RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-30 15:32:52 +01:00
Code Issues Projects Releases Wiki Activity
0917f4a575
llvm-mirror/test/Transforms/GVN/rle.ll

535 lines
11 KiB
LLVM
Raw Normal View History

enhance GVN to forward substitute a stored value to a load (and load -> load) when the base pointers must alias but when they are different types. This occurs very very frequently in 176.gcc and other code that uses bitfields a lot. llvm-svn: 82399
2009-09-20 21:03:47 +02:00
; RUN: opt < %s -gvn -S | FileCheck %s
; 32-bit little endian target.
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
;; Trivial RLE test.
define i32 @test0(i32 %V, i32* %P) {
store i32 %V, i32* %P
%A = load i32* %P
ret i32 %A
; CHECK: @test0
; CHECK: ret i32 %V
}
fix PR5016, a crash I introduced in GVN handing first class arrays and structs, which cannot be bitcast to integers. llvm-svn: 82460
2009-09-21 19:24:04 +02:00
;;===----------------------------------------------------------------------===;;
;; Tests for crashers
;;===----------------------------------------------------------------------===;;
;; PR5016
define i8 @crash0({i32, i32} %A, {i32, i32}* %P) {
store {i32, i32} %A, {i32, i32}* %P
%X = bitcast {i32, i32}* %P to i8*
%Y = load i8* %X
ret i8 %Y
}
enhance GVN to forward substitute a stored value to a load (and load -> load) when the base pointers must alias but when they are different types. This occurs very very frequently in 176.gcc and other code that uses bitfields a lot. llvm-svn: 82399
2009-09-20 21:03:47 +02:00
;;===----------------------------------------------------------------------===;;
;; Store -> Load and Load -> Load forwarding where src and dst are different
;; types, but where the base pointer is a must alias.
;;===----------------------------------------------------------------------===;;
;; i32 -> f32 forwarding.
define float @coerce_mustalias1(i32 %V, i32* %P) {
store i32 %V, i32* %P
%P2 = bitcast i32* %P to float*
%A = load float* %P2
ret float %A
; CHECK: @coerce_mustalias1
; CHECK-NOT: load
; CHECK: ret float
}
;; i32* -> float forwarding.
define float @coerce_mustalias2(i32* %V, i32** %P) {
store i32* %V, i32** %P
%P2 = bitcast i32** %P to float*
%A = load float* %P2
ret float %A
; CHECK: @coerce_mustalias2
; CHECK-NOT: load
; CHECK: ret float
}
;; float -> i32* forwarding.
define i32* @coerce_mustalias3(float %V, float* %P) {
store float %V, float* %P
%P2 = bitcast float* %P to i32**
%A = load i32** %P2
ret i32* %A
; CHECK: @coerce_mustalias3
; CHECK-NOT: load
; CHECK: ret i32*
}
;; i32 -> f32 load forwarding.
define float @coerce_mustalias4(i32* %P, i1 %cond) {
%A = load i32* %P
Move CoerceAvailableValueToLoadType earlier in GVN.cpp. Hook it up so that nonlocal and partially redundant loads can use it as well. The testcase shows examples of craziness this can handle. This triggers *many* times in 176.gcc. llvm-svn: 82403
2009-09-20 22:09:34 +02:00
enhance GVN to forward substitute a stored value to a load (and load -> load) when the base pointers must alias but when they are different types. This occurs very very frequently in 176.gcc and other code that uses bitfields a lot. llvm-svn: 82399
2009-09-20 21:03:47 +02:00
%P2 = bitcast i32* %P to float*
%B = load float* %P2
Move CoerceAvailableValueToLoadType earlier in GVN.cpp. Hook it up so that nonlocal and partially redundant loads can use it as well. The testcase shows examples of craziness this can handle. This triggers *many* times in 176.gcc. llvm-svn: 82403
