RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-21 20:12:56 +02:00
Code Issues Projects Releases Wiki Activity
44b06b0096
llvm-mirror/test/CodeGen/X86/block-placement.ll

396 lines
9.2 KiB
LLVM
Raw Normal View History

Add a very basic test for MachineBlockPlacement. This is essentially the canonical example I used when developing it, and is one of the primary motivating real-world use cases for __builtin_expect (when burried under a macro). I'm working on more test cases here, but I'm trying to make sure both that the pass is doing the right thing with the test cases and that they aren't too brittle to changes elsewhere in the code generation pipeline. Feedback and/or suggestions on how to test this are very welcome. Especially feedback on whether testing the block comments is a good strategy; I couldn't find any good examples to steal from but all the other ideas I had were a lot uglier or more fragile. llvm-svn: 142644
2011-10-21 10:01:56 +02:00
; RUN: llc -march=x86 -enable-block-placement < %s | FileCheck %s
declare void @error(i32 %i, i32 %a, i32 %b)
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
define i32 @test_ifchains(i32 %i, i32* %a, i32 %b) {
Add a very basic test for MachineBlockPlacement. This is essentially the canonical example I used when developing it, and is one of the primary motivating real-world use cases for __builtin_expect (when burried under a macro). I'm working on more test cases here, but I'm trying to make sure both that the pass is doing the right thing with the test cases and that they aren't too brittle to changes elsewhere in the code generation pipeline. Feedback and/or suggestions on how to test this are very welcome. Especially feedback on whether testing the block comments is a good strategy; I couldn't find any good examples to steal from but all the other ideas I had were a lot uglier or more fragile. llvm-svn: 142644
2011-10-21 10:01:56 +02:00
; Test a chain of ifs, where the block guarded by the if is error handling code
; that is not expected to run.
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
; CHECK: test_ifchains:
Add a very basic test for MachineBlockPlacement. This is essentially the canonical example I used when developing it, and is one of the primary motivating real-world use cases for __builtin_expect (when burried under a macro). I'm working on more test cases here, but I'm trying to make sure both that the pass is doing the right thing with the test cases and that they aren't too brittle to changes elsewhere in the code generation pipeline. Feedback and/or suggestions on how to test this are very welcome. Especially feedback on whether testing the block comments is a good strategy; I couldn't find any good examples to steal from but all the other ideas I had were a lot uglier or more fragile. llvm-svn: 142644
2011-10-21 10:01:56 +02:00
; CHECK: %entry
; CHECK: %else1
; CHECK: %else2
; CHECK: %else3
; CHECK: %else4
; CHECK: %exit
; CHECK: %then1
; CHECK: %then2
; CHECK: %then3
; CHECK: %then4
; CHECK: %then5
entry:
%gep1 = getelementptr i32* %a, i32 1
%val1 = load i32* %gep1
%cond1 = icmp ugt i32 %val1, 1
br i1 %cond1, label %then1, label %else1, !prof !0
then1:
call void @error(i32 %i, i32 1, i32 %b)
br label %else1
else1:
%gep2 = getelementptr i32* %a, i32 2
%val2 = load i32* %gep2
%cond2 = icmp ugt i32 %val2, 2
br i1 %cond2, label %then2, label %else2, !prof !0
then2:
call void @error(i32 %i, i32 1, i32 %b)
br label %else2
else2:
%gep3 = getelementptr i32* %a, i32 3
%val3 = load i32* %gep3
%cond3 = icmp ugt i32 %val3, 3
br i1 %cond3, label %then3, label %else3, !prof !0
then3:
call void @error(i32 %i, i32 1, i32 %b)
br label %else3
else3:
%gep4 = getelementptr i32* %a, i32 4
%val4 = load i32* %gep4
%cond4 = icmp ugt i32 %val4, 4
br i1 %cond4, label %then4, label %else4, !prof !0
then4:
call void @error(i32 %i, i32 1, i32 %b)
br label %else4
else4:
%gep5 = getelementptr i32* %a, i32 3
%val5 = load i32* %gep5
%cond5 = icmp ugt i32 %val5, 3
br i1 %cond5, label %then5, label %exit, !prof !0
then5:
call void @error(i32 %i, i32 1, i32 %b)
br label %exit
exit:
ret i32 %b
}
