Fix STG-> C-- of non-native switch with twos compliment

This is actually just the test. Have not gotten to the bottom of it yet. Previous commits are !4717 (merged) which is good to go.

mentioned in merge request !4717 (merged)

changed the description

The following switch:

byte_switch (bits8 x) {
    switch (x) {
    case 255: { return (7); }
    case 254: { return (15); }
    case   0: { return (5); }
    case   3: { return (9); }
    default:  { return (x); }
    }
}

returns the following values when called with values in [-4..4]:

❯ ./ByteSwitch 
-4
-3
15
-1
5
1
2
9
4

I.e. (-1) isn't correctly handled.

Cmm and x86-64 asm for the switch:

[byte_switch() { //  [R1]
         { info_tbls: []
           stack_info: arg_space: 8
         }
     {offset
       c7: // global
           _c1::I8 = %MO_XX_Conv_W64_W8(R1);
           //tick src<ByteSwitch_cmm.cmm:(3,23)-(11,1)>
           if (_c1::I8 < 255 :: W8) goto u8; else goto ub;
       u8: // global
           if (_c1::I8 < 254 :: W8) goto u9; else goto c4;
       u9: // global
           if (_c1::I8 < 1 :: W8) goto c5; else goto ua;
       c5: // global
           //tick src<ByteSwitch_cmm.cmm:7:13-27>
           R1 = 5;
           call (P64[Sp])(R1) args: 8, res: 0, upd: 8;
       ua: // global
           if (_c1::I8 != 3 :: W8) goto c2; else goto c6;
       c6: // global
           //tick src<ByteSwitch_cmm.cmm:8:13-27>
           R1 = 9;
           call (P64[Sp])(R1) args: 8, res: 0, upd: 8;
       c4: // global
           //tick src<ByteSwitch_cmm.cmm:6:15-30>
           R1 = 15;
           call (P64[Sp])(R1) args: 8, res: 0, upd: 8;
       ub: // global
           if (_c1::I8 < 256 :: W8) goto c3; else goto c2;
       c3: // global
           //tick src<ByteSwitch_cmm.cmm:5:15-29>
           R1 = 7;
           call (P64[Sp])(R1) args: 8, res: 0, upd: 8;
       c2: // global
           //tick src<ByteSwitch_cmm.cmm:9:14-28>
           R1 = %MO_XX_Conv_W8_W64(_c1::I8);
           call (P64[Sp])(R1) args: 8, res: 0, upd: 8;
     }
 }]

.section .text
.align 8
.globl byte_switch
.type byte_switch, @function
byte_switch:
_blk_c7:
	cmpb $-1,%bl
	jae _blk_ub
_blk_u8:
	cmpb $-2,%bl
	jb _blk_u9
_blk_c4:
	movl $15,%ebx
	jmp *(%rbp)
_blk_ub:
	cmpb $0,%bl
	jae _blk_c2
_blk_c3:
	movl $7,%ebx
	jmp *(%rbp)
_blk_u9:
	cmpb $1,%bl
	jae _blk_ua
_blk_c5:
	movl $5,%ebx
	jmp *(%rbp)
_blk_ua:
	cmpb $3,%bl
	jne _blk_c2
_blk_c6:
	movl $9,%ebx
	jmp *(%rbp)
_blk_c2:
	jmp *(%rbp)
	.size byte_switch, .-byte_switch

Block ub is wrong because 256 becomes 0 when it is converted to 8-bits. Tracing createSwitchPlan:

createSwitchPlan
  -- (range, m)
  ((0,18446744073709551615),fromList [(0,L7133701809754865669),(3,L7133701809754865670),(254,L7133701809754865668),(255,L7133701809754865667)])
  -- pieces
  [fromList [(0,L7133701809754865669)],fromList [(3,L7133701809754865670)],fromList [(254,L7133701809754865668)],fromList [(255,L7133701809754865667)]]
  -- flatPlan
  (Unconditionally L7133701809754865669,[(1,IfEqual 3 L7133701809754865670 (Unconditionally L7133701809754865666)),(254,Unconditionally L7133701809754865668),(255,Unconditionally L7133701809754865667),(256,Unconditionally L7133701809754865666)])
  -- plan
  IfLT False 255 (IfLT False 254 (IfLT False 1 (Unconditionally L7133701809754865669) (IfEqual 3 L7133701809754865670 (Unconditionally L7133701809754865666))) (Unconditionally L7133701809754865668)) (IfLT False 256 (Unconditionally L7133701809754865667) (Unconditionally L7133701809754865666))

The range is wrong. It seems like there are two issues:

• the test doesn't specify the bounds correctly in the switch
• as it is a reproducer of the original issue, the original code probably didn't specify the bounds correctly

I'll update the branch

resolved all threads

added 2 commits

• 37bef926 - Fix test
• fc0085b7 - StgToCmm: use correct bounds for switches on sized values

Compare with previous version

As always, thanks so much, @hsyl20!!

added 160 commits

• fc0085b7...acb188e0 - 158 commits from branch master
• b784a51e - Test non-native switch C-- with twos compliment
• 5798357d - StgToCmm: use correct bounds for switches on sized values

Compare with previous version

assigned to @marge-bot

marked this merge request as ready

changed title from Draft: Fix non-native switch C-- with twos compliment to Fix STG-> C-- of non-native switch with twos compliment

I will attempt to batch this MR (!6371 (closed))...

merged

Merged in 5798357d