Commit Graph

11 Commits

Author SHA1 Message Date
Hal Finkel cbf08925ef [PowerPC] Make use of the TargetRecip system
r238842 added the TargetRecip system for controlling use of reciprocal
estimates for sqrt and division using a set of parameters that can be set by
the frontend. Clang now supports a sophisticated -mrecip option, and this will
allow that option to effectively control the relevant code-generation
functionality of the PPC backend.

llvm-svn: 241985
2015-07-12 02:33:57 +00:00
Bill Schmidt 5c6cb813b6 [PowerPC] Avoid VSX FMA mutate when killed product reg = addend reg
With VSX enabled, test/CodeGen/PowerPC/recipest.ll exposes a bug in
the FMA mutation pass.  If we have a situation where a killed product
register is the same register as the FMA target, such as:

   %vreg5<def,tied1> = XSNMSUBADP %vreg5<tied0>, %vreg11, %vreg5,
                       %RM<imp-use>; VSFRC:%vreg5 F8RC:%vreg11 

then the substitution makes no sense.  We end up getting a crash when
we try to extend the interval associated with the killed product
register, as there is already a live range for %vreg5 there.  This
patch just disables the mutation under those circumstances.

Since recipest.ll generates different code with VMX enabled, I've
modified that test to use -mattr=-vsx.  I've borrowed the code from
that test that exposed the bug and placed it in fma-mutate.ll, where
it tests several mutation opportunities including the "bad" one.

llvm-svn: 220290
2014-10-21 13:02:37 +00:00
Sanjay Patel 3d497cd778 Improve sqrt estimate algorithm (fast-math)
This patch changes the fast-math implementation for calculating sqrt(x) from:
y = 1 / (1 / sqrt(x))
to:
y = x * (1 / sqrt(x))

This has 2 benefits: less code / faster code and one less estimate instruction 
that may lose precision.

The only target that will be affected (until http://reviews.llvm.org/D5658 is approved)
is PPC. The difference in codegen for PPC is 2 less flops for a single-precision sqrtf
or vector sqrtf and 4 less flops for a double-precision sqrt. 
We also eliminate a constant load and extra register usage.

Differential Revision: http://reviews.llvm.org/D5682

llvm-svn: 219445
2014-10-09 21:26:35 +00:00
Sanjay Patel 7bc9185ab5 Fast-math fold: x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
The motivation is to recognize code such as this from /llvm/projects/test-suite/SingleSource/Benchmarks/BenchmarkGame/n-body.c:

float distance = sqrt(dx * dx + dy * dy + dz * dz);
float mag = dt / (distance * distance * distance);

Without this patch, we don't match the sqrt as a reciprocal sqrt, so for PPC the new testcase in this patch produces:

   addis 3, 2, .LCPI4_2@toc@ha
   lfs 4, .LCPI4_2@toc@l(3)
   addis 3, 2, .LCPI4_1@toc@ha
   lfs 0, .LCPI4_1@toc@l(3)
   fcmpu 0, 1, 4
   beq 0, .LBB4_2
# BB#1:
   frsqrtes 4, 1
   addis 3, 2, .LCPI4_0@toc@ha
   lfs 5, .LCPI4_0@toc@l(3)
   fnmsubs 13, 1, 5, 1
   fmuls 6, 4, 4
   fmadds 1, 13, 6, 5
   fmuls 1, 4, 1
   fres 4, 1                <--- reciprocal of reciprocal square root
   fnmsubs 1, 1, 4, 0
   fmadds 4, 4, 1, 4
.LBB4_2:
   fmuls 1, 4, 2
   fres 2, 1
   fnmsubs 0, 1, 2, 0
   fmadds 0, 2, 0, 2
   fmuls 1, 3, 0
   blr

