[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.

llvm-svn: 293713

2 files changed, 7 insertions, 5 deletions

diff --git a/llvm/test/CodeGen/NVPTX/fast-math.ll b/llvm/test/CodeGen/NVPTX/fast-math.ll
index 528d2c02df5..f925d67434c 100644
--- a/llvm/test/CodeGen/NVPTX/fast-math.ll
+++ b/llvm/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/llvm/test/CodeGen/NVPTX/sqrt-approx.ll b/llvm/test/CodeGen/NVPTX/sqrt-approx.ll
index 5edf9e28a93..1e28db44b80 100644
--- a/llvm/test/CodeGen/NVPTX/sqrt-approx.ll
+++ b/llvm/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
 }