2009-09-20 22:09:34 +02:00
br i1 %cond, label %T, label %F
T:
enhance GVN to forward substitute a stored value to a load (and load -> load) when the base pointers must alias but when they are different types. This occurs very very frequently in 176.gcc and other code that uses bitfields a lot. llvm-svn: 82399
2009-09-20 21:03:47 +02:00
ret float %B
F:
%X = bitcast i32 %A to float
ret float %X
; CHECK: @coerce_mustalias4
; CHECK: %A = load i32* %P
; CHECK-NOT: load
; CHECK: ret float
; CHECK: F:
}
;; i32 -> i8 forwarding
define i8 @coerce_mustalias5(i32 %V, i32* %P) {
store i32 %V, i32* %P
%P2 = bitcast i32* %P to i8*
%A = load i8* %P2
ret i8 %A
; CHECK: @coerce_mustalias5
; CHECK-NOT: load
; CHECK: ret i8
}
;; i64 -> float forwarding
define float @coerce_mustalias6(i64 %V, i64* %P) {
store i64 %V, i64* %P
%P2 = bitcast i64* %P to float*
%A = load float* %P2
ret float %A
; CHECK: @coerce_mustalias6
; CHECK-NOT: load
; CHECK: ret float
}
;; i64 -> i8* (32-bit) forwarding
define i8* @coerce_mustalias7(i64 %V, i64* %P) {
store i64 %V, i64* %P
%P2 = bitcast i64* %P to i8**
%A = load i8** %P2
ret i8* %A
; CHECK: @coerce_mustalias7
; CHECK-NOT: load
; CHECK: ret i8*
}
Handle forwarding local memsets to loads. For example, we optimize this: short x(short *A) { memset(A, 1, sizeof(*A)*100); return A[42]; } to 'return 257' instead of doing the load. llvm-svn: 90695
2009-12-06 02:57:02 +01:00
; memset -> i16 forwarding.
define signext i16 @memset_to_i16_local(i16* %A) nounwind ssp {
entry:
%conv = bitcast i16* %A to i8*
tail call void @llvm.memset.i64(i8* %conv, i8 1, i64 200, i32 1)
%arrayidx = getelementptr inbounds i16* %A, i64 42
%tmp2 = load i16* %arrayidx
ret i16 %tmp2
; CHECK: @memset_to_i16_local
; CHECK-NOT: load
; CHECK: ret i16 257
}
; memset -> float forwarding.
define float @memset_to_float_local(float* %A, i8 %Val) nounwind ssp {
entry:
%conv = bitcast float* %A to i8* ; <i8*> [#uses=1]
tail call void @llvm.memset.i64(i8* %conv, i8 %Val, i64 400, i32 1)
%arrayidx = getelementptr inbounds float* %A, i64 42 ; <float*> [#uses=1]
%tmp2 = load float* %arrayidx ; <float> [#uses=1]
ret float %tmp2
; CHECK: @memset_to_float_local
; CHECK-NOT: load
; CHECK: zext
; CHECK-NEXT: shl
; CHECK-NEXT: or
; CHECK-NEXT: shl
; CHECK-NEXT: or
; CHECK-NEXT: bitcast
; CHECK-NEXT: ret float
}
add support for forwarding mem intrinsic values to non-local loads. llvm-svn: 90697
2009-12-06 05:54:31 +01:00
;; non-local memset -> i16 load forwarding.
define i16 @memset_to_i16_nonlocal0(i16* %P, i1 %cond) {
%P3 = bitcast i16* %P to i8*
br i1 %cond, label %T, label %F
T:
tail call void @llvm.memset.i64(i8* %P3, i8 1, i64 400, i32 1)
br label %Cont
F:
tail call void @llvm.memset.i64(i8* %P3, i8 2, i64 400, i32 1)
br label %Cont
Cont:
%P2 = getelementptr i16* %P, i32 4
%A = load i16* %P2
ret i16 %A
; CHECK: @memset_to_i16_nonlocal0
; CHECK: Cont:
; CHECK-NEXT: %A = phi i16 [ 514, %F ], [ 257, %T ]
; CHECK-NOT: load
; CHECK: ret i16 %A
}
constant fold loads from memcpy's from global constants. This is important because clang lowers nontrivial automatic struct/array inits to memcpy from a global array. llvm-svn: 90698
2009-12-06 06:29:56 +01:00
@GCst = constant {i32, float, i32 } { i32 42, float 14., i32 97 }
; memset -> float forwarding.
