diff options
author | Justin Lebar <jlebar@google.com> | 2017-01-31 23:08:57 +0000 |
---|---|---|
committer | Justin Lebar <jlebar@google.com> | 2017-01-31 23:08:57 +0000 |
commit | 6f09ea3f571cd2f4c39fff52deecec7cfc62d0c4 (patch) | |
tree | 235c9d4ca67913b3a77c113a4bdfd30648a3e8bf /test/CodeGen/NVPTX | |
parent | d99a5ea737c696917819bd764e5ecb960e15d0eb (diff) |
[NVPTX] Compute approx sqrt as 1/rsqrt(x) rather than x*rsqrt(x).
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.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@293713 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'test/CodeGen/NVPTX')
-rw-r--r-- | test/CodeGen/NVPTX/fast-math.ll | 4 | ||||
-rw-r--r-- | test/CodeGen/NVPTX/sqrt-approx.ll | 8 |
2 files changed, 7 insertions, 5 deletions
diff --git a/test/CodeGen/NVPTX/fast-math.ll b/test/CodeGen/NVPTX/fast-math.ll index 528d2c02df5..f925d67434c 100644 --- a/test/CodeGen/NVPTX/fast-math.ll +++ b/test/CodeGen/NVPTX/fast-math.ll @@ -40,11 +40,11 @@ define float @sqrt_div_fast_ftz(float %a, float %b) #0 #1 { } ; There are no fast-math or ftz versions of sqrt and div for f64. We use -; x * rsqrt(x) for sqrt(x), and emit a vanilla divide. +; reciprocal(rsqrt(x)) for sqrt(x), and emit a vanilla divide. ; CHECK-LABEL: sqrt_div_fast_ftz_f64( ; CHECK: rsqrt.approx.f64 -; CHECK: mul.f64 +; CHECK: rcp.approx.ftz.f64 ; CHECK: div.rn.f64 define double @sqrt_div_fast_ftz_f64(double %a, double %b) #0 #1 { %t1 = tail call double @llvm.sqrt.f64(double %a) diff --git a/test/CodeGen/NVPTX/sqrt-approx.ll b/test/CodeGen/NVPTX/sqrt-approx.ll index 5edf9e28a93..1e28db44b80 100644 --- a/test/CodeGen/NVPTX/sqrt-approx.ll +++ b/test/CodeGen/NVPTX/sqrt-approx.ll @@ -59,9 +59,11 @@ define float @test_sqrt_ftz(float %a) #0 #1 { ; CHECK-LABEL test_sqrt64 define double @test_sqrt64(double %a) #0 { -; There's no sqrt.approx.f64 instruction; we emit x * rsqrt.approx.f64(x). +; There's no sqrt.approx.f64 instruction; we emit +; reciprocal(rsqrt.approx.f64(x)). There's no non-ftz approximate reciprocal, +; so we just use the ftz version. ; CHECK: rsqrt.approx.f64 -; CHECK: mul.f64 +; CHECK: rcp.approx.ftz.f64 %ret = tail call double @llvm.sqrt.f64(double %a) ret double %ret } @@ -70,7 +72,7 @@ define double @test_sqrt64(double %a) #0 { define double @test_sqrt64_ftz(double %a) #0 #1 { ; There's no sqrt.approx.ftz.f64 instruction; we just use the non-ftz version. ; CHECK: rsqrt.approx.f64 -; CHECK: mul.f64 +; CHECK: rcp.approx.ftz.f64 %ret = tail call double @llvm.sqrt.f64(double %a) ret double %ret } |