s390x: Fix instruction encoding and disassembly format bugs (#5786)

- Fix encoding of the AHY instruction.
- Fix disassembly format of FIEBR, FIDBR, and LEDBRA instructions.

cranelift/codegen/src/isa/s390x/inst/emit.rs

@@ -1568,7 +1568,7 @@ impl Inst {
 
                 let (opcode_rx, opcode_rxy) = match alu_op {
                     ALUOp::Add32 => (Some(0x5a), Some(0xe35a)),        // A(Y)
-                    ALUOp::Add32Ext16 => (Some(0x4a), Some(0xe34a)),   // AH(Y)
+                    ALUOp::Add32Ext16 => (Some(0x4a), Some(0xe37a)),   // AH(Y)
                     ALUOp::Add64 => (None, Some(0xe308)),              // AG
                     ALUOp::Add64Ext16 => (None, Some(0xe338)),         // AGH
                     ALUOp::Add64Ext32 => (None, Some(0xe318)),         // AGF

cranelift/codegen/src/isa/s390x/inst/emit_tests.rs

@@ -587,7 +587,7 @@ fn test_s390x_binemit() {
                 flags: MemFlags::trusted(),
             },
         },
-        "E3102000004A",
+        "E3102000007A",
         "ahy %r1, 0(%r2)",
     ));
     insns.push((
@@ -7999,7 +7999,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B344008C",
-        "ledbra %f8, %f12, 0",
+        "ledbra %f8, 0, %f12, 0",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8029,7 +8029,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B357708C",
-        "fiebr %f8, %f12, 7",
+        "fiebr %f8, 7, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8039,7 +8039,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B35F708C",
-        "fidbr %f8, %f12, 7",
+        "fidbr %f8, 7, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8049,7 +8049,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B357608C",
-        "fiebr %f8, %f12, 6",
+        "fiebr %f8, 6, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8059,7 +8059,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B35F608C",
-        "fidbr %f8, %f12, 6",
+        "fidbr %f8, 6, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8069,7 +8069,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B357508C",
-        "fiebr %f8, %f12, 5",
+        "fiebr %f8, 5, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8079,7 +8079,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B35F508C",
-        "fidbr %f8, %f12, 5",
+        "fidbr %f8, 5, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8089,7 +8089,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B357408C",
-        "fiebr %f8, %f12, 4",
+        "fiebr %f8, 4, %f12",
     ));
     insns.push((
         Inst::FpuRound {
@@ -8099,7 +8099,7 @@ fn test_s390x_binemit() {
             rn: vr(12),
         },
         "B35F408C",
-        "fidbr %f8, %f12, 4",
+        "fidbr %f8, 4, %f12",
     ));
     insns.push((
         Inst::FpuRound {

cranelift/codegen/src/isa/s390x/inst/mod.rs

@@ -2359,11 +2359,16 @@ impl Inst {
                 let (rn, rn_fpr) = pretty_print_fpr(rn, allocs);
                 if opcode_fpr.is_some() && rd_fpr.is_some() && rn_fpr.is_some() {
                     format!(
-                        "{} {}, {}, {}",
+                        "{} {}, {}, {}{}",
                         opcode_fpr.unwrap(),
                         rd_fpr.unwrap(),
+                        mode,
                         rn_fpr.unwrap(),
-                        mode
+                        if opcode_fpr.unwrap().ends_with('a') {
+                            ", 0"
+                        } else {
+                            ""
+                        }
                     )
                 } else if opcode.starts_with('w') {
                     format!(

cranelift/filetests/filetests/isa/s390x/floating-point-arch13.clif

@@ -386,7 +386,7 @@ block0(v0: i64):
 ; block0:
 ;   ldgr %f2, %r2
 ;   wcdlgb %f4, %f2, 0, 3
-;   ledbra %f0, %f4, 4
+;   ledbra %f0, 4, %f4, 0
 ;   br %r14
 
 function %fcvt_from_sint_i64_f32(i64) -> f32 {
@@ -398,7 +398,7 @@ block0(v0: i64):
 ; block0:
 ;   ldgr %f2, %r2
 ;   wcdgb %f4, %f2, 0, 3
-;   ledbra %f0, %f4, 4
+;   ledbra %f0, 4, %f4, 0
 ;   br %r14
 
