[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
diff --git a/llvm/test/CodeGen/NVPTX/sqrt-approx.ll b/llvm/test/CodeGen/NVPTX/sqrt-approx.ll
index 5edf9e2..1e28db4 100644
--- a/llvm/test/CodeGen/NVPTX/sqrt-approx.ll
+++ b/llvm/test/CodeGen/NVPTX/sqrt-approx.ll
@@ -59,9 +59,11 @@
; 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_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
}