pulley: Flag all float registers as caller-saved (#11589)

This commit updates the ABI used by Pulley to consider all float
register as caller-saved instead of the previous classification which
was a split half/half. This change means that when exceptions are used,
for example, only the upper X registers need to be saved instead of the
upper F registers as well. This mirrors the x64 ABI for example which
clobbers all float registers on SystemV variants.

The motivation for this commit is to shrink the generated bytecode for
exception-using functions. Clobbering all X registers has special
support in the `push_frame_save` and `pop_frame_restore` opcodes where
the registers to save are a bitmask. With floats a separate
instruction-per-register is needed. While it's possible to add a macro
opcode to handle floats as well it was easier to update the ABI to match
x64. Basically I figure that if the ABI is good enough for x64 it's
probably fine for Pulley which already doesn't have great float
performance (all float ops are "extended opcodes").

Author: alexcrichton

cranelift/codegen/src/isa/pulley_shared/abi.rs

@@ -745,24 +745,7 @@ const DEFAULT_CALLEE_SAVES: PRegSet = PRegSet::empty()
     .with(px_reg(29))
     .with(px_reg(30))
     .with(px_reg(31))
-    // Float registers.
-    .with(pf_reg(16))
-    .with(pf_reg(17))
-    .with(pf_reg(18))
-    .with(pf_reg(19))
-    .with(pf_reg(20))
-    .with(pf_reg(21))
-    .with(pf_reg(22))
-    .with(pf_reg(23))
-    .with(pf_reg(24))
-    .with(pf_reg(25))
-    .with(pf_reg(26))
-    .with(pf_reg(27))
-    .with(pf_reg(28))
-    .with(pf_reg(29))
-    .with(pf_reg(30))
-    .with(pf_reg(31))
-    // Note: no vector registers are callee-saved.
+    // Note: no float/vector registers are callee-saved.
 ;
 
 fn compute_clobber_size(clobbers: &[Writable<RealReg>]) -> u32 {
@@ -799,7 +782,7 @@ const DEFAULT_CLOBBERS: PRegSet = PRegSet::empty()
     .with(px_reg(13))
     .with(px_reg(14))
     .with(px_reg(15))
-    // Float registers: the first 16 get clobbered.
+    // All float registers get clobbered.
     .with(pf_reg(0))
     .with(pf_reg(1))
     .with(pf_reg(2))
@@ -816,6 +799,22 @@ const DEFAULT_CLOBBERS: PRegSet = PRegSet::empty()
     .with(pf_reg(13))
     .with(pf_reg(14))
     .with(pf_reg(15))
+    .with(pf_reg(16))
+    .with(pf_reg(17))
+    .with(pf_reg(18))
+    .with(pf_reg(19))
+    .with(pf_reg(20))
+    .with(pf_reg(21))
+    .with(pf_reg(22))
+    .with(pf_reg(23))
+    .with(pf_reg(24))
+    .with(pf_reg(25))
+    .with(pf_reg(26))
+    .with(pf_reg(27))
+    .with(pf_reg(28))
+    .with(pf_reg(29))
+    .with(pf_reg(30))
+    .with(pf_reg(31))
     // All vector registers get clobbered.
     .with(pv_reg(0))
     .with(pv_reg(1))
@@ -955,7 +954,7 @@ fn create_reg_environment() -> MachineEnv {
 
     let preferred_regs_by_class: [Vec<PReg>; 3] = {
         let x_registers: Vec<PReg> = (0..16).map(|x| px_reg(x)).collect();
-        let f_registers: Vec<PReg> = (0..16).map(|x| pf_reg(x)).collect();
+        let f_registers: Vec<PReg> = (0..32).map(|x| pf_reg(x)).collect();
         let v_registers: Vec<PReg> = (0..32).map(|x| pv_reg(x)).collect();
         [x_registers, f_registers, v_registers]
     };
@@ -964,7 +963,7 @@ fn create_reg_environment() -> MachineEnv {
         let x_registers: Vec<PReg> = (16..XReg::SPECIAL_START)
             .map(|x| px_reg(x.into()))
             .collect();
-        let f_registers: Vec<PReg> = (16..32).map(|x| pf_reg(x)).collect();
+        let f_registers: Vec<PReg> = vec![];
         let v_registers: Vec<PReg> = vec![];
         [x_registers, f_registers, v_registers]
     };