Rewrite #3 of machine block placement. This is based somewhat on the second algorithm, but only loosely. It is more heavily based on the last discussion I had with Andy. It continues to walk from the inner-most loop outward, but there is a key difference. With this algorithm we ensure that as we visit each loop, the entire loop is merged into a single chain. At the end, the entire function is treated as a "loop", and merged into a single chain. This chain forms the desired sequence of blocks within the function. Switching to a single algorithm removes my biggest problem with the previous approaches -- they had different behavior depending on which system triggered the layout. Now there is exactly one algorithm and one basis for the decision making. The other key difference is how the chain is formed. This is based heavily on the idea Andy mentioned of keeping a worklist of blocks that are viable layout successors based on the CFG. Having this set allows us to consistently select the best layout successor for each block. It is expensive though. The code here remains very rough. There is a lot that needs to be done to clean up the code, and to make the runtime cost of this pass much lower. Very much WIP, but this was a giant chunk of code and I'd rather folks see it sooner than later. Everything remains behind a flag of course. I've added a couple of tests to exercise the issues that this iteration was motivated by: loop structure preservation. I've also fixed one test that was exhibiting the broken behavior of the previous version. llvm-svn: 144495
2011-11-13 12:20:44 +01:00
define i32 @test_loop_cold_blocks(i32 %i, i32* %a) {
; Check that we sink cold loop blocks after the hot loop body.
; CHECK: test_loop_cold_blocks:
; CHECK: %entry
; CHECK: %body1
; CHECK: %body2
; CHECK: %body3
; CHECK: %unlikely1
; CHECK: %unlikely2
; CHECK: %exit
entry:
br label %body1
body1:
%iv = phi i32 [ 0, %entry ], [ %next, %body3 ]
%base = phi i32 [ 0, %entry ], [ %sum, %body3 ]
%unlikelycond1 = icmp slt i32 %base, 42
br i1 %unlikelycond1, label %unlikely1, label %body2, !prof !0
unlikely1:
call void @error(i32 %i, i32 1, i32 %base)
br label %body2
body2:
%unlikelycond2 = icmp sgt i32 %base, 21
br i1 %unlikelycond2, label %unlikely2, label %body3, !prof !0
unlikely2:
call void @error(i32 %i, i32 2, i32 %base)
br label %body3
body3:
%arrayidx = getelementptr inbounds i32* %a, i32 %iv
%0 = load i32* %arrayidx
%sum = add nsw i32 %0, %base
%next = add i32 %iv, 1
%exitcond = icmp eq i32 %next, %i
br i1 %exitcond, label %exit, label %body1
exit:
ret i32 %sum
}
Add a very basic test for MachineBlockPlacement. This is essentially the canonical example I used when developing it, and is one of the primary motivating real-world use cases for __builtin_expect (when burried under a macro). I'm working on more test cases here, but I'm trying to make sure both that the pass is doing the right thing with the test cases and that they aren't too brittle to changes elsewhere in the code generation pipeline. Feedback and/or suggestions on how to test this are very welcome. Especially feedback on whether testing the block comments is a good strategy; I couldn't find any good examples to steal from but all the other ideas I had were a lot uglier or more fragile. llvm-svn: 142644
2011-10-21 10:01:56 +02:00
!0 = metadata !{metadata !"branch_weights", i32 4, i32 64}
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
Rewrite #3 of machine block placement. This is based somewhat on the second algorithm, but only loosely. It is more heavily based on the last discussion I had with Andy. It continues to walk from the inner-most loop outward, but there is a key difference. With this algorithm we ensure that as we visit each loop, the entire loop is merged into a single chain. At the end, the entire function is treated as a "loop", and merged into a single chain. This chain forms the desired sequence of blocks within the function. Switching to a single algorithm removes my biggest problem with the previous approaches -- they had different behavior depending on which system triggered the layout. Now there is exactly one algorithm and one basis for the decision making. The other key difference is how the chain is formed. This is based heavily on the idea Andy mentioned of keeping a worklist of blocks that are viable layout successors based on the CFG. Having this set allows us to consistently select the best layout successor for each block. It is expensive though. The code here remains very rough. There is a lot that needs to be done to clean up the code, and to make the runtime cost of this pass much lower. Very much WIP, but this was a giant chunk of code and I'd rather folks see it sooner than later. Everything remains behind a flag of course. I've added a couple of tests to exercise the issues that this iteration was motivated by: loop structure preservation. I've also fixed one test that was exhibiting the broken behavior of the previous version. llvm-svn: 144495
2011-11-13 12:20:44 +01:00
define i32 @test_loop_early_exits(i32 %i, i32* %a) {
; Check that we sink early exit blocks out of loop bodies.