define float @memcpy_to_float_local(float* %A) nounwind ssp {
entry:
%conv = bitcast float* %A to i8* ; <i8*> [#uses=1]
tail call void @llvm.memcpy.i64(i8* %conv, i8* bitcast ({i32, float, i32 }* @GCst to i8*), i64 12, i32 1)
%arrayidx = getelementptr inbounds float* %A, i64 1 ; <float*> [#uses=1]
%tmp2 = load float* %arrayidx ; <float> [#uses=1]
ret float %tmp2
; CHECK: @memcpy_to_float_local
; CHECK-NOT: load
; CHECK: ret float 1.400000e+01
}
add support for forwarding mem intrinsic values to non-local loads. llvm-svn: 90697
2009-12-06 05:54:31 +01:00
Handle forwarding local memsets to loads. For example, we optimize this: short x(short *A) { memset(A, 1, sizeof(*A)*100); return A[42]; } to 'return 257' instead of doing the load. llvm-svn: 90695
2009-12-06 02:57:02 +01:00
declare void @llvm.memset.i64(i8* nocapture, i8, i64, i32) nounwind
constant fold loads from memcpy's from global constants. This is important because clang lowers nontrivial automatic struct/array inits to memcpy from a global array. llvm-svn: 90698
2009-12-06 06:29:56 +01:00
declare void @llvm.memcpy.i64(i8* nocapture, i8* nocapture, i64, i32) nounwind
Handle forwarding local memsets to loads. For example, we optimize this: short x(short *A) { memset(A, 1, sizeof(*A)*100); return A[42]; } to 'return 257' instead of doing the load. llvm-svn: 90695
2009-12-06 02:57:02 +01:00
Move CoerceAvailableValueToLoadType earlier in GVN.cpp. Hook it up so that nonlocal and partially redundant loads can use it as well. The testcase shows examples of craziness this can handle. This triggers *many* times in 176.gcc. llvm-svn: 82403
2009-09-20 22:09:34 +02:00
;; non-local i32/float -> i8 load forwarding.
define i8 @coerce_mustalias_nonlocal0(i32* %P, i1 %cond) {
%P2 = bitcast i32* %P to float*
%P3 = bitcast i32* %P to i8*
br i1 %cond, label %T, label %F
T:
store i32 42, i32* %P
br label %Cont
F:
store float 1.0, float* %P2
br label %Cont
Cont:
%A = load i8* %P3
ret i8 %A
; CHECK: @coerce_mustalias_nonlocal0
; CHECK: Cont:
; CHECK: %A = phi i8 [
; CHECK-NOT: load
; CHECK: ret i8 %A
}
add support for forwarding mem intrinsic values to non-local loads. llvm-svn: 90697
2009-12-06 05:54:31 +01:00
Improve GVN to be able to forward substitute a small load from a piece of a large store when both are in the same block. This allows clang to compile the testcase in PR4216 to this code: _test_bitfield: movl 4(%esp), %eax movl %eax, %ecx andl $-65536, %ecx orl $32962, %eax andl $40186, %eax orl %ecx, %eax ret This is not ideal, but is a whole lot better than the code produced by llvm-gcc: _test_bitfield: movw $-32574, %ax orw 4(%esp), %ax andw $-25350, %ax movw %ax, 4(%esp) movw 7(%esp), %cx shlw $8, %cx movzbl 6(%esp), %edx orw %cx, %dx movzwl %dx, %ecx shll $16, %ecx movzwl %ax, %eax orl %ecx, %eax ret and dramatically better than that produced by gcc 4.2: _test_bitfield: pushl %ebx call L3 "L00000000001$pb": L3: popl %ebx movl 8(%esp), %eax leal 0(,%eax,4), %edx sarb $7, %dl movl %eax, %ecx andl $7168, %ecx andl $-7201, %ebx movzbl %dl, %edx andl $1, %edx sall $5, %edx orl %ecx, %ebx orl %edx, %ebx andl $24, %eax andl $-58336, %ebx orl %eax, %ebx orl $32962, %ebx movl %ebx, %eax popl %ebx ret llvm-svn: 82439
2009-09-21 07:57:11 +02:00
;; non-local i32/float -> i8 load forwarding. This also tests that the "P3"
;; bitcast equivalence can be properly phi translated.
