2016-08-15 21:27:49 +08:00
|
|
|
; RUN: opt %loadPolly -polly-scops -polly-invariant-load-hoisting=true -analyze < %s | FileCheck %s
|
|
|
|
|
; RUN: opt %loadPolly -polly-codegen -polly-invariant-load-hoisting=true -S < %s | FileCheck %s --check-prefix=CODEGEN
|
2015-10-18 20:39:19 +08:00
|
|
|
;
|
|
|
|
|
; void f(int *I0, int *I1, int *V) {
|
|
|
|
|
; for (int i = 0; i < 1000; i++) {
|
|
|
|
|
; if ((long)(*I0) == 0)
|
|
|
|
|
; V[i] += *I1;
|
|
|
|
|
; }
|
|
|
|
|
; }
|
|
|
|
|
;
|
Model zext-extend instructions
A zero-extended value can be interpreted as a piecewise defined signed
value. If the value was non-negative it stays the same, otherwise it
is the sum of the original value and 2^n where n is the bit-width of
the original (or operand) type. Examples:
zext i8 127 to i32 -> { [127] }
zext i8 -1 to i32 -> { [256 + (-1)] } = { [255] }
zext i8 %v to i32 -> [v] -> { [v] | v >= 0; [256 + v] | v < 0 }
However, LLVM/Scalar Evolution uses zero-extend (potentially lead by a
truncate) to represent some forms of modulo computation. The left-hand side
of the condition in the code below would result in the SCEV
"zext i1 <false, +, true>for.body" which is just another description
of the C expression "i & 1 != 0" or, equivalently, "i % 2 != 0".
for (i = 0; i < N; i++)
if (i & 1 != 0 /* == i % 2 */)
/* do something */
If we do not make the modulo explicit but only use the mechanism described
above we will get the very restrictive assumption "N < 3", because for all
values of N >= 3 the SCEVAddRecExpr operand of the zero-extend would wrap.
Alternatively, we can make the modulo in the operand explicit in the
resulting piecewise function and thereby avoid the assumption on N. For the
example this would result in the following piecewise affine function:
{ [i0] -> [(1)] : 2*floor((-1 + i0)/2) = -1 + i0;
[i0] -> [(0)] : 2*floor((i0)/2) = i0 }
To this end we can first determine if the (immediate) operand of the
zero-extend can wrap and, in case it might, we will use explicit modulo
semantic to compute the result instead of emitting non-wrapping assumptions.
Note that operands with large bit-widths are less likely to be negative
because it would result in a very large access offset or loop bound after the
zero-extend. To this end one can optimistically assume the operand to be
positive and avoid the piecewise definition if the bit-width is bigger than
some threshold (here MaxZextSmallBitWidth).
We choose to go with a hybrid solution of all modeling techniques described
above. For small bit-widths (up to MaxZextSmallBitWidth) we will model the
wrapping explicitly and use a piecewise defined function. However, if the
bit-width is bigger than MaxZextSmallBitWidth we will employ overflow
assumptions and assume the "former negative" piece will not exist.
llvm-svn: 267408
2016-04-25 22:01:36 +08:00
|
|
|
; CHECK: Assumed Context:
|
|
|
|
|
; CHECK-NEXT: [loadI0] -> { : }
|
|
|
|
|
; CHECK-NEXT: Invalid Context:
|
|
|
|
|
; CHECK-NEXT: [loadI0] -> { : loadI0 < 0 }
|
2015-10-18 20:39:19 +08:00
|
|
|
;
|
Model zext-extend instructions
A zero-extended value can be interpreted as a piecewise defined signed
value. If the value was non-negative it stays the same, otherwise it
is the sum of the original value and 2^n where n is the bit-width of
the original (or operand) type. Examples:
zext i8 127 to i32 -> { [127] }
zext i8 -1 to i32 -> { [256 + (-1)] } = { [255] }
zext i8 %v to i32 -> [v] -> { [v] | v >= 0; [256 + v] | v < 0 }
However, LLVM/Scalar Evolution uses zero-extend (potentially lead by a
truncate) to represent some forms of modulo computation. The left-hand side
of the condition in the code below would result in the SCEV
"zext i1 <false, +, true>for.body" which is just another description
of the C expression "i & 1 != 0" or, equivalently, "i % 2 != 0".
for (i = 0; i < N; i++)
if (i & 1 != 0 /* == i % 2 */)
/* do something */
If we do not make the modulo explicit but only use the mechanism described
above we will get the very restrictive assumption "N < 3", because for all
values of N >= 3 the SCEVAddRecExpr operand of the zero-extend would wrap.
Alternatively, we can make the modulo in the operand explicit in the
resulting piecewise function and thereby avoid the assumption on N. For the
example this would result in the following piecewise affine function:
{ [i0] -> [(1)] : 2*floor((-1 + i0)/2) = -1 + i0;
[i0] -> [(0)] : 2*floor((i0)/2) = i0 }
To this end we can first determine if the (immediate) operand of the
zero-extend can wrap and, in case it might, we will use explicit modulo
semantic to compute the result instead of emitting non-wrapping assumptions.