; CHECK: test_loop_early_exits:
; CHECK: %entry
; CHECK: %body1
; CHECK: %body2
; CHECK: %body3
; CHECK: %body4
; CHECK: %exit
; CHECK: %bail1
; CHECK: %bail2
; CHECK: %bail3
entry:
br label %body1
body1:
%iv = phi i32 [ 0, %entry ], [ %next, %body4 ]
%base = phi i32 [ 0, %entry ], [ %sum, %body4 ]
%bailcond1 = icmp eq i32 %base, 42
br i1 %bailcond1, label %bail1, label %body2
bail1:
ret i32 -1
body2:
%bailcond2 = icmp eq i32 %base, 43
br i1 %bailcond2, label %bail2, label %body3
bail2:
ret i32 -2
body3:
%bailcond3 = icmp eq i32 %base, 44
br i1 %bailcond3, label %bail3, label %body4
bail3:
ret i32 -3
body4:
%arrayidx = getelementptr inbounds i32* %a, i32 %iv
%0 = load i32* %arrayidx
%sum = add nsw i32 %0, %base
%next = add i32 %iv, 1
%exitcond = icmp eq i32 %next, %i
br i1 %exitcond, label %exit, label %body1
exit:
ret i32 %sum
}
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
define i32 @test_loop_align(i32 %i, i32* %a) {
; Check that we provide basic loop body alignment with the block placement
; pass.
; CHECK: test_loop_align:
; CHECK: %entry
Don't hard code the desired alignment for loops -- it isn't 16-bytes on all x86 systems. Sorry for the breakage. llvm-svn: 142656
2011-10-21 18:41:39 +02:00
; CHECK: .align [[ALIGN:[0-9]+]],
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
; CHECK-NEXT: %body
; CHECK: %exit
entry:
br label %body
body:
%iv = phi i32 [ 0, %entry ], [ %next, %body ]
%base = phi i32 [ 0, %entry ], [ %sum, %body ]
%arrayidx = getelementptr inbounds i32* %a, i32 %iv
%0 = load i32* %arrayidx
%sum = add nsw i32 %0, %base
%next = add i32 %iv, 1
%exitcond = icmp eq i32 %next, %i
br i1 %exitcond, label %exit, label %body
exit:
ret i32 %sum
}
define i32 @test_nested_loop_align(i32 %i, i32* %a, i32* %b) {
; Check that we provide nested loop body alignment.
