Regalloc solver: check that a variable doesn't exist to test if it can be added (fixes #1123);

This situation could be triggered that can_add_var would return true
while a variable was already added for the given register.

For instance, when we have a reassignment (because of a fixed register
input requirement) and a fixed input conflict on the same fixed
register, this register will not be available in the regs_in set after
inputs_done (because of the fixed input conflict diversion) but will
have its own variable.

cranelift/codegen/src/regalloc/solver.rs

@@ -626,7 +626,7 @@ impl Solver {
     /// constraints if the value has already been added as a variable or fixed assignment.
     fn add_live_var(&mut self, value: Value, rc: RegClass, from: RegUnit, live_through: bool) {
         // Check for existing entries for this value.
-        if self.regs_in.is_avail(rc, from) {
+        if !self.can_add_var(rc, from) {
             // There could be an existing variable entry.
             if let Some(v) = self.vars.iter_mut().find(|v| v.value == value) {
                 // We have an existing variable entry for `value`. Combine the constraints.
@@ -969,6 +969,7 @@ impl Solver {
     /// Check if `value` can be added as a variable to help find a solution.
     pub fn can_add_var(&mut self, constraint: RegClass, from: RegUnit) -> bool {
         !self.regs_in.is_avail(constraint, from)
+            && !self.vars.iter().any(|var| var.from == Some(from))
     }
 }
 

cranelift/filetests/filetests/regalloc/solver-fixedconflict-var-2.clif

@@ -0,0 +1,100 @@
+test compile
+set opt_level=speed
+set enable_pinned_reg=true
+target x86_64 haswell
+
+function u0:0(i32, i32, i32, i64 vmctx) -> i64 uext system_v {
+ebb0(v0: i32, v1: i32, v2: i32, v3: i64):
+    v236 = iconst.i32 0x4de9_bd37
+    v424 = iconst.i32 0
+    jump ebb37(v424)
+
+ebb37(v65: i32):
+    v433 = iconst.i32 0
+    jump ebb40(v433)
+
+ebb40(v70: i32):
+    v75 = iconst.i32 0
+    v259 = iconst.i32 0
+    v78 -> v259
+    v449 = iconst.i32 0
+    v450, v451 = x86_sdivmodx v75, v449, v259
+    v79 -> v450
+    v269 = iconst.i32 0
+    v270 = ushr_imm v269, 31
+    v271 = iadd v269, v270
+    v98 -> v271
+    v100 = iconst.i32 -31
+    v272 = iconst.i32 0x4de9_bd37
+    v490, v273 = x86_smulx v100, v272
+    v493 = iconst.i32 0
+    jump ebb61(v493)
+
+ebb61(v103: i32):
+    v104 = iconst.i32 -23
+    v105 = iconst.i32 -23
+    v106 = popcnt v105
+    v500 = sshr_imm v104, 31
+    v501 = iconst.i32 0
+    jump ebb64(v501)
+
+ebb64(v107: i32):
+    v108 = iconst.i32 0
+    v109 = iconst.i32 0
+    v278 = iconst.i32 0
+    v507, v279 = x86_smulx v109, v278
+    v280 = isub v279, v109
+    v281 = sshr_imm v280, 11
+    v282 = iconst.i32 0
+    v283 = iadd v281, v282
+    v111 -> v283
+    v112 = rotr v108, v283
+    jump ebb65
+
+ebb65:
+    v509 = iconst.i32 0
+    v510, v511 = x86_sdivmodx v107, v509, v112
+    v113 -> v510
+    v114 = iconst.i32 0
+    v517 = iconst.i32 0
+    v518, v519 = x86_sdivmodx v103, v517, v114
+    v115 -> v518
+    v534 = iconst.i32 0
+    v122 -> v534
+    v541 = iconst.i32 0
+    v542, v543 = x86_sdivmodx v271, v541, v122
+    v123 -> v542
+    v289 = iconst.i32 0
+    v125 -> v289
+    v550 = iconst.i32 0
+    v551, v552 = x86_sdivmodx v79, v550, v289
+    v126 -> v551
+    v130 = iconst.i32 0
+    v558 = iconst.i32 0
+    v559, v560 = x86_sdivmodx v70, v558, v130
+    v131 -> v559
+    v305 = iconst.i32 0
+    v140 -> v305
+    v577 = iconst.i32 0
+    v578, v579 = x86_sdivmodx v65, v577, v305
+    v141 -> v578
+    v166 = iconst.i32 0
+    v167 = iconst.i32 -31
+    v318 = iconst.i32 0x4de9_bd37
+    v650, v319 = x86_smulx v167, v318
+    v320 = isub v319, v167
+    v321 = sshr_imm v320, 4
+    v322 = iconst.i32 0
+    v323 = iadd v321, v322
+    v169 -> v323
+    v652 = iconst.i32 0
+    v653, v654 = x86_sdivmodx v166, v652, v323
+    v170 -> v653
+    v171 = iconst.i32 -23
+    v172 = iconst.i32 -23
+    v173 = popcnt v172
+    v174 = popcnt v173
+    v660 = sshr_imm v171, 31
+    v661, v662 = x86_sdivmodx v171, v660, v174
+    trap user0
+}

