mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-23 11:13:28 +01:00
8abf11ea97
Summary: If both vector args to vselect are concat_vectors and the condition is constant and picks half a vector from each argument, convert the vselect into a concat_vectors. Added a test. The ConvertSelectToConcatVector is assuming it doesn't get vselects with arguments of, for example, <undef, undef, true, true>. Those get taken care of in the checks above its call. Reviewers: nadav, delena, grosbach, hfinkel Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D3916 llvm-svn: 209929
195 lines
7.7 KiB
LLVM
195 lines
7.7 KiB
LLVM
; RUN: llc < %s
|
|
|
|
%Domain = type { i8*, i32, i32*, i32, i32, i32*, %Domain* }
|
|
@AConst = constant i32 123 ; <i32*> [#uses=1]
|
|
|
|
; Test setting values of different constants in registers.
|
|
;
|
|
define void @testConsts(i32 %N, float %X) {
|
|
%a = add i32 %N, 1 ; <i32> [#uses=0]
|
|
%i = add i32 %N, 12345678 ; <i32> [#uses=0]
|
|
%b = add i16 4, 3 ; <i16> [#uses=0]
|
|
%c = fadd float %X, 0.000000e+00 ; <float> [#uses=0]
|
|
%d = fadd float %X, 0x400921CAC0000000 ; <float> [#uses=0]
|
|
%f = add i32 -1, 10 ; <i32> [#uses=0]
|
|
%g = add i16 20, -1 ; <i16> [#uses=0]
|
|
%j = add i16 -1, 30 ; <i16> [#uses=0]
|
|
%h = add i8 40, -1 ; <i8> [#uses=0]
|
|
%k = add i8 -1, 50 ; <i8> [#uses=0]
|
|
ret void
|
|
}
|
|
|
|
; A SetCC whose result is used should produce instructions to
|
|
; compute the boolean value in a register. One whose result
|
|
; is unused will only generate the condition code but not
|
|
; the boolean result.
|
|
;
|
|
define void @unusedBool(i32* %x, i32* %y) {
|
|
icmp eq i32* %x, %y ; <i1>:1 [#uses=1]
|
|
xor i1 %1, true ; <i1>:2 [#uses=0]
|
|
icmp ne i32* %x, %y ; <i1>:3 [#uses=0]
|
|
ret void
|
|
}
|
|
|
|
; A constant argument to a Phi produces a Cast instruction in the
|
|
; corresponding predecessor basic block. This checks a few things:
|
|
; -- phi arguments coming from the bottom of the same basic block
|
|
; (they should not be forward substituted in the machine code!)
|
|
; -- code generation for casts of various types
|
|
; -- use of immediate fields for integral constants of different sizes
|
|
; -- branch on a constant condition
|
|
;
|
|
define void @mergeConstants(i32* %x, i32* %y) {
|
|
; <label>:0
|
|
br label %Top
|
|
|
|
Top: ; preds = %Next, %Top, %0
|
|
phi i32 [ 0, %0 ], [ 1, %Top ], [ 524288, %Next ] ; <i32>:1 [#uses=0]
|
|
phi float [ 0.000000e+00, %0 ], [ 1.000000e+00, %Top ], [ 2.000000e+00, %Next ] ; <float>:2 [#uses=0]
|
|
phi double [ 5.000000e-01, %0 ], [ 1.500000e+00, %Top ], [ 2.500000e+00, %Next ]
|
|
phi i1 [ true, %0 ], [ false, %Top ], [ true, %Next ] ; <i1>:4 [#uses=0]
|
|
br i1 true, label %Top, label %Next
|
|
|
|
Next: ; preds = %Top
|
|
br label %Top
|
|
}
|
|
|
|
|
|
|
|
; A constant argument to a cast used only once should be forward substituted
|
|
; and loaded where needed, which happens is:
|
|
; -- User of cast has no immediate field
|
|
; -- User of cast has immediate field but constant is too large to fit
|
|
; or constant is not resolved until later (e.g., global address)
|
|
; -- User of cast uses it as a call arg. or return value so it is an implicit
|
|
; use but has to be loaded into a virtual register so that the reg.
|
|
; allocator can allocate the appropriate phys. reg. for it
|
|
;
|
|
define i32* @castconst(float) {
|
|
%castbig = trunc i64 99999999 to i32 ; <i32> [#uses=1]
|
|
%castsmall = trunc i64 1 to i32 ; <i32> [#uses=1]
|
|
%usebig = add i32 %castbig, %castsmall ; <i32> [#uses=0]
|
|
%castglob = bitcast i32* @AConst to i64* ; <i64*> [#uses=1]
|
|
%dummyl = load i64* %castglob ; <i64> [#uses=0]
|
|
%castnull = inttoptr i64 0 to i32* ; <i32*> [#uses=1]
|
|
ret i32* %castnull
|
|
}
|
|
|
|
; Test branch-on-comparison-with-zero, in two ways:
|
|
; 1. can be folded
|
|
; 2. cannot be folded because result of comparison is used twice
|
|
;
|
|
define void @testbool(i32 %A, i32 %B) {
|
|
br label %Top
|
|
|
|
Top: ; preds = %loop, %0
|
|
%D = add i32 %A, %B ; <i32> [#uses=2]
|
|
%E = sub i32 %D, -4 ; <i32> [#uses=1]
|
|
%C = icmp sle i32 %E, 0 ; <i1> [#uses=1]
|
|
br i1 %C, label %retlbl, label %loop
|
|
|
|
loop: ; preds = %loop, %Top
|
|
%F = add i32 %A, %B ; <i32> [#uses=0]
|
|
%G = sub i32 %D, -4 ; <i32> [#uses=1]
|
|
%D.upgrd.1 = icmp sle i32 %G, 0 ; <i1> [#uses=1]
|
|
%E.upgrd.2 = xor i1 %D.upgrd.1, true ; <i1> [#uses=1]
|
|
br i1 %E.upgrd.2, label %loop, label %Top
|
|
|
|
retlbl: ; preds = %Top
|
|
ret void
|
|
}
|
|
|
|
|
|
;; Test use of a boolean result in cast operations.