define i8 @coerce_mustalias_nonlocal1(i32* %P, i1 %cond) {
%P2 = bitcast i32* %P to float*
br i1 %cond, label %T, label %F
T:
store i32 42, i32* %P
br label %Cont
F:
store float 1.0, float* %P2
br label %Cont
Cont:
%P3 = bitcast i32* %P to i8*
%A = load i8* %P3
ret i8 %A
;; FIXME: This is disabled because this caused a miscompile in the llvm-gcc
;; bootstrap, see r82411
;
; HECK: @coerce_mustalias_nonlocal1
; HECK: Cont:
; HECK: %A = phi i8 [
; HECK-NOT: load
; HECK: ret i8 %A
}
Move CoerceAvailableValueToLoadType earlier in GVN.cpp. Hook it up so that nonlocal and partially redundant loads can use it as well. The testcase shows examples of craziness this can handle. This triggers *many* times in 176.gcc. llvm-svn: 82403
2009-09-20 22:09:34 +02:00
;; non-local i32 -> i8 partial redundancy load forwarding.
define i8 @coerce_mustalias_pre0(i32* %P, i1 %cond) {
%P3 = bitcast i32* %P to i8*
br i1 %cond, label %T, label %F
T:
store i32 42, i32* %P
br label %Cont
F:
br label %Cont
Cont:
%A = load i8* %P3
ret i8 %A
; CHECK: @coerce_mustalias_pre0
; CHECK: F:
; CHECK: load i8* %P3
; CHECK: Cont:
; CHECK: %A = phi i8 [
; CHECK-NOT: load
; CHECK: ret i8 %A
}
Improve GVN to be able to forward substitute a small load from a piece of a large store when both are in the same block. This allows clang to compile the testcase in PR4216 to this code: _test_bitfield: movl 4(%esp), %eax movl %eax, %ecx andl $-65536, %ecx orl $32962, %eax andl $40186, %eax orl %ecx, %eax ret This is not ideal, but is a whole lot better than the code produced by llvm-gcc: _test_bitfield: movw $-32574, %ax orw 4(%esp), %ax andw $-25350, %ax movw %ax, 4(%esp) movw 7(%esp), %cx shlw $8, %cx movzbl 6(%esp), %edx orw %cx, %dx movzwl %dx, %ecx shll $16, %ecx movzwl %ax, %eax orl %ecx, %eax ret and dramatically better than that produced by gcc 4.2: _test_bitfield: pushl %ebx call L3 "L00000000001$pb": L3: popl %ebx movl 8(%esp), %eax leal 0(,%eax,4), %edx sarb $7, %dl movl %eax, %ecx andl $7168, %ecx andl $-7201, %ebx movzbl %dl, %edx andl $1, %edx sall $5, %edx orl %ecx, %ebx orl %edx, %ebx andl $24, %eax andl $-58336, %ebx orl %eax, %ebx orl $32962, %ebx movl %ebx, %eax popl %ebx ret llvm-svn: 82439
2009-09-21 07:57:11 +02:00
;;===----------------------------------------------------------------------===;;
;; Store -> Load and Load -> Load forwarding where src and dst are different
;; types, and the reload is an offset from the store pointer.
;;===----------------------------------------------------------------------===;;
enable non-local analysis and PRE of large store -> little load. This doesn't kick in too much because of phi translation issues, but this can be resolved in the future. llvm-svn: 82447
2009-09-21 08:48:08 +02:00
;; i32 -> i8 forwarding.
add pr# llvm-svn: 82440
2009-09-21 07:57:47 +02:00
;; PR4216
Improve GVN to be able to forward substitute a small load from a piece of a large store when both are in the same block. This allows clang to compile the testcase in PR4216 to this code: _test_bitfield: movl 4(%esp), %eax movl %eax, %ecx andl $-65536, %ecx orl $32962, %eax andl $40186, %eax orl %ecx, %eax ret This is not ideal, but is a whole lot better than the code produced by llvm-gcc: _test_bitfield: movw $-32574, %ax orw 4(%esp), %ax andw $-25350, %ax movw %ax, 4(%esp) movw 7(%esp), %cx shlw $8, %cx movzbl 6(%esp), %edx orw %cx, %dx movzwl %dx, %ecx shll $16, %ecx movzwl %ax, %eax orl %ecx, %eax ret and dramatically better than that produced by gcc 4.2: _test_bitfield: pushl %ebx call L3 "L00000000001$pb": L3: popl %ebx movl 8(%esp), %eax leal 0(,%eax,4), %edx sarb $7, %dl movl %eax, %ecx andl $7168, %ecx andl $-7201, %ebx movzbl %dl, %edx andl $1, %edx sall $5, %edx orl %ecx, %ebx orl %edx, %ebx andl $24, %eax andl $-58336, %ebx orl %eax, %ebx orl $32962, %ebx movl %ebx, %eax popl %ebx ret llvm-svn: 82439
2009-09-21 07:57:11 +02:00
define i8 @coerce_offset0(i32 %V, i32* %P) {
store i32 %V, i32* %P
%P2 = bitcast i32* %P to i8*
%P3 = getelementptr i8* %P2, i32 2
%A = load i8* %P3
ret i8 %A
; CHECK: @coerce_offset0
; CHECK-NOT: load
; CHECK: ret i8
}
enable non-local analysis and PRE of large store -> little load. This doesn't kick in too much because of phi translation issues, but this can be resolved in the future. llvm-svn: 82447
2009-09-21 08:48:08 +02:00
;; non-local i32/float -> i8 load forwarding.
