7 * <https://bugs.libre-soc.org/show_bug.cgi?id=561>
8 * <https://bugs.libre-soc.org/show_bug.cgi?id=572>
9 * <https://bugs.libre-soc.org/show_bug.cgi?id=571>
10 * <https://llvm.org/devmtg/2016-11/Slides/Emerson-ScalableVectorizationinLLVMIR.pdf>
11 * <https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc#vector-loads-and-stores>
13 Vectorisation of Load and Store requires creation, from scalar operations,
14 a number of different types:
16 * fixed stride (contiguous sequence with no gaps)
17 * element strided (sequential but regularly offset, with gaps)
18 * vector indexed (vector of base addresses and vector of offsets)
19 * fail-first on the same (where it makes sense to do so)
21 OpenPOWER Load/Store operations may be seen from [[isa/fixedload]] and
22 [[isa/fixedstore]] pseudocode to be of the form:
28 and for immediate variants:
34 Thus in the first example, the source registers may each be independently
35 marked as scalar or vector, and likewise the destination; in the second
36 example only the one source and one dest may be marked as scalar or
39 Thus we can see that Vector Indexed may be covered, and, as demonstrated
40 with the pseudocode below, the immediate can be set to the element width
41 in order to give unit or element stride. With there being no way to tell which from the Scalar opcode, the choice is provided instead by the SV Context.
44 # op_width: lb=1, lh=2, lw=4, ld=8
45 op_load(RT, RA, op_width, immed, svctx, RAupdate):
46 ps = get_pred_val(FALSE, RA); # predication on src
47 pd = get_pred_val(FALSE, RT); # ... AND on dest
48 for (i=0, j=0, u=0; i < VL && j < VL;):
49 # skip nonpredicates elements
50 if (RA.isvec) while (!(ps & 1<<i)) i++;
51 if (RAupdate.isvec) while (!(ps & 1<<u)) u++;
52 if (RT.isvec) while (!(pd & 1<<j)) j++;
53 if svctx.ldstmode == elementstride:
57 elif svctx.ldstmode == unitstride:
62 # quirky Vector indexed mode but with an immediate
66 # standard scalar mode (but predicated)
67 # no stride multiplier means VSPLAT mode
74 if RAupdate: ireg[RAupdate+u] = EA;
76 ireg[RT+j] <= MEM[EA];
78 break # destination scalar, end now
84 function op_ldx(RT, RA, RB, RAupdate=False) # LD not VLD!
85 ps = get_pred_val(FALSE, RA); # predication on src
86 pd = get_pred_val(FALSE, RT); # ... AND on dest
87 for (i=0, j=0, k=0, u=0; i < VL && j < VL && k < VL):
88 # skip nonpredicated RA, RB and RT
89 if (RA.isvec) while (!(ps & 1<<i)) i++;
90 if (RAupdate.isvec) while (!(ps & 1<<u)) u++;
91 if (RB.isvec) while (!(ps & 1<<k)) k++;
92 if (RT.isvec) while (!(pd & 1<<j)) j++;
93 EA = ireg[RA+i] + ireg[RB+k] # indexed address
94 if RAupdate: ireg[RAupdate+u] = EA
95 ireg[RT+j] <= MEM[EA];
97 break # destination scalar, end immediately
98 if (!RA.isvec && !RB.isvec)
104 Note in both cases that [[sv/svp64]] allows RA in "update" mode (`ldux`) to be effectively a completely different register from RA-as-a-source. This because there is room in svp64 to extend RA-as-src as well as RA-as-dest, both independently as scalar or vector *and* independently extending their range.
106 # Determining the LD/ST Modes
108 A minor complication (caused by the retro-fitting of modern Vector
109 features to a Scalar ISA) is that certain features do not exactly make
110 sense or are considered a security risk. Fail-first on Vector Indexed
111 allows attackers to probe large numbers of pages from userspace, where
112 strided fail-first (by creating contiguous sequential LDs) does not.
114 In addition, even in other modes, Vector source RA makes no sense for
115 computing offsets, and reduce mode even less. Realistically we need
116 an alternative table meaning for [[sv/svp64]] mode.
121 - vector immed(RA) nonsense.