; CHECK: test_nested_loop_align:
; CHECK: %entry
Don't hard code the desired alignment for loops -- it isn't 16-bytes on all x86 systems. Sorry for the breakage. llvm-svn: 142656
2011-10-21 18:41:39 +02:00
; CHECK: .align [[ALIGN]],
Rewrite #3 of machine block placement. This is based somewhat on the second algorithm, but only loosely. It is more heavily based on the last discussion I had with Andy. It continues to walk from the inner-most loop outward, but there is a key difference. With this algorithm we ensure that as we visit each loop, the entire loop is merged into a single chain. At the end, the entire function is treated as a "loop", and merged into a single chain. This chain forms the desired sequence of blocks within the function. Switching to a single algorithm removes my biggest problem with the previous approaches -- they had different behavior depending on which system triggered the layout. Now there is exactly one algorithm and one basis for the decision making. The other key difference is how the chain is formed. This is based heavily on the idea Andy mentioned of keeping a worklist of blocks that are viable layout successors based on the CFG. Having this set allows us to consistently select the best layout successor for each block. It is expensive though. The code here remains very rough. There is a lot that needs to be done to clean up the code, and to make the runtime cost of this pass much lower. Very much WIP, but this was a giant chunk of code and I'd rather folks see it sooner than later. Everything remains behind a flag of course. I've added a couple of tests to exercise the issues that this iteration was motivated by: loop structure preservation. I've also fixed one test that was exhibiting the broken behavior of the previous version. llvm-svn: 144495
2011-11-13 12:20:44 +01:00
; CHECK-NEXT: %loop.body.1
Don't hard code the desired alignment for loops -- it isn't 16-bytes on all x86 systems. Sorry for the breakage. llvm-svn: 142656
2011-10-21 18:41:39 +02:00
; CHECK: .align [[ALIGN]],
Add loop aligning to MachineBlockPlacement based on review discussion so it's a bit more plausible to use this instead of CodePlacementOpt. The code for this was shamelessly stolen from CodePlacementOpt, and then trimmed down a bit. There doesn't seem to be much utility in returning true/false from this pass as we may or may not have rewritten all of the blocks. Also, the statistic of counting how many loops were aligned doesn't seem terribly important so I removed it. If folks would like it to be included, I'm happy to add it back. This was probably the most egregious of the missing features, and now I'm going to start gathering some performance numbers and looking at specific loop structures that have different layout between the two. Test is updated to include both basic loop alignment and nested loop alignment. llvm-svn: 142645
2011-10-21 10:57:37 +02:00
; CHECK-NEXT: %inner.loop.body
; CHECK-NOT: .align
; CHECK: %exit
entry:
br label %loop.body.1
loop.body.1:
%iv = phi i32 [ 0, %entry ], [ %next, %loop.body.2 ]
%arrayidx = getelementptr inbounds i32* %a, i32 %iv
%bidx = load i32* %arrayidx
br label %inner.loop.body
inner.loop.body:
%inner.iv = phi i32 [ 0, %loop.body.1 ], [ %inner.next, %inner.loop.body ]
%base = phi i32 [ 0, %loop.body.1 ], [ %sum, %inner.loop.body ]
%scaled_idx = mul i32 %bidx, %iv
%inner.arrayidx = getelementptr inbounds i32* %b, i32 %scaled_idx
%0 = load i32* %inner.arrayidx
%sum = add nsw i32 %0, %base
%inner.next = add i32 %iv, 1
%inner.exitcond = icmp eq i32 %inner.next, %i
br i1 %inner.exitcond, label %loop.body.2, label %inner.loop.body
loop.body.2:
%next = add i32 %iv, 1
%exitcond = icmp eq i32 %next, %i
br i1 %exitcond, label %exit, label %loop.body.1
exit:
ret i32 %sum
}
Teach machine block placement to cope with unnatural loops. These don't get loop info structures associated with them, and so we need some way to make forward progress selecting and placing basic blocks. The technique used here is pretty brutal -- it just scans the list of blocks looking for the first unplaced candidate. It keeps placing blocks like this until the CFG becomes tractable. The cost is somewhat unfortunate, it requires allocating a vector of all basic block pointers eagerly. I have some ideas about how to simplify and optimize this, but I'm trying to get the logic correct first. Thanks to Benjamin Kramer for the reduced test case out of GCC. Sadly there are other bugs that GCC is tickling that I'm reducing and working on now. llvm-svn: 144516
2011-11-14 01:00:35 +01:00
define void @unnatural_cfg1() {
; Test that we can handle a loop with an inner unnatural loop at the end of
; a function. This is a gross CFG reduced out of the single source GCC.