define i8 @coerce_offset_nonlocal0(i32* %P, i1 %cond) {
%P2 = bitcast i32* %P to float*
%P3 = bitcast i32* %P to i8*
%P4 = getelementptr i8* %P3, i32 2
br i1 %cond, label %T, label %F
T:
store i32 42, i32* %P
br label %Cont
F:
store float 1.0, float* %P2
br label %Cont
Cont:
%A = load i8* %P4
ret i8 %A
; CHECK: @coerce_offset_nonlocal0
; CHECK: Cont:
; CHECK: %A = phi i8 [
; CHECK-NOT: load
; CHECK: ret i8 %A
}
;; non-local i32 -> i8 partial redundancy load forwarding.
define i8 @coerce_offset_pre0(i32* %P, i1 %cond) {
%P3 = bitcast i32* %P to i8*
%P4 = getelementptr i8* %P3, i32 2
br i1 %cond, label %T, label %F
T:
store i32 42, i32* %P
br label %Cont
F:
br label %Cont
Cont:
%A = load i8* %P4
ret i8 %A
; CHECK: @coerce_offset_pre0
; CHECK: F:
; CHECK: load i8* %P4
; CHECK: Cont:
; CHECK: %A = phi i8 [
; CHECK-NOT: load
; CHECK: ret i8 %A
}
Improve GVN to be able to forward substitute a small load from a piece of a large store when both are in the same block. This allows clang to compile the testcase in PR4216 to this code: _test_bitfield: movl 4(%esp), %eax movl %eax, %ecx andl $-65536, %ecx orl $32962, %eax andl $40186, %eax orl %ecx, %eax ret This is not ideal, but is a whole lot better than the code produced by llvm-gcc: _test_bitfield: movw $-32574, %ax orw 4(%esp), %ax andw $-25350, %ax movw %ax, 4(%esp) movw 7(%esp), %cx shlw $8, %cx movzbl 6(%esp), %edx orw %cx, %dx movzwl %dx, %ecx shll $16, %ecx movzwl %ax, %eax orl %ecx, %eax ret and dramatically better than that produced by gcc 4.2: _test_bitfield: pushl %ebx call L3 "L00000000001$pb": L3: popl %ebx movl 8(%esp), %eax leal 0(,%eax,4), %edx sarb $7, %dl movl %eax, %ecx andl $7168, %ecx andl $-7201, %ebx movzbl %dl, %edx andl $1, %edx sall $5, %edx orl %ecx, %ebx orl %edx, %ebx andl $24, %eax andl $-58336, %ebx orl %eax, %ebx orl $32962, %ebx movl %ebx, %eax popl %ebx ret llvm-svn: 82439
2009-09-21 07:57:11 +02:00
merge two tests. llvm-svn: 90691
2009-12-06 02:47:24 +01:00
define i32 @chained_load(i32** %p) {
block1:
%z = load i32** %p
br i1 true, label %block2, label %block3
block2:
%a = load i32** %p
br label %block4
block3:
%b = load i32** %p
br label %block4
block4:
%c = load i32** %p
%d = load i32* %c
ret i32 %d
; CHECK: @chained_load
; CHECK: %z = load i32** %p
; CHECK-NOT: load
; CHECK: %d = load i32* %z
; CHECK-NEXT: ret i32 %d
}
Improve GVN to be able to forward substitute a small load from a piece of a large store when both are in the same block. This allows clang to compile the testcase in PR4216 to this code: _test_bitfield: movl 4(%esp), %eax movl %eax, %ecx andl $-65536, %ecx orl $32962, %eax andl $40186, %eax orl %ecx, %eax ret This is not ideal, but is a whole lot better than the code produced by llvm-gcc: _test_bitfield: movw $-32574, %ax orw 4(%esp), %ax andw $-25350, %ax movw %ax, 4(%esp) movw 7(%esp), %cx shlw $8, %cx movzbl 6(%esp), %edx orw %cx, %dx movzwl %dx, %ecx shll $16, %ecx movzwl %ax, %eax orl %ecx, %eax ret and dramatically better than that produced by gcc 4.2: _test_bitfield: pushl %ebx call L3 "L00000000001$pb": L3: popl %ebx movl 8(%esp), %eax leal 0(,%eax,4), %edx sarb $7, %dl movl %eax, %ecx andl $7168, %ecx andl $-7201, %ebx movzbl %dl, %edx andl $1, %edx sall $5, %edx orl %ecx, %ebx orl %edx, %ebx andl $24, %eax andl $-58336, %ebx orl %eax, %ebx orl $32962, %ebx movl %ebx, %eax popl %ebx ret llvm-svn: 82439