cranelift/filetests/filetests/isa/pulley32/call.clif

@@ -16,7 +16,7 @@ block0:
 ;   push_frame
 ; block0:
 ;   xzero x2
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p2i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }], clobbers: PRegSet { bits: [65534, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p2i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }], clobbers: PRegSet { bits: [65534, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   xone x0
 ;   pop_frame
 ;   ret
@@ -43,7 +43,7 @@ block0:
 ;   push_frame
 ; block0:
 ;   xzero x2
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p2i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I32) }], clobbers: PRegSet { bits: [65534, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p2i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I32) }], clobbers: PRegSet { bits: [65534, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   xone x0
 ;   pop_frame
 ;   ret
@@ -75,7 +75,7 @@ block0:
 ;   xone x4
 ;   xconst8 x5, 2
 ;   xconst8 x6, 3
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p3i), XReg(p4i), XReg(p5i), XReg(p6i)] }, uses: [], defs: [], clobbers: PRegSet { bits: [65535, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p3i), XReg(p4i), XReg(p5i), XReg(p6i)] }, uses: [], defs: [], clobbers: PRegSet { bits: [65535, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   pop_frame
 ;   ret
 ;
@@ -103,7 +103,7 @@ block0:
 ; VCode:
 ;   push_frame
 ; block0:
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }, CallRetPair { vreg: Writable { reg: p1i }, location: Reg(p1i, types::I64) }, CallRetPair { vreg: Writable { reg: p2i }, location: Reg(p2i, types::I64) }, CallRetPair { vreg: Writable { reg: p3i }, location: Reg(p3i, types::I64) }], clobbers: PRegSet { bits: [65520, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }, CallRetPair { vreg: Writable { reg: p1i }, location: Reg(p1i, types::I64) }, CallRetPair { vreg: Writable { reg: p2i }, location: Reg(p2i, types::I64) }, CallRetPair { vreg: Writable { reg: p3i }, location: Reg(p3i, types::I64) }], clobbers: PRegSet { bits: [65520, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   xadd64 x4, x0, x2
 ;   xadd64 x3, x1, x3
 ;   xadd64 x0, x4, x3
@@ -149,7 +149,7 @@ block0:
 ;   xmov x11, x14
 ;   xmov x12, x14
 ;   xmov x13, x14
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p14i), XReg(p14i), XReg(p14i), XReg(p14i)] }, uses: [CallArgPair { vreg: p4i, preg: p4i }, CallArgPair { vreg: p5i, preg: p5i }, CallArgPair { vreg: p6i, preg: p6i }, CallArgPair { vreg: p7i, preg: p7i }, CallArgPair { vreg: p8i, preg: p8i }, CallArgPair { vreg: p9i, preg: p9i }, CallArgPair { vreg: p10i, preg: p10i }, CallArgPair { vreg: p11i, preg: p11i }, CallArgPair { vreg: p12i, preg: p12i }, CallArgPair { vreg: p13i, preg: p13i }, CallArgPair { vreg: p14i, preg: p14i }], defs: [], clobbers: PRegSet { bits: [65535, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p14i), XReg(p14i), XReg(p14i), XReg(p14i)] }, uses: [CallArgPair { vreg: p4i, preg: p4i }, CallArgPair { vreg: p5i, preg: p5i }, CallArgPair { vreg: p6i, preg: p6i }, CallArgPair { vreg: p7i, preg: p7i }, CallArgPair { vreg: p8i, preg: p8i }, CallArgPair { vreg: p9i, preg: p9i }, CallArgPair { vreg: p10i, preg: p10i }, CallArgPair { vreg: p11i, preg: p11i }, CallArgPair { vreg: p12i, preg: p12i }, CallArgPair { vreg: p13i, preg: p13i }, CallArgPair { vreg: p14i, preg: p14i }], defs: [], clobbers: PRegSet { bits: [65535, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   pop_frame_restore 64, {}
 ;   ret
 ;
@@ -217,7 +217,7 @@ block0:
 ;   push_frame_save 112, {x16, x17, x18, x19, x26, x27, x28, x29}
 ; block0:
 ;   x12 = load_addr OutgoingArg(0)
