x*rsqrt(x) returns NaN for x == 0, whereas 1/rsqrt(x) returns 0, as
desired.
Verified that the particular nvptx approximate instructions here do in
fact return 0 for x = 0.
llvm-svn: 293713
Summary:
The affected transforms all implicitly use associativity of addition,
for which we usually require unsafe math to be enabled.
The "Aggressive" flag is only meant to convey information about the
performance of the fused ops relative to a fmul+fadd sequence.
Fixes Bug 31626.
Reviewers: spatel, hfinkel, mehdi_amini, arsenm, tstellarAMD
Subscribers: jholewinski, nemanjai, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D28675
llvm-svn: 293635
Summary:
This lets us lower to sqrt.approx and rsqrt.approx under more
circumstances.
* Now we emit sqrt.approx and rsqrt.approx for calls to @llvm.sqrt.f32,
when fast-math is enabled. Previously, we only would emit it for
calls to @llvm.nvvm.sqrt.f. (With this patch we no longer emit
sqrt.approx for calls to @llvm.nvvm.sqrt.f; we rely on intcombine to
simplify llvm.nvvm.sqrt.f into llvm.sqrt.f32.)
* Now we emit the ftz version of rsqrt.approx when ftz is enabled.
Previously, we only emitted rsqrt.approx when ftz was disabled.
Reviewers: hfinkel
Subscribers: llvm-commits, tra, jholewinski
Differential Revision: https://reviews.llvm.org/D28508
llvm-svn: 293605
Support for barrier synchronization between a subset of threads
in a CTA through one of sixteen explicitly specified barriers.
These intrinsics are not directly exposed in CUDA but are
critical for forthcoming support of OpenMP on NVPTX GPUs.
The intrinsics allow the synchronization of an arbitrary
(multiple of 32) number of threads in a CTA at one of 16
distinct barriers. The two intrinsics added are as follows:
call void @llvm.nvvm.barrier.n(i32 10)
waits for all threads in a CTA to arrive at named barrier #10.
call void @llvm.nvvm.barrier(i32 15, i32 992)
waits for 992 threads in a CTA to arrive at barrier #15.
Detailed description of these intrinsics are available in the PTX manual.
http://docs.nvidia.com/cuda/parallel-thread-execution/#parallel-synchronization-and-communication-instructions
Reviewers: hfinkel, jlebar
Differential Revision: https://reviews.llvm.org/D17657
llvm-svn: 293384
There's no neg.f16 instruction, so negation has to
be done via subtraction from zero.
Differential Revision: https://reviews.llvm.org/D28876
llvm-svn: 292452
Summary:
This change also lets us use max.{s,u}16. There's a vague warning in a
test about this maybe being less efficient, but I could not come up with
a case where the resulting SASS (sm_35 or sm_60) was different with or
without max.{s,u}16. It's true that nvcc seems to emit only
max.{s,u}32, but even ptxas 7.0 seems to have no problem generating
efficient SASS from max.{s,u}16 (the casts up to i32 and back down to
i16 seem to be implicit and nops, happening via register aliasing).
In the absence of evidence, better to have fewer special cases, emit
more straightforward code, etc. In particular, if a new GPU has 16-bit
min/max instructions, we want to be able to use them.
Reviewers: tra
Subscribers: jholewinski, llvm-commits
Differential Revision: https://reviews.llvm.org/D28732
llvm-svn: 292304
Summary: Previously we lowered it literally, to shifts and xors.
Reviewers: tra
Subscribers: jholewinski, llvm-commits
Differential Revision: https://reviews.llvm.org/D28722
llvm-svn: 292303
Summary:
Avoid an unnecessary conversion operation when using the result of
ctpop.i32 or ctpop.i16 as an i32, as in both cases the ptx instruction
we run returns an i32.
(Previously if we used the value as an i32, we'd do an unnecessary
zext+trunc.)
Reviewers: tra
Subscribers: jholewinski, llvm-commits
Differential Revision: https://reviews.llvm.org/D28721
llvm-svn: 292302
Summary:
* Disable "ctlz speculation", which inserts a branch on every ctlz(x) which
has defined behavior on x == 0 to check whether x is, in fact zero.
* Add DAG patterns that avoid re-truncating or re-expanding the result
of the 16- and 64-bit ctz instructions.
Reviewers: tra
Subscribers: llvm-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D28719
llvm-svn: 292299
Summary:
Previously there were three ways to inform the NVVMReflect pass whether
you wanted to flush denormals to zero:
* An LLVM command-line option
* Parameters to the NVVMReflect constructor
* Metadata on the module itself.
This change removes the first two, leaving only the third.
The motivation for this change, aside from simplifying things, is that
we want LLVM to be aware of whether it's operating in FTZ mode, so other
passes can use this information. Ideally we'd have a target-generic
piece of metadata on the module. This change moves us in that
direction.
Reviewers: tra
Subscribers: jholewinski, llvm-commits
Differential Revision: https://reviews.llvm.org/D28700
llvm-svn: 292068
Only scalar half-precision operations are supported at the moment.
- Adds general support for 'half' type in NVPTX.
- fp16 math operations are supported on sm_53+ GPUs only
(can be disabled with --nvptx-no-f16-math).
- Type conversions to/from fp16 are supported on all GPU variants.
- On GPU variants that do not have full fp16 support (or if it's disabled),
fp16 operations are promoted to fp32 and results are converted back
to fp16 for storage.
Differential Revision: https://reviews.llvm.org/D28540
llvm-svn: 291956
Previously we'd always lower @llvm.{sin,cos}.f32 to {sin.cos}.approx.f32
instruction even when unsafe FP math was not allowed.