2009-09-21 07:57:11 +02:00
Small and carefully crafted testcase showing a miscompilation by GVN that I'm working on. This is manifesting as a miscompile of 255.vortex on some targets. No check lines yet because it fails. llvm-svn: 90520
2009-12-04 03:12:12 +01:00
declare i1 @cond() readonly
declare i1 @cond2() readonly
define i32 @phi_trans2() {
Switch GVN and memdep to use PHITransAddr, which correctly handles phi translation of complex expressions like &A[i+1]. This has the following benefits: 1. The phi translation logic is all contained in its own class with a strong interface and verification that it is self consistent. 2. The logic is more correct than before. Previously, if intermediate expressions got PHI translated, we'd miss the update and scan for the wrong pointers in predecessor blocks. @phi_trans2 is a testcase for this. 3. We have a lot less code in memdep. We can handle phi translation across blocks of things like @phi_trans3, which is pretty insane :). This patch should fix the miscompiles of 255.vortex, and I tested it with a bootstrap of llvm-gcc, llvm-test and dejagnu of course. llvm-svn: 90926
2009-12-09 02:59:31 +01:00
; CHECK: @phi_trans2
Small and carefully crafted testcase showing a miscompilation by GVN that I'm working on. This is manifesting as a miscompile of 255.vortex on some targets. No check lines yet because it fails. llvm-svn: 90520
2009-12-04 03:12:12 +01:00
entry:
%P = alloca i32, i32 400
br label %F1
F1:
%A = phi i32 [1, %entry], [2, %F]
%cond2 = call i1 @cond()
br i1 %cond2, label %T1, label %TY
T1:
%P2 = getelementptr i32* %P, i32 %A
%x = load i32* %P2
%cond = call i1 @cond2()
br i1 %cond, label %TX, label %F
F:
%P3 = getelementptr i32* %P, i32 2
store i32 17, i32* %P3
store i32 42, i32* %P2 ; Provides "P[A]".
br label %F1
TX:
Switch GVN and memdep to use PHITransAddr, which correctly handles phi translation of complex expressions like &A[i+1]. This has the following benefits: 1. The phi translation logic is all contained in its own class with a strong interface and verification that it is self consistent. 2. The logic is more correct than before. Previously, if intermediate expressions got PHI translated, we'd miss the update and scan for the wrong pointers in predecessor blocks. @phi_trans2 is a testcase for this. 3. We have a lot less code in memdep. We can handle phi translation across blocks of things like @phi_trans3, which is pretty insane :). This patch should fix the miscompiles of 255.vortex, and I tested it with a bootstrap of llvm-gcc, llvm-test and dejagnu of course. llvm-svn: 90926
2009-12-09 02:59:31 +01:00
; This load should not be compiled to 'ret i32 42'. An overly clever
; implementation of GVN would see that we're returning 17 if the loop
; executes once or 42 if it executes more than that, but we'd have to do
; loop restructuring to expose this, and GVN shouldn't do this sort of CFG
; transformation.