122 - unit-stride/el-stride needed on immed(RA)
124 modes for immed(RA) version:
130 - vector RA is "banned"
132 | 0-1 | 2 | 3 4 | description |
133 | --- | --- |---------|-------------------------- |
134 | 00 | str | sz dz | normal mode |
135 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
136 | 01 | inv | str RC1 | Rc=0: ffirst z/nonz |
137 | 10 | N | sz str | sat mode: N=0/1 u/s |
138 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
139 | 11 | inv | str RC1 | Rc=0: pred-result z/nonz |
141 The `str` bit is only relevant when `RA.isvec` is clear: this indicates
142 whether stride is unit or element:
145 svctx.ldstmode = indexed
147 svctx.ldstmode = unitstride
149 svctx.ldstmode = elementstride
151 Thr modes for RA+RB indexed version are slightly different:
157 - vector RA or RB is "banned"
160 | 0-1 | 2 | 3 4 | description |
161 | --- | --- |---------|-------------------------- |
162 | 00 | 0 | sz dz | normal mode |
163 | 00 | rsv | rsvd | reserved |
164 | 01 | inv | CR-bit | Rc=1: ffirst CR sel |
165 | 01 | inv | sz RC1 | Rc=0: ffirst z/nonz |
166 | 10 | N | sz dz | sat mode: N=0/1 u/s |
167 | 11 | inv | CR-bit | Rc=1: pred-result CR sel |
168 | 11 | inv | sz RC1 | Rc=0: pred-result z/nonz |
170 imm(RA) RT.v RA.v no stride allowed
171 imm(RA) RY.s RA.v no stride allowed
172 imm(RA) RT.v RA.s stride-select needed
173 imm(RA) RT.s RA.s not vectorised
174 RA,RB RT.v RA/RB.v ffirst banned
175 RA,RB RT.s RA/RB.v ffirst banned
176 RA,RB RT.v RA/RB.s vsplat activated
177 RA,RB RT.s RA/RB.s not vectirised
179 # LOAD/STORE Elwidths <a name="ldst"></a>
181 Loads and Stores are almost unique in that the OpenPOWER Scalar ISA
182 provides a width for the operation (lb, lh, lw, ld). Only `extsb` and
183 others like it provide an explicit operation width. There are therefore
184 *three* widths involved:
186 * operation width (lb=8, lh=16, lw=32, ld=64)
187 s src elelent width override
188 * destination element width override
190 Some care is therefore needed to express and make clear the transformations,
191 which are expressly in this order:
193 * Load at the operation width (lb/lh/lw/ld) as usual
194 * byte-reversal as usual
195 * Non-saturated mode:
196 - zero-extension or truncation from operation width to source elwidth
197 - zero/truncation to dest elwidth
199 - Sign-extension or truncation from operation width to source width
200 - signed/unsigned saturation down to dest elwidth
202 In order to respect OpenPOWER v3.0B Scalar behaviour the memory side
203 is treated effectively as completely separate and distinct from SV
204 augmentation. This is primarily down to quirks surrounding LE/BE and
205 byte-reversal in OpenPOWER.
207 Note the following regarding the pseudocode to follow:
209 * `scalar identity behaviour` SV Context parameter conditions turn this
210 into a straight absolute fully-compliant Scalar v3.0B LD operation
211 * `brev` selects whether the operation is the byte-reversed variant (`ldbrx`
213 * `op_width` specifies the operation width (`lb`, `lh`, `lw`, `ld`) as
214 a "normal" part of Scalar v3.0B LD
215 * `imm_offs` specifies the immediate offset `ld r3, imm_offs(r5)`, again
216 as a "normal" part of Scalar v3.0B LD
217 * `svctx` specifies the SV Context and includes VL as well as
218 source and destination elwidth overrides.
220 Below is the pseudocode for Unit-Strided LD (which includes Vector capability).
222 Note that twin predication, predication-zeroing, saturation
223 and other modes have all been removed, for clarity and simplicity:
225 # LD not VLD! (ldbrx if brev=True)
226 # this covers unit stride mode and a type of vector offset
227 function op_ld(RT, RA, brev, op_width, imm_offs, svctx)
228 for (int i = 0, int j = 0; i < svctx.VL && j < svctx.VL;):
230 if not svctx.unit/el-strided:
231 # strange vector mode, compute 64 bit address which is
232 # not polymorphic! elwidth hardcoded to 64 here
233 srcbase = get_polymorphed_reg(RA, 64, i)
235 # unit / element stride mode, compute 64 bit address
236 srcbase = get_polymorphed_reg(RA, 64, 0)
237 # adjust for unit/el-stride
240 # takes care of (merges) processor LE/BE and ld/ldbrx
241 bytereverse = brev XNOR MSR.LE
243 # read the underlying memory
244 memread <= mem[srcbase + imm_offs];
246 # optionally performs byteswap at op width
248 memread = byteswap(memread, op_width)
252 if svpctx.saturation_mode:
253 ... saturation adjustment...
255 # truncate/extend to over-ridden source width.
256 memread = adjust_wid(memread, op_width, svctx.src_elwidth)
258 # takes care of inserting memory-read (now correctly byteswapped)
259 # into regfile underlying LE-defined order, into the right place
260 # within the NEON-like register, respecting destination element
261 # bitwidth, and the element index (j)
262 set_polymorphed_reg(RT, svctx.dest_bitwidth, j, memread)
264 # increments both src and dest element indices (no predication here)
270 In the [[sv/propagation]] page the concept of "Remapping" is described.
271 Whilst it is expensive to set up (2 64-bit opcodes minimum) it provides
272 a way to arbitrarily perform 1D, 2D and 3D "remapping" of up to 64
273 elements worth of LDs or STs. The usual interest in such re-mapping
274 is for example in separating out 24-bit RGB channel data into separate
275 contiguous registers. NEON covers this as shown in the diagram below:
277 <img src="https://community.arm.com/cfs-file/__key/communityserver-blogs-components-weblogfiles/00-00-00-21-42/Loading-RGB-data-with-structured-load.png" >
279 Remap easily covers this capability, and with dest
280 elwidth overrides and saturation may do so with built-in conversion that
281 would normally require additional width-extension, sign-extension and
282 min/max Vectorised instructions as post-processing stages.
284 Thus we do not need to provide specialist LD/ST "Structure Packed" opcodes
285 because the generic abstracted concept of "Remapping", when applied to
286 LD/ST, will give that same capability, with far more flexibility.