Clang-generated IR is not affected by this as it uses precise sin/cos
from CUDA's libdevice when unsafe math is disabled.
Differential Revision: https://reviews.llvm.org/D28619
llvm-svn: 291936
Summary:
Previously if you had
* a function with the fast-math-enabled attr, followed by
* a function without the fast-math attr,
the second function would inherit the first function's fast-math-ness.
This means that mixing fast-math and non-fast-math functions in a module
was completely broken unless you explicitly annotated every
non-fast-math function with "unsafe-fp-math"="false". This appears to
have been broken since r176986 (March 2013), when the resetTargetOptions
function was introduced.
This patch tests the correct behavior as best we can. I don't think I
can test FPDenormalMode and NoTrappingFPMath, because they aren't used
in any backends during function lowering. Surprisingly, I also can't
find any uses at all of LessPreciseFPMAD affecting generated code.
The NVPTX/fast-math.ll test changes are an expected result of fixing
this bug. When FMA is disabled, we emit add as "add.rn.f32", which
prevents fma combining. Before this patch, fast-math was enabled in all
functions following the one which explicitly enabled it on itself, so we
were emitting plain "add.f32" where we should have generated
"add.rn.f32".
Reviewers: mkuper
Subscribers: hfinkel, majnemer, jholewinski, nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D28507
llvm-svn: 291618
We used the logBase2 of the high instead of the ceilLogBase2 resulting
in the wrong result for certain values. For example, it resulted in an
i1 AssertZExt when the exclusive portion of the range was 3.
llvm-svn: 291196
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
This reapplies r289902 with additional testcase upgrades and a change
to the Bitcode record for DIGlobalVariable, that makes upgrading the
old format unambiguous also for variables without DIExpressions.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 290153
This reverts commit 289920 (again).
I forgot to implement a Bitcode upgrade for the case where a DIGlobalVariable
has not DIExpression. Unfortunately it is not possible to safely upgrade
these variables without adding a flag to the bitcode record indicating which
version they are.
My plan of record is to roll the planned follow-up patch that adds a
unit: field to DIGlobalVariable into this patch before recomitting.
This way we only need one Bitcode upgrade for both changes (with a
version flag in the bitcode record to safely distinguish the record
formats).
Sorry for the churn!
llvm-svn: 289982
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
This reapplies r289902 with additional testcase upgrades.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 289920
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
<rdar://problem/29250149>
https://llvm.org/bugs/show_bug.cgi?id=31013
Differential Revision: https://reviews.llvm.org/D26769
llvm-svn: 289902
Summary:
This has been replaced by the NVPTXInferAddressSpaces pass. We've had
the new one as the default with the old one accessible via a flag for
some months now, and we've had no problems.
Reviewers: tra
Subscribers: llvm-commits, jholewinski, jingyue, mgorny
Differential Revision: https://reviews.llvm.org/D26165
llvm-svn: 285642
Summary:
In isel, transform
Num % Den
into
Num - (Num / Den) * Den
if the result of Num / Den is already available.
Reviewers: tra
Subscribers: hfinkel, llvm-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D26090
llvm-svn: 285461
Summary: In getArgumentAlignment check if the ImmutableCallSite pointer CS is non-null before dereferencing. If CS is 0x0 fall back to the ABI type alignment else compute the alignment as before.
Reviewers: eliben, jpienaar
Subscribers: jlebar, vchuravy, cfe-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D9168
llvm-svn: 282045
This patch reverses the edge from DIGlobalVariable to GlobalVariable.
This will allow us to more easily preserve debug info metadata when
manipulating global variables.
Fixes PR30362. A program for upgrading test cases is attached to that
bug.
Differential Revision: http://reviews.llvm.org/D20147
llvm-svn: 281284
Summary:
With this change (plus some changes to prevent !invariant from being
clobbered within llvm), clang will be able to model the __ldg CUDA
builtin as an invariant load, rather than as a target-specific llvm
intrinsic. This will let the optimizer play with these loads --
specifically, we should be able to vectorize them in the load-store
vectorizer.
Reviewers: tra
Subscribers: jholewinski, hfinkel, llvm-commits, chandlerc
Differential Revision: https://reviews.llvm.org/D23477
llvm-svn: 281152
Summary:
Previously these only worked via NVPTX-specific intrinsics.
This change will allow us to convert these target-specific intrinsics
into the general LLVM versions, allowing existing LLVM passes to reason
about their behavior.
It also gets us some minor codegen improvements as-is, from situations
where we canonicalize code into one of these llvm intrinsics.
Reviewers: majnemer
Subscribers: llvm-commits, jholewinski, tra
Differential Revision: https://reviews.llvm.org/D24300
llvm-svn: 281092
Summary:
This switches us to use a different, more powerful algorithm for address
space inference. I've tested this locally and it seems to work great.
Once we're more confident in it, we can remove the old pass altogether.
Reviewers: jingyue
Subscribers: llvm-commits, tra, jholewinski
Differential Revision: https://reviews.llvm.org/D23694
llvm-svn: 279317
This bring LLVM-generated PTX closer to what nvcc generates and avoids
triggering issues in ptxas.
For instance, ptxas does not accept .s16 (or .u16) registers as operands
for .fp16 instructions.
Differential Revision: https://reviews.llvm.org/D23460
llvm-svn: 278568
Avoid unnecessary spills of byval arguments of device functions to
local space on SASS level and subsequent pointer conversion to generic
address space that follows. Instead, make a local copy in IR, provide
a way to access arguments directly, and let LLVM optimize the copy away
when possible.
Differential Review: https://reviews.llvm.org/D21421
llvm-svn: 276153
Justin Lebar