; CHECK: TX:
; CHECK: ret i32 %x
ret i32 %x
Small and carefully crafted testcase showing a miscompilation by GVN that I'm working on. This is manifesting as a miscompile of 255.vortex on some targets. No check lines yet because it fails. llvm-svn: 90520
2009-12-04 03:12:12 +01:00
TY:
ret i32 0
}
Switch GVN and memdep to use PHITransAddr, which correctly handles phi translation of complex expressions like &A[i+1]. This has the following benefits: 1. The phi translation logic is all contained in its own class with a strong interface and verification that it is self consistent. 2. The logic is more correct than before. Previously, if intermediate expressions got PHI translated, we'd miss the update and scan for the wrong pointers in predecessor blocks. @phi_trans2 is a testcase for this. 3. We have a lot less code in memdep. We can handle phi translation across blocks of things like @phi_trans3, which is pretty insane :). This patch should fix the miscompiles of 255.vortex, and I tested it with a bootstrap of llvm-gcc, llvm-test and dejagnu of course. llvm-svn: 90926
2009-12-09 02:59:31 +01:00
define i32 @phi_trans3(i32* %p) {
; CHECK: @phi_trans3
block1:
br i1 true, label %block2, label %block3
block2:
store i32 87, i32* %p
br label %block4
block3:
%p2 = getelementptr i32* %p, i32 43
store i32 97, i32* %p2
br label %block4
block4:
%A = phi i32 [-1, %block2], [42, %block3]
br i1 true, label %block5, label %exit
; CHECK: block4:
; CHECK-NEXT: %D = phi i32 [ 87, %block2 ], [ 97, %block3 ]
; CHECK-NOT: load
block5:
%B = add i32 %A, 1
br i1 true, label %block6, label %exit
block6:
%C = getelementptr i32* %p, i32 %B
br i1 true, label %block7, label %exit
block7:
%D = load i32* %C
ret i32 %D
; CHECK: block7:
; CHECK-NEXT: ret i32 %D
exit:
ret i32 -1
}
Small and carefully crafted testcase showing a miscompilation by GVN that I'm working on. This is manifesting as a miscompile of 255.vortex on some targets. No check lines yet because it fails. llvm-svn: 90520
2009-12-04 03:12:12 +01:00
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
define i8 @phi_trans4(i8* %p) {
; CHECK: @phi_trans4
entry:
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
%X3 = getelementptr i8* %p, i32 192
store i8 192, i8* %X3
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
%X = getelementptr i8* %p, i32 4
%Y = load i8* %X
br label %loop
loop:
%i = phi i32 [4, %entry], [192, %loop]
%X2 = getelementptr i8* %p, i32 %i
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
%Y2 = load i8* %X2
; CHECK: loop:
; CHECK-NEXT: %Y2 = phi i8 [ %Y, %entry ], [ 0, %loop ]
; CHECK-NOT: load i8
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
%cond = call i1 @cond2()
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
%Z = bitcast i8 *%X3 to i32*
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
store i32 0, i32* %Z
br i1 %cond, label %loop, label %out
out:
%R = add i8 %Y, %Y2
ret i8 %R
}
define i8 @phi_trans5(i8* %p) {
; CHECK: @phi_trans5
entry:
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
%X4 = getelementptr i8* %p, i32 2
store i8 19, i8* %X4
%X = getelementptr i8* %p, i32 4
%Y = load i8* %X
br label %loop
loop:
%i = phi i32 [4, %entry], [3, %cont]
%X2 = getelementptr i8* %p, i32 %i
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
%Y2 = load i8* %X2 ; Ensure this load is not being incorrectly replaced.
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
%cond = call i1 @cond2()
br i1 %cond, label %cont, label %out
cont:
%Z = getelementptr i8* %X2, i32 -1
%Z2 = bitcast i8 *%Z to i32*
store i32 50462976, i32* %Z2 ;; (1 << 8) | (2 << 16) | (3 << 24)
fix hte last remaining known (by me) phi translation bug. When we reanalyze clobbers to forward pieces of large stores to small loads, we need to consider the properly phi translated pointer in the store block. llvm-svn: 90978
2009-12-09 19:21:46 +01:00
; CHECK: store i32
; CHECK-NEXT: getelementptr i8* %p, i32 3
; CHECK-NEXT: load i8*
the code in GVN that tries to forward large loads to small stores is not phi translating, thus it miscompiles really crazy testcases. This is from inspection, I haven't seen this in the wild. llvm-svn: 90930
2009-12-09 03:43:05 +01:00
br label %loop
out:
%R = add i8 %Y, %Y2
ret i8 %R
}
Reference in New Issue Copy Permalink