Re: [fpc-devel] no exceptions from arithmetical operations on AArch64
The compiler adds an explicit comparison to ensure that in case of an integer division by zero, a run time error is raised anyway. The same is done on PowerPC, which doesn't trigger an exception for integer division by zero either. As jumps in many cases are slow due to queue issues, this might be a severe performance hit. It's an interesting problem. There may be cases in which you don't need to know about the exception immediately, in which case you could avoid conditional branches by doing a conditional select in the loop, say, and only testing after the loop whether a division by zero has occurred. However, most branch predictors can cope quite well with a branch that is almost never taken. In the case of floating-point operations, the AArch64 hardware does something like that for you: it sets a sticky bit when the exceptional thing has happened. However, testing whether that bit has been set is probably more expensive than checking whether the divisor is zero so you'd want to put off the test until after several operations. How much freedom do you have in the choice of which exceptions you detect and when? Just for a check list, the AArch64 cumulative exception bits are: IXC : Inexact UFC : Underflow OFC : Overflow DZC : Division by Zero IOC : Invalid Operation (I don't have much practical experience of programming with floating-point operations, though I have at times studied what AArch64 instructions do.) Edmund ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
Re: [fpc-devel] no exceptions from arithmetical operations on AArch64
On 05/29/2015 08:37 AM, Jonas Maebe wrote: The compiler adds an explicit comparison to ensure that in case of an integer division by zero, a run time error is raised anyway. The same is done on PowerPC, which doesn't trigger an exception for integer division by zero either. As jumps in many cases are slow due to queue issues, this might be a severe performance hit. How much slower is the code after introducing the check ? Can the check be disable for high performance code snippets ? What does gcc work on that behave ? -Michael ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
Re: [fpc-devel] no exceptions from arithmetical operations on AArch64
On 05/29/2015 10:22 AM, Edmund Grimley Evans wrote: There may be cases in which you don't need to know about the exception immediately. In fact - especially with high-performance applications - there are cases where the algorithm you are doing guarantees that the divisor is not zero, so testing does not make sense. -Michael ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
Re: [fpc-devel] no exceptions from arithmetical operations on AArch64
Edmund Grimley Evans wrote on Fri, 29 May 2015: How much freedom do you have in the choice of which exceptions you detect and when? In theory: all the freedom in the world. E.g., when using the x87 instructions on x86 hardware, floating point exceptions are (currently) only thrown when the next x87 instruction starts executing following the one that caused the exception. This can be in a completely different part of the program, if you're unlucky, and is due to the origins of the x87 as an independent/off-chip co-processor. For a long time, I've had the idea of adding a compiler switch that optionally emits an fwait instruction after every x87 instruction that might cause an exception, but never got around it. In practice, I think that throwing the exception at the point where it occurs is best. It's true that it's not ideal from a performance perspective, but in that case you can just disable the generation of exception-generating code. We already have an architecture-independent way to raise an exception if the FPU detected an error while its exceptions were masked (using the Math unit as described in my previously referenced post). Therefore you can perform manual checking only at the points you care about in that case (with the caveat that's also mentioned in that previous post, namely that currently this kind of checking is only implemented on AArch64). Just for a check list, the AArch64 cumulative exception bits are: IXC : Inexact UFC : Underflow OFC : Overflow DZC : Division by Zero IOC : Invalid Operation These are more or less the same on all architectures. The reason is probably that they're mandated by the IEEE 754 standard: http://en.wikipedia.org/wiki/IEEE_floating_point#Exception_handling Jonas ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
Re: [fpc-devel] no exceptions from arithmetical operations on AArch64
Edmund Grimley Evans wrote: Looking through the tests that fail on aarch64-linux I found ten that depend on an arithmetical operation causing an exception, usually division by zero: tbs/tb0262 test/texception4 test/units/math/tmask test/units/math/tmask2 webtbs/tw3157 webtbs/tw3160a webtbs/tw3160b webtbs/tw3160c webtbs/tw3161 webtbs/tw4100 In AArch64 the instruction for integer division never causes an exception, The compiler adds an explicit comparison to ensure that in case of an integer division by zero, a run time error is raised anyway. The same is done on PowerPC, which doesn't trigger an exception for integer division by zero either. and floating-point arithmetic can only cause an exception if an optional part of the architecture is enabled; most hardware doesn't even have the option. It turns out that this is also optional in the regular ARM architecture, although there most implementations do seem to support it. An AArch64 cpu without the support running plain ARM code won't trigger floating point exceptions either though. Clearly the compiler could generate code to check for exceptions but it would be very inefficient when those exceptions are not required. The idea is to add a code generation option for this, so the programmer can decide whether they want the exceptions (and associated performance overhead) or not. I can't find an explicit statement in the documentation but this example suggests that exceptions from arithmetical operations are not required by the language definition: http://www.freepascal.org/docs-html/ref/refse101.html#x212-22200017.1 While the example does not have the intention of demonstrating that, it is in fact true that according to the ISO Extended Pascal standard nothing special needs to happen in case of a division by zero: a) A term of the form i div j shall be an error if j is zero (http://www.pascal-central.com/docs/iso10206.pdf , section 6.8.3.2 — the same goes for i / j, which is the floating point division). b) definition of Error: A violation by a program of the requirements of this International Standard that a processor is permitted to leave undetected (section 3.2) However, we generally mainly aim for Turbo Pascal and Delphi compatibility, and most code written for those compilers (and for FPC on other platforms) assumes that it is at least possible to configure the platform to raise floating point exceptions. So, can those tests be disabled on AArch64, by adding { %skipcpu=aarch64 }? I don't think that's the best approach. Here are two possible alternatives: a) we add the math unit to the uses clause, check whether floating point exceptions can be enabled (see http://lists.freepascal.org/pipermail/fpc-devel/2015-February/035397.html for the details) and if not, skip the test (that would also cover running the tests when compiled for ARM) b) we leave them as they are until support has been added to the code generator to generate explicit exception checks after each floating point operation that may trigger one (and then specify that this option should be used when compiling those tests; the compiler will ignore it on platforms that don't need it) My preference goes to option b). I haven't started working on that yet though. Jonas ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
[fpc-devel] no exceptions from arithmetical operations on AArch64
Looking through the tests that fail on aarch64-linux I found ten that depend on an arithmetical operation causing an exception, usually division by zero: tbs/tb0262 test/texception4 test/units/math/tmask test/units/math/tmask2 webtbs/tw3157 webtbs/tw3160a webtbs/tw3160b webtbs/tw3160c webtbs/tw3161 webtbs/tw4100 In AArch64 the instruction for integer division never causes an exception, and floating-point arithmetic can only cause an exception if an optional part of the architecture is enabled; most hardware doesn't even have the option. Clearly the compiler could generate code to check for exceptions but it would be very inefficient when those exceptions are not required. I can't find an explicit statement in the documentation but this example suggests that exceptions from arithmetical operations are not required by the language definition: http://www.freepascal.org/docs-html/ref/refse101.html#x212-22200017.1 So, can those tests be disabled on AArch64, by adding { %skipcpu=aarch64 }? Edmund ___ fpc-devel maillist - fpc-devel@lists.freepascal.org http://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel