Thursday, July 27, 2006
GCC Summit part 1.25
Pictures from the GCC Summit from LXO:
http://www.ic.unicamp.br/~oliva/images/pictures/2006-06-GPLv3-Barcelona-Zurich-GCCSummit-Ottawa/
http://www.ic.unicamp.br/~oliva/images/pictures/2006-06-GPLv3-Barcelona-Zurich-GCCSummit-Ottawa/
Sunday, July 23, 2006
Japan Part 3, the end
So I am back home now in Foster City where it got to 108F (42C) yesterday. Man it was hot. But that is not what this posting is about. Anyways Japan was fun, I wish I did not have trouble sleeping. I also wished I was not preoccupied with other stuff unrelated to work or even this trip, I would have enjoyed it even more.
The SCEI building was a couple buildings down from the Honda HQ, I have pictures. On Monday the holiday, I took a walk from Shibuya subway station to the office, well I did not know where I was going but I ended up at my office. This walk was nice because I went by places where hardly anyone was and it was quiet which meant I could think about the things troubling me, which I am not going to say them here at least not yet. It is a long and weird story and I don't want to tell it just yet until I find out more (the story will be called GCC summit part 2 when I get around to writting it down). Anyways so I took this walk, it got me thinking, am I ready to settle down and have a family. I don't know, I don't think I am ready just yet; work just takes too much time out of me currently. Plus working with GCC's bugs takes the other part of my spare time but then again having a girl-friend would be useful so I could have some spare time and actually get away from GCC sometime. Ok, maybe I gave part of the long story up but oh well.
On Tuesday night, the whole toolchain group went out drinking. It was fun, though I did drink the most as it was a welcome party for me. Thursday for lunch, we had Indian, one of the best Indian food I had in years. One of the dishes was spicy, though not much, it was good. We had to wait an half an hour to get seated. Friday, I spend most of the day at the airport waiting for the plane. Well that is not fully true as I had two Fridays. My plane left at 6pm JST (Tokyo time) and got in around 11am PDT (California TIme). I went into work to have lunch with Alex and we talked for about an hour and an half and then he had to leave to go. I went home from lunch and decided to sleep at 4pm, god I was tried, I had got less than an hour on the plane ride and that was 9 hours itself so I was up almost 24hours, it was crazy. So I woke up around 11pm that same night. I could not sleep for another 3 hours and then I forced myself to go to bed and get some more sleep. I woke up at some early time on Saturday, I had enough sleep for the day.
I am running into bugs or maybe mis-features in the GCC code about reload. I wish reload would just die. Anyways I am thinking about seeing about the infrastructure which was added for 4.2.0 which might help fix this issue but that would mean I need to backport the patch and the fixes related to it to 4.1.1 which I am regreating doing. Also on the plane ride back, I remembered that the code I was writting would only work for targets that set strict alignment. Also from look at the code gen because I need to add splats, it might get funny looking but it still does gets rid of the load hit store reload issue which is much worse than the other code gen. I might even figure out a way of not needing the splats but the splats make it easy to store stuff to memory on a strict alignment target as you just use stvewx instead of having to do a load and then a perm and then a store of the whole vector.
The SCEI building was a couple buildings down from the Honda HQ, I have pictures. On Monday the holiday, I took a walk from Shibuya subway station to the office, well I did not know where I was going but I ended up at my office. This walk was nice because I went by places where hardly anyone was and it was quiet which meant I could think about the things troubling me, which I am not going to say them here at least not yet. It is a long and weird story and I don't want to tell it just yet until I find out more (the story will be called GCC summit part 2 when I get around to writting it down). Anyways so I took this walk, it got me thinking, am I ready to settle down and have a family. I don't know, I don't think I am ready just yet; work just takes too much time out of me currently. Plus working with GCC's bugs takes the other part of my spare time but then again having a girl-friend would be useful so I could have some spare time and actually get away from GCC sometime. Ok, maybe I gave part of the long story up but oh well.
On Tuesday night, the whole toolchain group went out drinking. It was fun, though I did drink the most as it was a welcome party for me. Thursday for lunch, we had Indian, one of the best Indian food I had in years. One of the dishes was spicy, though not much, it was good. We had to wait an half an hour to get seated. Friday, I spend most of the day at the airport waiting for the plane. Well that is not fully true as I had two Fridays. My plane left at 6pm JST (Tokyo time) and got in around 11am PDT (California TIme). I went into work to have lunch with Alex and we talked for about an hour and an half and then he had to leave to go. I went home from lunch and decided to sleep at 4pm, god I was tried, I had got less than an hour on the plane ride and that was 9 hours itself so I was up almost 24hours, it was crazy. So I woke up around 11pm that same night. I could not sleep for another 3 hours and then I forced myself to go to bed and get some more sleep. I woke up at some early time on Saturday, I had enough sleep for the day.
I am running into bugs or maybe mis-features in the GCC code about reload. I wish reload would just die. Anyways I am thinking about seeing about the infrastructure which was added for 4.2.0 which might help fix this issue but that would mean I need to backport the patch and the fixes related to it to 4.1.1 which I am regreating doing. Also on the plane ride back, I remembered that the code I was writting would only work for targets that set strict alignment. Also from look at the code gen because I need to add splats, it might get funny looking but it still does gets rid of the load hit store reload issue which is much worse than the other code gen. I might even figure out a way of not needing the splats but the splats make it easy to store stuff to memory on a strict alignment target as you just use stvewx instead of having to do a load and then a perm and then a store of the whole vector.
Friday, July 14, 2006
Japan part 2
So here I am in Japan on a Friday night in the hotel. Well I have Saturday night and Sunday night to part as Monday is a holiday. So I was going to meet a friend of the OZ labs but I missed him because the subway station was layed out stupidly. Anyways I was trying to get to the JR south exit but I went through a gate to get a train and could not get out so I had to take a train and then try to get back, hoping I was not going to get lost. So I got on the first train out and made it to a different station (the line which I took did not exist on the map too which was fun). It took an hour to get back to the correct station. He left after the half an hour mark. OH well.
On Thursday I did meet camel from #llvm, and we had okonomiyaki. It was nice.
Tonight (since it is now Saturday at 1pm), I might go out on the town but I am not really awake yet and I don't feel that good.
Work has been good, I am finally getting stuff done but running into bugs which people have not run into before which sucks. First it was BIT_FIELD_REF on the left hand side, and then vec_set was not being used and then an ICE due to BIT_FIELD_REF < VIEW_CONVERT_EXPR uses the wrong mode (and even the fix I came for that is incorrect as it produces horible code in some cases, well a simple one of a+b or two way different modes).
On Thursday I did meet camel from #llvm, and we had okonomiyaki. It was nice.
Tonight (since it is now Saturday at 1pm), I might go out on the town but I am not really awake yet and I don't feel that good.
Work has been good, I am finally getting stuff done but running into bugs which people have not run into before which sucks. First it was BIT_FIELD_REF on the left hand side, and then vec_set was not being used and then an ICE due to BIT_FIELD_REF < VIEW_CONVERT_EXPR uses the wrong mode (and even the fix I came for that is incorrect as it produces horible code in some cases, well a simple one of a+b or two way different modes).
Wednesday, July 12, 2006
Japan part 1
So far Tokyo has been fun. Though I have not eaten dinner at all because I have not been able to decide what to eat. It is hot and humid here. Even sometimes it gets hot in the office so much I decide to head home which means going in an even hotter and more humid walk.
I have been working on an interesting optimization for GCC but I have ran into a couple of other optimizations[1] that need to be done before this one actually starts working. The optimization is to work around a hazard with the Cell processor. The idea is just like doing the floating point in the SSE unit but instead doing it in the VMX unit on the Cell. This optimization allows for the compiler to do some scalar computation in vectors instead of storing it off to the stack.
A co-work has suggested a new type but that seems to defet the propose of the optimization in that the code should not need to be changed
[1]http://gcc.gnu.org/ml/gcc/2006-07/msg00215.html
I have been working on an interesting optimization for GCC but I have ran into a couple of other optimizations[1] that need to be done before this one actually starts working. The optimization is to work around a hazard with the Cell processor. The idea is just like doing the floating point in the SSE unit but instead doing it in the VMX unit on the Cell. This optimization allows for the compiler to do some scalar computation in vectors instead of storing it off to the stack.
A co-work has suggested a new type but that seems to defet the propose of the optimization in that the code should not need to be changed
[1]http://gcc.gnu.org/ml/gcc/2006-07/msg00215.html
Monday, July 03, 2006
GCC summit part 1
Ok, this post is going to be a bit weird but here it goes.
What was I thinking during Friday night? What did I do correct and why?
Anyways the GCC summit was good. Most of the talks were boring but getting to know people and their faces is always good.
The flight out to Ottawa sucked, it took about 14 hours longer than excepted. With an overnight stay in Chicago which I had to pay which is crap. And then during the next lag of the flight in Chicago, it took an extra two hours because the plane was delayed.
What was I thinking during Friday night? What did I do correct and why?
Anyways the GCC summit was good. Most of the talks were boring but getting to know people and their faces is always good.
The flight out to Ottawa sucked, it took about 14 hours longer than excepted. With an overnight stay in Chicago which I had to pay which is crap. And then during the next lag of the flight in Chicago, it took an extra two hours because the plane was delayed.