|
|
;; Requires converting a condition code result into a 0/1 value in a reg.
|
|
;;
|
|
define i32 @castbool(i32 %A, i32 %B) {
|
|
bb0:
|
|
%cond213 = icmp slt i32 %A, %B ; <i1> [#uses=1]
|
|
%cast110 = zext i1 %cond213 to i8 ; <i8> [#uses=1]
|
|
%cast109 = zext i8 %cast110 to i32 ; <i32> [#uses=1]
|
|
ret i32 %cast109
|
|
}
|
|
|
|
;; Test use of a boolean result in arithmetic and logical operations.
|
|
;; Requires converting a condition code result into a 0/1 value in a reg.
|
|
;;
|
|
define i1 @boolexpr(i1 %b, i32 %N) {
|
|
%b2 = icmp sge i32 %N, 0 ; <i1> [#uses=1]
|
|
%b3 = and i1 %b, %b2 ; <i1> [#uses=1]
|
|
ret i1 %b3
|
|
}
|
|
|
|
; Test branch on floating point comparison
|
|
;
|
|
define void @testfloatbool(float %x, float %y) {
|
|
br label %Top
|
|
|
|
Top: ; preds = %Top, %0
|
|
%p = fadd float %x, %y ; <float> [#uses=1]
|
|
%z = fsub float %x, %y ; <float> [#uses=1]
|
|
%b = fcmp ole float %p, %z ; <i1> [#uses=2]
|
|
%c = xor i1 %b, true ; <i1> [#uses=0]
|
|
br i1 %b, label %Top, label %goon
|
|
|
|
goon: ; preds = %Top
|
|
ret void
|
|
}
|
|
|
|
|
|
; Test cases where an LLVM instruction requires no machine
|
|
; instructions (e.g., cast int* to long). But there are 2 cases:
|
|
; 1. If the result register has only a single use and the use is in the
|
|
; same basic block, the operand will be copy-propagated during
|
|
; instruction selection.
|
|
; 2. If the result register has multiple uses or is in a different
|
|
; basic block, it cannot (or will not) be copy propagated during
|
|
; instruction selection. It will generate a
|
|
; copy instruction (add-with-0), but this copy should get coalesced
|
|
; away by the register allocator.
|
|
;
|
|
define i32 @checkForward(i32 %N, i32* %A) {
|
|
bb2:
|
|
%reg114 = shl i32 %N, 2 ; <i32> [#uses=1]
|
|
%cast115 = sext i32 %reg114 to i64 ; <i64> [#uses=1]
|
|
%cast116 = ptrtoint i32* %A to i64 ; <i64> [#uses=1]
|
|
%reg116 = add i64 %cast116, %cast115 ; <i64> [#uses=1]
|
|
%castPtr = inttoptr i64 %reg116 to i32* ; <i32*> [#uses=1]
|
|
%reg118 = load i32* %castPtr ; <i32> [#uses=1]
|
|
%cast117 = sext i32 %reg118 to i64 ; <i64> [#uses=2]
|
|
%reg159 = add i64 1234567, %cast117 ; <i64> [#uses=0]
|
|
%reg160 = add i64 7654321, %cast117 ; <i64> [#uses=0]
|
|
ret i32 0
|
|
}
|
|
|
|
|
|
; Test case for unary NOT operation constructed from XOR.
|
|
;
|
|
define void @checkNot(i1 %b, i32 %i) {
|
|
%notB = xor i1 %b, true ; <i1> [#uses=1]
|
|
%notI = xor i32 %i, -1 ; <i32> [#uses=2]
|
|
%F = icmp sge i32 %notI, 100 ; <i1> [#uses=1]
|
|
%J = add i32 %i, %i ; <i32> [#uses=1]
|
|
%andNotB = and i1 %F, %notB ; <i1> [#uses=0]
|
|
%andNotI = and i32 %J, %notI ; <i32> [#uses=0]
|
|
%notB2 = xor i1 true, %b ; <i1> [#uses=0]
|
|
%notI2 = xor i32 -1, %i ; <i32> [#uses=0]
|
|
ret void
|
|
}
|
|
|
|
; Test case for folding getelementptr into a load/store
|
|
;
|
|
define i32 @checkFoldGEP(%Domain* %D, i64 %idx) {
|
|
%reg841 = getelementptr %Domain* %D, i64 0, i32 1 ; <i32*> [#uses=1]
|
|
%reg820 = load i32* %reg841 ; <i32> [#uses=1]
|
|
ret i32 %reg820
|
|
}
|
|
|
|
; Test case for scalarising a 1 element vselect
|
|
;
|
|
define <1 x i32> @checkScalariseVSELECT(<1 x i32> %a, <1 x i32> %b) {
|
|
%cond = icmp uge <1 x i32> %a, %b
|
|
%s = select <1 x i1> %cond, <1 x i32> %a, <1 x i32> %b
|
|
ret <1 x i32> %s
|
|
}
|