; CHECK: unnatural_cfg1
; CHECK: %entry
; CHECK: %loop.body1
; CHECK: %loop.body2
Rather than trying to use the loop block sequence *or* the function block sequence when recovering from unanalyzable control flow constructs, *always* use the function sequence. I'm not sure why I ever went down the path of trying to use the loop sequence, it is fundamentally not the correct sequence to use. We're trying to preserve the incoming layout in the cases of unreasonable control flow, and that is only encoded at the function level. We already have a filter to select *exactly* the sub-set of blocks within the function that we're trying to form into a chain. The resulting code layout is also significantly better because of this. In several places we were ending up with completely unreasonable control flow constructs due to the ordering chosen by the loop structure for its internal storage. This change removes a completely wasteful vector of basic blocks, saving memory allocation in the common case even though it costs us CPU in the fairly rare case of unnatural loops. Finally, it fixes the latest crasher reduced out of GCC's single source. Thanks again to Benjamin Kramer for the reduction, my bugpoint skills failed at it. llvm-svn: 144627
2011-11-15 07:26:43 +01:00
; CHECK: %loop.body3
Teach machine block placement to cope with unnatural loops. These don't get loop info structures associated with them, and so we need some way to make forward progress selecting and placing basic blocks. The technique used here is pretty brutal -- it just scans the list of blocks looking for the first unplaced candidate. It keeps placing blocks like this until the CFG becomes tractable. The cost is somewhat unfortunate, it requires allocating a vector of all basic block pointers eagerly. I have some ideas about how to simplify and optimize this, but I'm trying to get the logic correct first. Thanks to Benjamin Kramer for the reduced test case out of GCC. Sadly there are other bugs that GCC is tickling that I'm reducing and working on now. llvm-svn: 144516
2011-11-14 01:00:35 +01:00
entry:
br label %loop.header
loop.header:
br label %loop.body1
loop.body1:
br i1 undef, label %loop.body3, label %loop.body2
loop.body2:
%ptr = load i32** undef, align 4
br label %loop.body3
loop.body3:
%myptr = phi i32* [ %ptr2, %loop.body5 ], [ %ptr, %loop.body2 ], [ undef, %loop.body1 ]
%bcmyptr = bitcast i32* %myptr to i32*
%val = load i32* %bcmyptr, align 4
%comp = icmp eq i32 %val, 48
br i1 %comp, label %loop.body4, label %loop.body5
loop.body4:
br i1 undef, label %loop.header, label %loop.body5
loop.body5:
%ptr2 = load i32** undef, align 4
br label %loop.body3
}
Fix an overflow bug in MachineBranchProbabilityInfo. This pass relied on the sum of the edge weights not overflowing uint32, and crashed when they did. This is generally safe as BranchProbabilityInfo tries to provide this guarantee. However, the CFG can get modified during codegen in a way that grows the *sum* of the edge weights. This doesn't seem unreasonable (imagine just adding more blocks all with the default weight of 16), but it is hard to come up with a case that actually triggers 32-bit overflow. Fortuately, the single-source GCC build is good at this. The solution isn't very pretty, but its no worse than the previous code. We're already summing all of the edge weights on each query, we can sum them, check for an overflow, compute a scale, and sum them again. I've included a *greatly* reduced test case out of the GCC source that triggers it. It's a pretty lame test, as it clearly is just barely triggering the overflow. I'd like to have something that is much more definitive, but I don't understand the fundamental pattern that triggers an explosion in the edge weight sums. The buggy code is duplicated within this file. I'll colapse them into a single implementation in a subsequent commit. llvm-svn: 144526
2011-11-14 09:50:16 +01:00
Rather than trying to use the loop block sequence *or* the function block sequence when recovering from unanalyzable control flow constructs, *always* use the function sequence. I'm not sure why I ever went down the path of trying to use the loop sequence, it is fundamentally not the correct sequence to use. We're trying to preserve the incoming layout in the cases of unreasonable control flow, and that is only encoded at the function level. We already have a filter to select *exactly* the sub-set of blocks within the function that we're trying to form into a chain. The resulting code layout is also significantly better because of this. In several places we were ending up with completely unreasonable control flow constructs due to the ordering chosen by the loop structure for its internal storage. This change removes a completely wasteful vector of basic blocks, saving memory allocation in the common case even though it costs us CPU in the fairly rare case of unnatural loops. Finally, it fixes the latest crasher reduced out of GCC's single source. Thanks again to Benjamin Kramer for the reduction, my bugpoint skills failed at it. llvm-svn: 144627
2011-11-15 07:26:43 +01:00
define void @unnatural_cfg2() {
; Test that we can handle a loop with a nested natural loop *and* an unnatural
; loop. This was reduced from a crash on block placement when run over
; single-source GCC.
; CHECK: unnatural_cfg2
; CHECK: %entry
; CHECK: %loop.header
; CHECK: %loop.body1
; CHECK: %loop.body2
; CHECK: %loop.body3
; CHECK: %loop.inner1.begin
; The end block is folded with %loop.body3...