 function %fcvt_from_uint_i8_f64(i8) -> f64 {

cranelift/filetests/filetests/isa/s390x/floating-point.clif

@@ -285,7 +285,7 @@ block0(v0: f64):
 }
 
 ; block0:
-;   ledbra %f0, %f0, 0
+;   ledbra %f0, 0, %f0, 0
 ;   br %r14
 
 function %ceil_f32(f32) -> f32 {
@@ -295,7 +295,7 @@ block0(v0: f32):
 }
 
 ; block0:
-;   fiebr %f0, %f0, 6
+;   fiebr %f0, 6, %f0
 ;   br %r14
 
 function %ceil_f64(f64) -> f64 {
@@ -305,7 +305,7 @@ block0(v0: f64):
 }
 
 ; block0:
-;   fidbr %f0, %f0, 6
+;   fidbr %f0, 6, %f0
 ;   br %r14
 
 function %floor_f32(f32) -> f32 {
@@ -315,7 +315,7 @@ block0(v0: f32):
 }
 
 ; block0:
-;   fiebr %f0, %f0, 7
+;   fiebr %f0, 7, %f0
 ;   br %r14
 
 function %floor_f64(f64) -> f64 {
@@ -325,7 +325,7 @@ block0(v0: f64):
 }
 
 ; block0:
-;   fidbr %f0, %f0, 7
+;   fidbr %f0, 7, %f0
 ;   br %r14
 
 function %trunc_f32(f32) -> f32 {
@@ -335,7 +335,7 @@ block0(v0: f32):
 }
 
 ; block0:
-;   fiebr %f0, %f0, 5
+;   fiebr %f0, 5, %f0
 ;   br %r14
 
 function %trunc_f64(f64) -> f64 {
@@ -345,7 +345,7 @@ block0(v0: f64):
 }
 
 ; block0:
-;   fidbr %f0, %f0, 5
+;   fidbr %f0, 5, %f0
 ;   br %r14
 
 function %nearest_f32(f32) -> f32 {
@@ -355,7 +355,7 @@ block0(v0: f32):
 }
 
 ; block0:
-;   fiebr %f0, %f0, 4
+;   fiebr %f0, 4, %f0
 ;   br %r14
 
 function %nearest_f64(f64) -> f64 {
@@ -365,7 +365,7 @@ block0(v0: f64):
 }
 
 ; block0:
-;   fidbr %f0, %f0, 4
+;   fidbr %f0, 4, %f0
 ;   br %r14
 
 function %fma_f32(f32, f32, f32) -> f32 {
@@ -732,7 +732,7 @@ block0(v0: i8):
 ;   llgcr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdlgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_sint_i8_f32(i8) -> f32 {
@@ -745,7 +745,7 @@ block0(v0: i8):
 ;   lgbr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_uint_i16_f32(i16) -> f32 {
@@ -758,7 +758,7 @@ block0(v0: i16):
 ;   llghr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdlgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_sint_i16_f32(i16) -> f32 {
@@ -771,7 +771,7 @@ block0(v0: i16):
 ;   lghr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_uint_i32_f32(i32) -> f32 {
@@ -784,7 +784,7 @@ block0(v0: i32):
 ;   llgfr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdlgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_sint_i32_f32(i32) -> f32 {
@@ -797,7 +797,7 @@ block0(v0: i32):
 ;   lgfr %r4, %r2
 ;   ldgr %f4, %r4
 ;   wcdgb %f6, %f4, 0, 3
-;   ledbra %f0, %f6, 4
+;   ledbra %f0, 4, %f6, 0
 ;   br %r14
 
 function %fcvt_from_uint_i64_f32(i64) -> f32 {
@@ -809,7 +809,7 @@ block0(v0: i64):
 ; block0:
 ;   ldgr %f2, %r2
 ;   wcdlgb %f4, %f2, 0, 3
-;   ledbra %f0, %f4, 4
+;   ledbra %f0, 4, %f4, 0
 ;   br %r14
 
 function %fcvt_from_sint_i64_f32(i64) -> f32 {
@@ -821,7 +821,7 @@ block0(v0: i64):
 ; block0:
 ;   ldgr %f2, %r2
 ;   wcdgb %f4, %f2, 0, 3
-;   ledbra %f0, %f4, 4
+;   ledbra %f0, 4, %f4, 0
 ;   br %r14
 
 function %fcvt_from_uint_i8_f64(i8) -> f64 {