cranelift/filetests/filetests/regalloc/solver-fixedconflict-var-3.clif

@@ -0,0 +1,137 @@
+test compile
+set opt_level=speed
+set enable_pinned_reg=true
+target x86_64 haswell
+
+function u0:0(i32, i32, i32, i64 vmctx) -> i64 uext system_v {
+ebb0(v0: i32, v1: i32, v2: i32, v3: i64):
+    v5 = iconst.i32 -8
+    v114 = iconst.i32 0
+    v16 = iconst.i32 -8
+    v17 = popcnt v16
+    v192 = ifcmp_imm v17, -1
+    brif eq v192, ebb12
+    trap user0
+
+ebb12:
+    v122 = iconst.i32 0
+    v123 = ushr_imm v122, 31
+    v124 = iadd v122, v123
+    v20 -> v124
+    v25 = iconst.i32 -19
+    v204 = iconst.i32 0
+    v31 -> v204
+    v210 = ifcmp_imm v31, -1
+    brif eq v210, ebb18
+    trap user0
+
+ebb18:
+    v215 = iconst.i32 0
+    jump ebb19(v215)
+
+ebb19(v32: i32):
+    v35 = iconst.i32 0
+    v218 = ifcmp_imm v35, -1
+    brif eq v218, ebb21
+    trap user0
+
+ebb21:
+    v223 = iconst.i32 0
+    jump ebb22(v223)
+
+ebb22(v36: i32):
+    v136 = iconst.i32 0
+    v40 -> v136
+    v227 = ifcmp_imm v136, -1
+    brif eq v227, ebb24
+    trap user0
+
+ebb24:
+    v232 = iconst.i32 0
+    jump ebb25(v232)
+
+ebb25(v41: i32):
+    v142 = iconst.i32 0
+    v45 -> v142
+    v236 = ifcmp_imm v142, -1
+    brif eq v236, ebb27
+    trap user0
+
+ebb27:
+    v241 = iconst.i32 0
+    jump ebb28(v241)
+
+ebb28(v46: i32):
+    v49 = iconst.i32 0
+    v244 = ifcmp_imm v49, -1
+    brif eq v244, ebb30
+    trap user0
+
+ebb30:
+    v254 = iconst.i32 0
+    v53 -> v254
+    v54 = iconst.i32 -23
+    v55 = popcnt v54
+    v143 = iconst.i32 0x4de9_bd37
+    v260, v144 = x86_smulx v55, v143
+    v145 = iconst.i32 0
+    v146 = sshr_imm v145, 4
+    v147 = iconst.i32 0
+    v148 = iadd v146, v147
+    v57 -> v148
+    v58 = ishl v53, v148
+    jump ebb35
+
+ebb35:
+    v262 = iconst.i32 0
+    v263, v264 = x86_sdivmodx v46, v262, v58
+    v59 -> v263
+    v270 = iconst.i32 0
+    v271, v272 = x86_sdivmodx v41, v270, v59
+    v60 -> v271
+    v61 = f32const 0.0
+    v280 = iconst.i32 0
+    v281 = fcmp uno v61, v61
+    brnz v281, ebb41(v280)
+    trap user0
+
+ebb41(v62: i32):
+    v157 = iconst.i32 0
+    v158 = sshr_imm v157, 4
+    v159 = iconst.i32 0
+    v160 = iadd v158, v159
+    v75 -> v160
+    v308 = ifcmp_imm v160, -1
+    brif eq v308, ebb52
+    trap user0
+
+ebb52:
+    v87 = iconst.i32 -23
+    v88 = iconst.i32 -23
+    v89 = popcnt v88
+    v161 = iconst.i32 0x4de9_bd37
+    v324, v162 = x86_smulx v89, v161
+    v163 = isub v162, v89
+    v164 = sshr_imm v163, 4
+    v165 = iconst.i32 0
+    v166 = iadd v164, v165
+    v91 -> v166
+    v326 = iconst.i32 0
+    v327, v328 = x86_sdivmodx v87, v326, v166
+    v92 -> v327
+    v351 = iconst.i32 0
+    v99 -> v351
+    v358 = iconst.i32 0
+    v359, v360 = x86_sdivmodx v36, v358, v99
+    v100 -> v359
+    v102 = iconst.i32 0
+    v103 = rotr.i32 v32, v102
+    v366 = iconst.i32 0
+    v367, v368 = x86_sdivmodx v25, v366, v103
+    v104 -> v367
+    v383 = iconst.i32 0
+    v107 -> v383
+    v390 = iconst.i32 0
+    v391, v392 = x86_sdivmodx v124, v390, v107
+    trap user0
+}

