HI,
By splitting the assert I am getting the following IR:
*define dso_local i32 @main() #0 {entry: %c = alloca double, align 8 %0 =
call double @llvm.ceil.f64(double 0xFFF0000000000000) store double %0,
double* %c, align 8 %call = call i32 @__isinff(float 0x7FF0000000000000)
#4 %call1 = call i32 (i32, ...) bitcast (i32 (...)* @assert to i32 (i32,
...)*)(i32 %call) %1 = load double, double* %c, align 8 %2 = bitcast
double %1 to i64 %3 = icmp slt i64 %2, 0 %4 = zext i1 %3 to i32 %call2 =
call i32 (i32, ...) bitcast (i32 (...)* @assert to i32 (i32, ...)*)(i32
%4) ret i32 0}*
It's not the most optimal but still doesn't have as much unnecessary code.
My main concern is why are signle, double and extended precision floats
being checked for 'isinf', when the original code only asks for double, and
this behaviour seems to be induced by the logical and operator.
Thanks,
Akash.
On Thu, Jan 30, 2020 at 2:17 PM Craig Topper <craig.topper at gmail.com> wrote:
> That code looks like it was compiled without any optimizations enabled.
> The IR without optimizations is meant to be easy for the frontend to
> generate. Even with the code split into two asserts, without optimizations
> enabled I see a lot of the same things unnecessary things.
>
> ~Craig
>
>
> On Thu, Jan 30, 2020 at 12:25 AM Akash Banerjee via cfe-dev <
> cfe-dev at lists.llvm.org> wrote:
>
>> Please consider the following C code:
>>
>>
>>
>>
>> *#include <math.h>int main(void) { double c = ceil(-INFINITY);
>> assert(isinf(INFINITY) && signbit(c));}*
>>
>> The following LLVM IR is generated for the above code by clang:
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> *define dso_local i32 @main() #0 {entry: %retval = alloca i32, align 4
>> %c = alloca double, align 8 store i32 0, i32* %retval, align 4 %0 = call
>> double @llvm.ceil.f64(double 0xFFF0000000000000) store double %0, double*
>> %c, align 8 br i1 true, label %cond.true, label %cond.false
>> //Unnecessary conditional branchcond.true:
>> ; preds = %entry %call = call i32 @__isinff(float
>> 0x7FF0000000000000) #4 %tobool = icmp ne i32 %call, 0 br i1 %tobool,
>> label %land.rhs, label %land.endcond.false:
>> ; preds = %entry** //Unnecessary Block*
>>
>>
>> * br i1 false, label %cond.true1, label %cond.false4cond.true1:
>> ; preds = %cond.false*
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> * //Unnecessary Block %call2 = call i32 @__isinf(double
>> 0x7FF0000000000000) #4 %tobool3 = icmp ne i32 %call2, 0 br i1 %tobool3,
>> label %land.rhs, label %land.endcond.false4:
>> ; preds = %cond.false //Unnecessary Block %call5 = call i32
>> @__isinfl(x86_fp80 0xK7FFF8000000000000000) #4 %tobool6 = icmp ne i32
>> %call5, 0 br i1 %tobool6, label %land.rhs, label %land.endland.rhs:
>> ; preds = %cond.false4, %cond.true1,
>> %cond.true %1 = load double, double* %c, align 8 %2 = bitcast double %1
>> to i64 %3 = icmp slt i64 %2, 0 br label %land.endland.end:
>> ; preds = %land.rhs, %cond.false4, %cond.true1,
>> %cond.true %4 = phi i1 [ false, %cond.false4 ], [ false, %cond.true1 ], [
>> false, %cond.true ], [ %3, %land.rhs ] %land.ext = zext i1 %4 to i32
>> %call7 = call i32 (i32, ...) bitcast (i32 (...)* @assert to i32 (i32,
>> ...)*)(i32 %land.ext) %5 = load i32, i32* %retval, align 4 ret i32 %5}*
>>
>> This seems to happen because of the logical and operator in the assert,
>> and instead, if the assert is split into two separate asserts:
>> * assert(isinf(INFINITY));*
>> * assert(signbit(c));*
>> , then the IR generated is fine.
>>
>> Thanks,
>> Akash.
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20200130/c1ea9082/attachment.html>