a zero value.

=> To enforce this, make ConstantUndef::emit() issue an error
message. Do this in the x86 backend, as other backends may decide to
treat undef values differently.

3. The regalloc_evict_non_overlap test was loading from an undef
pointer. Subzero was not handling this correctly (the undef pointer was
being emitted without being legalized), but it does not have to handle
this case since PNaCl IR disallows undef pointers.

=> Fix the regalloc_evict_non_overlap test to use an inttoptr instead of
directly loading from the undef pointer. Also, add an assert in
IceTargetLoweringX8632::FormMemoryOperand() to make sure that undef
pointers are never encountered.

BUG=none
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/432613002
  33. 30 Jul, 2014 3 commits