cranelift/filetests/filetests/regalloc/solver-fixedconflict-var.clif

@@ -0,0 +1,173 @@
+test compile
+set opt_level=speed
+set enable_pinned_reg=true
+target x86_64 haswell
+
+;; Test for the issue #1123; https://github.com/CraneStation/cranelift/issues/1123
+
+function u0:0(i32, i32, i32, i64 vmctx) -> i64 uext system_v {
+ebb0(v0: i32, v1: i32, v2: i32, v3: i64):
+    v351 = iconst.i32 0x4de9_bd37
+    v31 = iconst.i32 -23
+    v35 = iconst.i32 0
+    v36 = iconst.i32 -31
+    v357 = iconst.i32 0x4de9_bd37
+    v530, v358 = x86_smulx v36, v357
+    v359 = isub v358, v36
+    v360 = sshr_imm v359, 4
+    v361 = iconst.i32 0
+    v362 = iadd v360, v361
+    v38 -> v362
+    v532 = sshr_imm v35, 31
+    v533, v534 = x86_sdivmodx v35, v532, v362
+    v39 -> v533
+    v53 = iconst.i32 0
+    v547 = ifcmp_imm v53, -1
+    brif eq v547, ebb30
+    trap user0
+
+ebb30:
+    v75 = iconst.i32 0
+    v581 = ifcmp_imm v75, -1
+    brif eq v581, ebb42
+    trap user0
+
+ebb42:
+    v136 = iconst.i32 0
+    v691 = ifcmp_imm v136, -1
+    brif eq v691, ebb81
+    trap user0
+
+ebb81:
+    v158 = iconst.i32 0
+    v725 = ifcmp_imm v158, -1
+    brif eq v725, ebb93
+    trap user0
+
+ebb93:
+    v760 = iconst.i32 0
+    jump ebb106(v760)
+
+ebb106(v175: i32):
+    v179 = iconst.i32 0
+    v180 = icmp_imm eq v179, 0
+    v183 = iconst.i32 0
+    v766 = ifcmp_imm v183, -1
+    brif eq v766, ebb108
+    trap user0
+
+ebb108:
+    v771 = iconst.i32 0
+    jump ebb109(v771)
+
+ebb109(v184: i32):
+    v785 = iconst.i32 0
+    v193 -> v785
+    v791 = ifcmp_imm v193, -1
+    brif eq v791, ebb117
+    trap user0
+
+ebb117:
+    v796 = iconst.i32 0
+    jump ebb118(v796)
+
+ebb118(v194: i32):
+    v203 = iconst.i32 -63
+    v809 = iconst.i32 0
+    v207 -> v809
+    v815 = ifcmp_imm v207, -1
+    brif eq v815, ebb126
+    trap user0
+
+ebb126:
+    v209 = iconst.i32 0
+    v823 = ifcmp_imm v209, -1
+    brif eq v823, ebb129
+    trap user0
+
+ebb129:
+    v213 = iconst.i32 -23
+    v214 = iconst.i32 -19
+    v215 = icmp_imm eq v214, 0
+    v216 = bint.i32 v215
+    v217 = popcnt v216
+    v435 = iconst.i32 0x7df7_df7d
