RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 11:13:28 +01:00
Code Issues Projects Releases Wiki Activity

[NVPTX] Add (1.0 / sqrt(x)) => rsqrt(x) generation when allowable by FP flags

Browse Source 
llvm-svn: 185178
This commit is contained in:
Justin Holewinski 2013-06-28 17:58:13 +00:00
parent f17855a9dc
commit 434a514175
2 changed files with 25 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
lib/Target/NVPTX/NVPTXInstrInfo.td
View File

@ -136,7 +136,8 @@ def hasLDG : Predicate<"Subtarget.hasLDG()">;
def hasLDU : Predicate<"Subtarget.hasLDU()">;
def hasGenericLdSt : Predicate<"Subtarget.hasGenericLdSt()">;
def doF32FTZ : Predicate<"UseF32FTZ">;
def doF32FTZ : Predicate<"UseF32FTZ==1">;
def doNoF32FTZ : Predicate<"UseF32FTZ==0">;
def doFMAF32 : Predicate<"doFMAF32">;
def doFMAF32_ftz : Predicate<"(doFMAF32 && UseF32FTZ)">;
@ -765,6 +766,16 @@ def FDIV32ri_prec : NVPTXInst<(outs Float32Regs:$dst),
(fdiv Float32Regs:$a, fpimm:$b))]>,
Requires<[reqPTX20]>;
//
// F32 rsqrt
//
def RSQRTF32approx1r : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$b),
"rsqrt.approx.f32 \t$dst, $b;", []>;
def: Pat<(fdiv FloatConst1, (int_nvvm_sqrt_f Float32Regs:$b)),
(RSQRTF32approx1r Float32Regs:$b)>,
Requires<[do_DIVF32_FULL, do_SQRTF32_APPROX, doNoF32FTZ]>;
multiclass FPCONTRACT32<string OpcStr, Predicate Pred> {
def rrr : NVPTXInst<(outs Float32Regs:$dst),

13
test/CodeGen/NVPTX/rsqrt.ll Normal file
View File

@ -0,0 +1,13 @@
; RUN: llc < %s -march=nvptx -mcpu=sm_20 -nvptx-prec-divf32=1 -nvptx-prec-sqrtf32=0 | FileCheck %s
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
declare float @llvm.nvvm.sqrt.f(float)
define float @foo(float %a) {
; CHECK: rsqrt.approx.f32
%val = tail call float @llvm.nvvm.sqrt.f(float %a)
%ret = fdiv float 1.0, %val
ret float %ret
}