; CHECK-NOT: %loop.inner1.end
; CHECK: %loop.body4
; CHECK: %loop.inner2.begin
; The loop.inner2.end block is folded
; CHECK: %bail
entry:
br label %loop.header
loop.header:
%comp0 = icmp eq i32* undef, null
br i1 %comp0, label %bail, label %loop.body1
loop.body1:
%val0 = load i32** undef, align 4
br i1 undef, label %loop.body2, label %loop.inner1.begin
loop.body2:
br i1 undef, label %loop.body4, label %loop.body3
loop.body3:
%ptr1 = getelementptr inbounds i32* %val0, i32 0
%castptr1 = bitcast i32* %ptr1 to i32**
%val1 = load i32** %castptr1, align 4
br label %loop.inner1.begin
loop.inner1.begin:
%valphi = phi i32* [ %val2, %loop.inner1.end ], [ %val1, %loop.body3 ], [ %val0, %loop.body1 ]
%castval = bitcast i32* %valphi to i32*
%comp1 = icmp eq i32 undef, 48
br i1 %comp1, label %loop.inner1.end, label %loop.body4
loop.inner1.end:
%ptr2 = getelementptr inbounds i32* %valphi, i32 0
%castptr2 = bitcast i32* %ptr2 to i32**
%val2 = load i32** %castptr2, align 4
br label %loop.inner1.begin
loop.body4.dead:
br label %loop.body4
loop.body4:
%comp2 = icmp ult i32 undef, 3
br i1 %comp2, label %loop.inner2.begin, label %loop.end
loop.inner2.begin:
br i1 false, label %loop.end, label %loop.inner2.end
loop.inner2.end:
%comp3 = icmp eq i32 undef, 1769472
br i1 %comp3, label %loop.end, label %loop.inner2.begin
loop.end:
br label %loop.header
bail:
unreachable
}
Fix an overflow bug in MachineBranchProbabilityInfo. This pass relied on the sum of the edge weights not overflowing uint32, and crashed when they did. This is generally safe as BranchProbabilityInfo tries to provide this guarantee. However, the CFG can get modified during codegen in a way that grows the *sum* of the edge weights. This doesn't seem unreasonable (imagine just adding more blocks all with the default weight of 16), but it is hard to come up with a case that actually triggers 32-bit overflow. Fortuately, the single-source GCC build is good at this. The solution isn't very pretty, but its no worse than the previous code. We're already summing all of the edge weights on each query, we can sum them, check for an overflow, compute a scale, and sum them again. I've included a *greatly* reduced test case out of the GCC source that triggers it. It's a pretty lame test, as it clearly is just barely triggering the overflow. I'd like to have something that is much more definitive, but I don't understand the fundamental pattern that triggers an explosion in the edge weight sums. The buggy code is duplicated within this file. I'll colapse them into a single implementation in a subsequent commit. llvm-svn: 144526
2011-11-14 09:50:16 +01:00
define i32 @problematic_switch() {
; This function's CFG caused overlow in the machine branch probability
; calculation, triggering asserts. Make sure we don't crash on it.
; CHECK: problematic_switch
entry:
switch i32 undef, label %exit [
i32 879, label %bogus
i32 877, label %step
i32 876, label %step
i32 875, label %step
i32 874, label %step
i32 873, label %step
i32 872, label %step
i32 868, label %step
i32 867, label %step
i32 866, label %step
i32 861, label %step
i32 860, label %step
i32 856, label %step
i32 855, label %step
i32 854, label %step
i32 831, label %step
i32 830, label %step
i32 829, label %step
i32 828, label %step
i32 815, label %step
i32 814, label %step
i32 811, label %step
i32 806, label %step
i32 805, label %step
i32 804, label %step
i32 803, label %step
i32 802, label %step
i32 801, label %step
i32 800, label %step
i32 799, label %step
i32 798, label %step
i32 797, label %step
i32 796, label %step
i32 795, label %step
]
bogus:
unreachable
step:
br label %exit
exit:
%merge = phi i32 [ 3, %step ], [ 6, %entry ]
ret i32 %merge
}
Reference in New Issue Copy Permalink