Note that operands with large bit-widths are less likely to be negative
because it would result in a very large access offset or loop bound after the
zero-extend. To this end one can optimistically assume the operand to be
positive and avoid the piecewise definition if the bit-width is bigger than
some threshold (here MaxZextSmallBitWidth).
We choose to go with a hybrid solution of all modeling techniques described
above. For small bit-widths (up to MaxZextSmallBitWidth) we will model the
wrapping explicitly and use a piecewise defined function. However, if the
bit-width is bigger than MaxZextSmallBitWidth we will employ overflow
assumptions and assume the "former negative" piece will not exist.
llvm-svn: 267408
2016-04-25 22:01:36 +08:00
|
|
|
; CHECK: p0: %loadI0
|
|
|
|
|
;
|
|
|
|
|
; CHECK: Stmt_if_then
|
|
|
|
|
; CHECK-NEXT: Domain :=
|
|
|
|
|
; CHECK-NEXT: [loadI0] -> { Stmt_if_then[i0] : loadI0 = 0 and 0 <= i0 <= 999 };
|
2015-10-18 20:39:19 +08:00
|
|
|
;
|
|
|
|
|
; CODEGEN: polly.preload.begin:
|
2016-06-12 03:17:15 +08:00
|
|
|
; CODEGEN-NEXT: %polly.access.I0 = getelementptr i32, i32* %I0, i64 0
|
2015-10-18 20:39:19 +08:00
|
|
|
; CODEGEN-NEXT: %polly.access.I0.load = load i32, i32* %polly.access.I0
|
|
|
|
|
; CODEGEN-NEXT: store i32 %polly.access.I0.load, i32* %loadI0.preload.s2a
|
2016-06-12 03:17:15 +08:00
|
|
|
; CODEGEN-NEXT: %0 = sext i32 %polly.access.I0.load to i64
|
|
|
|
|
; CODEGEN-NEXT: %1 = icmp eq i64 %0, 0
|
2015-10-18 20:39:19 +08:00
|
|
|
; CODEGEN-NEXT: br label %polly.preload.cond
|
|
|
|
|
;
|
|
|
|
|
; CODEGEN: polly.preload.cond:
|
2016-06-12 03:17:15 +08:00
|
|
|
; CODEGEN-NEXT: br i1 %1, label %polly.preload.exec, label %polly.preload.merge
|
2015-10-18 20:39:19 +08:00
|
|
|
;
|
|
|
|
|
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
|
|
|
|
|
|
|
|
|
|
define void @f(i32* %I0, i32* %I1, i32* %V) {
|
|
|
|
|
entry:
|
|
|
|
|
br label %for.cond
|
|
|
|
|
|
|
|
|
|
for.cond: ; preds = %for.inc, %entry
|
|
|
|
|
%indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ]
|
|
|
|
|
%exitcond = icmp ne i64 %indvars.iv, 1000
|
|
|
|
|
br i1 %exitcond, label %for.body, label %for.end
|
|
|
|
|
|
|
|
|
|
for.body: ; preds = %for.cond
|
|
|
|
|
%loadI1a = load i32, i32* %I0, align 4
|
|
|
|
|
%arrayidx = getelementptr inbounds i32, i32* %V, i64 %indvars.iv
|
|
|
|
|
%loadI1a1 = load i32, i32* %arrayidx, align 4
|
|
|
|
|
%add = add nsw i32 %loadI1a1, %loadI1a
|
|
|
|
|
store i32 %add, i32* %arrayidx, align 4
|
|
|
|
|
%loadI0 = load i32, i32* %I0, align 4
|
|
|
|
|
%loadI0ext = zext i32 %loadI0 to i64
|
|
|
|
|
%cmp1 = icmp eq i64 %loadI0ext, 0
|
|
|
|
|
br i1 %cmp1, label %if.then, label %if.end
|
|
|
|
|
|
|
|
|
|
if.then: ; preds = %for.body
|
|
|
|
|
%loadI1b = load i32, i32* %I1, align 4
|
|
|
|
|
%arrayidx4 = getelementptr inbounds i32, i32* %V, i64 %indvars.iv
|
|
|
|
|
%loadI1a4 = load i32, i32* %arrayidx4, align 4
|
|
|
|
|
%add5 = add nsw i32 %loadI1a4, %loadI1b
|
|
|
|
|
store i32 %add5, i32* %arrayidx4, align 4
|
|
|
|
|
br label %if.end
|
|
|
|
|
|
|
|
|
|
if.end: ; preds = %if.then, %for.body
|
|
|
|
|
br label %for.inc
|
|
|
|
|
|
|
|
|
|
for.inc: ; preds = %if.end
|
|
|
|
|
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
|
|
|
|
|
br label %for.cond
|
|
|
|
|
|
|
|
|
|
for.end: ; preds = %for.cond
|
|
|
|
|
ret void
|
|
|
|
|
}
|