+    v831, v436 = x86_smulx v217, v435
+    v437 = isub v436, v217
+    v438 = sshr_imm v437, 5
+    v439 = ushr_imm v438, 31
+    v440 = iadd v438, v439
+    v219 -> v440
+    v220 = rotr v213, v440
+    v229 = iconst.i32 0
+    v841 = iconst.i32 0
+    v842, v843 = x86_sdivmodx v194, v841, v229
+    v230 -> v842
+    v849 = iconst.i32 0
+    v850, v851 = x86_sdivmodx v184, v849, v230
+    v231 -> v850
+    v232 = iconst.i32 0
+    v857 = iconst.i32 0
+    v858, v859 = x86_sdivmodx v175, v857, v232
+    v233 -> v858
+    v915 = iconst.i32 0
+    jump ebb163(v915)
+
+ebb163(v253: i32):
+    v255 = iconst.i32 0
+    v256 = iconst.i32 -23
+    v257 = iconst.i32 -19
+    v258 = icmp_imm eq v257, 0
+    v259 = bint.i32 v258
+    v260 = popcnt v259
+    v447 = iconst.i32 0x7df7_df7d
+    v921, v448 = x86_smulx v260, v447
+    v449 = isub v448, v260
+    v450 = sshr_imm v449, 5
+    v451 = ushr_imm v450, 31
+    v452 = iadd v450, v451
+    v262 -> v452
+    v263 = rotr v256, v452
+    v264 = popcnt v263
+    v265 = popcnt v264
+    v266 = popcnt v265
+    v267 = rotr v255, v266
+    v268 = popcnt v267
+    v923 = iconst.i32 0
+    v924, v925 = x86_sdivmodx v253, v923, v268
+    v269 -> v924
+    v276 = iconst.i32 0
+    v277 = iconst.i32 -63
+    v278 = popcnt v277
+    v947 = iconst.i32 0
+    v948, v949 = x86_sdivmodx v276, v947, v278
+    v279 -> v948
+    v309 = iconst.i32 0
+    v310 = iconst.i32 0
+    v311 = iconst.i32 0
+    v312 = icmp_imm eq v311, 0
+    v313 = bint.i32 v312
+    v314 = rotr v310, v313
+    v315 = iconst.i32 -31
+    v464 = iconst.i32 0
+    v1020, v465 = x86_smulx v315, v464
+    v466 = isub v465, v315
+    v467 = sshr_imm v466, 4
+    v468 = iconst.i32 0
+    v469 = iadd v467, v468
+    v317 -> v469
+    v1022 = iconst.i32 0
+    v1023, v1024 = x86_sdivmodx v314, v1022, v469
+    v318 -> v1023
+    v320 = iconst.i32 0
+    v321 = iconst.i32 -19
+    v322 = popcnt v321
+    v1030 = iconst.i32 0
+    v1031, v1032 = x86_sdivmodx v320, v1030, v322
+    v323 -> v1031
+    v1047 = iconst.i32 0
+    v325 -> v1047
+    v1054 = sshr_imm v309, 31
+    v1055, v1056 = x86_sdivmodx v309, v1054, v325
+    trap user0
+}