After the patch, this simplifies to:

frsqrtes 0, 1
addis 3, 2, .LCPI4_1@toc@ha
fres 5, 2
lfs 4, .LCPI4_1@toc@l(3)
addis 3, 2, .LCPI4_0@toc@ha
lfs 7, .LCPI4_0@toc@l(3)
fnmsubs 13, 1, 4, 1
fmuls 6, 0, 0
fnmsubs 2, 2, 5, 7
fmadds 1, 13, 6, 4
fmadds 2, 5, 2, 5
fmuls 0, 0, 1
fmuls 0, 0, 2
fmuls 1, 3, 0
blr

Differential Revision: http://reviews.llvm.org/D5628

llvm-svn: 219139
2014-10-06 19:31:18 +00:00
Sanjay Patel bdf1e38856 Refactor reciprocal and reciprocal square root estimate into target-independent functions (part 2).
This is purely refactoring. No functional changes intended. PowerPC is the only target
that is currently using this interface.

The ultimate goal is to allow targets other than PowerPC (certainly X86 and Aarch64) to turn this:

z = y / sqrt(x)

into:

z = y * rsqrte(x)

And:

z = y / x

into:

z = y * rcpe(x)

using whatever HW magic they can use. See http://llvm.org/bugs/show_bug.cgi?id=20900 .

There is one hook in TargetLowering to get the target-specific opcode for an estimate instruction
along with the number of refinement steps needed to make the estimate usable.

Differential Revision: http://reviews.llvm.org/D5484

llvm-svn: 218553
2014-09-26 23:01:47 +00:00
Sanjay Patel 4bc685c206 tighten up checks
We manage to generate all of the matching instructions (and a lot more) via
the reciprocal optimization function - even if we completely remove the square
root optimization. With CHECK_NEXT, we assure that we're executing the
expected square root optimization paths and not generating extra insts.

llvm-svn: 218284
2014-09-22 22:46:44 +00:00
Sanjay Patel 5cf7561d21 remove unnecessary labels; NFC
llvm-svn: 218278
2014-09-22 21:52:53 +00:00
Hal Finkel 1e2e3ea584 Make the PPC fast-math sqrt expansion safe at 0
In fast-math mode sqrt(x) is calculated using the fast expansion of the
reciprocal of the reciprocal sqrt expansion. The reciprocal and reciprocal
sqrt expansions use the associated estimate instructions along with some Newton
iterations. Unfortunately, as a result, sqrt(0) was being calculated as NaN,
which is not correct. Now we explicitly return a result of zero if the input is
zero.

llvm-svn: 190624
2013-09-12 19:04:12 +00:00
Bill Schmidt 22f9191979 Use new CHECK-DAG support to stabilize CodeGen/PowerPC/recipest.ll
While testing some experimental code to add vector-scalar registers to
PowerPC, I noticed that a couple of independent instructions were
flipped by the scheduler.  The new CHECK-DAG support is perfect for
avoiding this problem.

llvm-svn: 182020
2013-05-16 16:15:18 +00:00
Hal Finkel f96c18e3bc PPC: Improve code generation for mixed-precision reciprocal sqrt
The DAGCombine logic that recognized a/sqrt(b) and transformed it into
a multiplication by the reciprocal sqrt did not handle cases where the
sqrt and the division were separated by an fpext or fptrunc.

llvm-svn: 178801
2013-04-04 22:44:12 +00:00
Hal Finkel 2e10331057 Use PPC reciprocal estimates with Newton iteration in fast-math mode
When unsafe FP math operations are enabled, we can use the fre[s] and
frsqrte[s] instructions, which generate reciprocal (sqrt) estimates, together
with some Newton iteration, in order to quickly generate floating-point
division and sqrt results. All of these instructions are separately optional,
and so each has its own feature flag (except for the Altivec instructions,
which are covered under the existing Altivec flag). Doing this is not only
faster than using the IEEE-compliant fdiv/fsqrt instructions, but allows these
computations to be pipelined with other computations in order to hide their
overall latency.

I've also added a couple of missing fnmsub patterns which turned out to be
missing (but are necessary for good code generation of the Newton iterations).
Altivec needs a similar fix, but that will probably be more complicated because
fneg is expanded for Altivec's v4f32.

llvm-svn: 178617
2013-04-03 04:01:11 +00:00