-;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p12i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }, CallRetPair { vreg: Writable { reg: p1i }, location: Reg(p1i, types::I64) }, CallRetPair { vreg: Writable { reg: p2i }, location: Reg(p2i, types::I64) }, CallRetPair { vreg: Writable { reg: p3i }, location: Reg(p3i, types::I64) }, CallRetPair { vreg: Writable { reg: p4i }, location: Reg(p4i, types::I64) }, CallRetPair { vreg: Writable { reg: p5i }, location: Reg(p5i, types::I64) }, CallRetPair { vreg: Writable { reg: p6i }, location: Reg(p6i, types::I64) }, CallRetPair { vreg: Writable { reg: p7i }, location: Reg(p7i, types::I64) }, CallRetPair { vreg: Writable { reg: p8i }, location: Reg(p8i, types::I64) }, CallRetPair { vreg: Writable { reg: p9i }, location: Reg(p9i, types::I64) }, CallRetPair { vreg: Writable { reg: p10i }, location: Reg(p10i, types::I64) }, CallRetPair { vreg: Writable { reg: p11i }, location: Reg(p11i, types::I64) }, CallRetPair { vreg: Writable { reg: p12i }, location: Reg(p12i, types::I64) }, CallRetPair { vreg: Writable { reg: p13i }, location: Reg(p13i, types::I64) }, CallRetPair { vreg: Writable { reg: p14i }, location: Reg(p14i, types::I64) }, CallRetPair { vreg: Writable { reg: p27i }, location: Stack(OutgoingArg(0), types::I64) }, CallRetPair { vreg: Writable { reg: p19i }, location: Stack(OutgoingArg(8), types::I64) }, CallRetPair { vreg: Writable { reg: p29i }, location: Stack(OutgoingArg(16), types::I64) }, CallRetPair { vreg: Writable { reg: p16i }, location: Stack(OutgoingArg(24), types::I64) }, CallRetPair { vreg: Writable { reg: p17i }, location: Stack(OutgoingArg(32), types::I64) }, CallRetPair { vreg: Writable { reg: p18i }, location: Stack(OutgoingArg(40), types::I64) }], clobbers: PRegSet { bits: [32768, 65535, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   call CallInfo { dest: PulleyCall { name: TestCase(%g), args: [XReg(p12i)] }, uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }, CallRetPair { vreg: Writable { reg: p1i }, location: Reg(p1i, types::I64) }, CallRetPair { vreg: Writable { reg: p2i }, location: Reg(p2i, types::I64) }, CallRetPair { vreg: Writable { reg: p3i }, location: Reg(p3i, types::I64) }, CallRetPair { vreg: Writable { reg: p4i }, location: Reg(p4i, types::I64) }, CallRetPair { vreg: Writable { reg: p5i }, location: Reg(p5i, types::I64) }, CallRetPair { vreg: Writable { reg: p6i }, location: Reg(p6i, types::I64) }, CallRetPair { vreg: Writable { reg: p7i }, location: Reg(p7i, types::I64) }, CallRetPair { vreg: Writable { reg: p8i }, location: Reg(p8i, types::I64) }, CallRetPair { vreg: Writable { reg: p9i }, location: Reg(p9i, types::I64) }, CallRetPair { vreg: Writable { reg: p10i }, location: Reg(p10i, types::I64) }, CallRetPair { vreg: Writable { reg: p11i }, location: Reg(p11i, types::I64) }, CallRetPair { vreg: Writable { reg: p12i }, location: Reg(p12i, types::I64) }, CallRetPair { vreg: Writable { reg: p13i }, location: Reg(p13i, types::I64) }, CallRetPair { vreg: Writable { reg: p14i }, location: Reg(p14i, types::I64) }, CallRetPair { vreg: Writable { reg: p27i }, location: Stack(OutgoingArg(0), types::I64) }, CallRetPair { vreg: Writable { reg: p19i }, location: Stack(OutgoingArg(8), types::I64) }, CallRetPair { vreg: Writable { reg: p29i }, location: Stack(OutgoingArg(16), types::I64) }, CallRetPair { vreg: Writable { reg: p16i }, location: Stack(OutgoingArg(24), types::I64) }, CallRetPair { vreg: Writable { reg: p17i }, location: Stack(OutgoingArg(32), types::I64) }, CallRetPair { vreg: Writable { reg: p18i }, location: Stack(OutgoingArg(40), types::I64) }], clobbers: PRegSet { bits: [32768, 4294967295, 4294967295, 0] }, callee_conv: Fast, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   xadd64 x26, x0, x1
 ;   xadd64 x28, x2, x3
 ;   xadd64 x2, x4, x5
@@ -291,7 +291,7 @@ block0(v0: i32):
 ; VCode:
 ;   push_frame
 ; block0:
-;   indirect_call x0, CallInfo { dest: XReg(p0i), uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }], clobbers: PRegSet { bits: [65534, 65535, 4294967295, 0] }, callee_conv: Tail, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
+;   indirect_call x0, CallInfo { dest: XReg(p0i), uses: [], defs: [CallRetPair { vreg: Writable { reg: p0i }, location: Reg(p0i, types::I64) }], clobbers: PRegSet { bits: [65534, 4294967295, 4294967295, 0] }, callee_conv: Tail, caller_conv: Fast, callee_pop_size: 0, try_call_info: None }
 ;   pop_frame
 ;   ret
 ;