diff --git a/litex/soc/cores/cpu/picorv32/core.py b/litex/soc/cores/cpu/picorv32/core.py
index 7fe830d66..93f3358a2 100644
--- a/litex/soc/cores/cpu/picorv32/core.py
+++ b/litex/soc/cores/cpu/picorv32/core.py
@@ -9,10 +9,12 @@ class PicoRV32(Module):
     name = "picorv32"
     endianness = "little"
     gcc_triple = ("riscv64-unknown-elf", "riscv32-unknown-elf")
-    gcc_flags = "-D__picorv32__ -mno-save-restore -march=rv32im -mabi=ilp32"
+    gcc_flags_template = "-D__picorv32__ -mno-save-restore -march=rv32{ext} -mabi=ilp32"
     linker_output_format = "elf32-littleriscv"
 
     def __init__(self, platform, progaddr_reset, variant):
+        self.gcc_flags = ""
+
         self.reset = Signal()
         self.ibus = i = wishbone.Interface()
         self.dbus = d = wishbone.Interface()
@@ -29,31 +31,53 @@ class PicoRV32(Module):
         mem_wstrb = Signal(4)
         mem_rdata = Signal(32)
 
+        # PicoRV32 parameters. To create a new variant, modify this dictionary
+        # and change the desired parameters.
+        picorv32_params = {
+            "p_ENABLE_COUNTERS" : 1,
+            "p_ENABLE_COUNTERS64" : 1,
+            # Changing REGS has no effect as on FPGAs, the registers are
+            # implemented using a register file stored in DPRAM.
+            "p_ENABLE_REGS_16_31" : 1,
+            "p_ENABLE_REGS_DUALPORT" : 1,
+            "p_LATCHED_MEM_RDATA" : 0,
+            "p_TWO_STAGE_SHIFT" : 1,
+            "p_TWO_CYCLE_COMPARE" : 0,
+            "p_TWO_CYCLE_ALU" : 0,
+            "p_CATCH_MISALIGN" : 1,
+            "p_CATCH_ILLINSN" : 1,
+            "p_ENABLE_PCPI" : 0,
+            "p_ENABLE_MUL" : 1,
+            "p_ENABLE_DIV" : 1,
+            "p_ENABLE_FAST_MUL" : 0,
+            "p_ENABLE_IRQ" : 1,
+            "p_ENABLE_IRQ_QREGS" : 1,
+            "p_ENABLE_IRQ_TIMER" : 1,
+            "p_ENABLE_TRACE" : 0,
+            "p_MASKED_IRQ" : 0x00000000,
+            "p_LATCHED_IRQ" : 0xffffffff,
+            "p_PROGADDR_RESET" : progaddr_reset,
+            "p_PROGADDR_IRQ" : progaddr_reset + 0x00000010,
+            "p_STACKADDR" : 0xffffffff
+        }
+
+        if variant == None:
+            self.gcc_flags = PicoRV32.gcc_flags_template.format(ext="im")
+        elif variant == "minimal":
+            picorv32_params.update({
+                "p_ENABLE_COUNTERS" : 0,
+                "p_ENABLE_COUNTERS64" : 0,
+                "p_TWO_STAGE_SHIFT" : 0,
+                "p_CATCH_MISALIGN" : 0,
+                "p_ENABLE_MUL" : 0,
+                "p_ENABLE_DIV" : 0,
+                "p_ENABLE_IRQ_TIMER" : 0
+            })
+            self.gcc_flags = PicoRV32.gcc_flags_template.format(ext="i")
+
         self.specials += Instance("picorv32",
-            # parameters
-            p_ENABLE_COUNTERS=1,
-            p_ENABLE_COUNTERS64=1,
-            p_ENABLE_REGS_16_31=1,
-            p_ENABLE_REGS_DUALPORT=1,
-            p_LATCHED_MEM_RDATA=0,
-            p_TWO_STAGE_SHIFT=1,
-            p_TWO_CYCLE_COMPARE=0,
-            p_TWO_CYCLE_ALU=0,
-            p_CATCH_MISALIGN=1,
-            p_CATCH_ILLINSN=1,
-            p_ENABLE_PCPI=0,
-            p_ENABLE_MUL=1,
-            p_ENABLE_DIV=1,
-            p_ENABLE_FAST_MUL=0,
-            p_ENABLE_IRQ=1,
-            p_ENABLE_IRQ_QREGS=1,
-            p_ENABLE_IRQ_TIMER=1,
-            p_ENABLE_TRACE=0,
-            p_MASKED_IRQ=0x00000000,
-            p_LATCHED_IRQ=0xffffffff,
-            p_PROGADDR_RESET=progaddr_reset,
-            p_PROGADDR_IRQ=0x00000010,
-            p_STACKADDR=0xffffffff,
+            # parameters dictionary
+            **picorv32_params,
 
             # clock / reset
             i_clk=ClockSignal(),
diff --git a/litex/soc/software/libbase/crt0-picorv32.S b/litex/soc/software/libbase/crt0-picorv32.S
index 1d6faa851..7e4c712b3 100644
--- a/litex/soc/software/libbase/crt0-picorv32.S
+++ b/litex/soc/software/libbase/crt0-picorv32.S
@@ -18,7 +18,23 @@ _start:
 
 .org 0x00000010 # IRQ
 _irq_vector:
-  j _irq
+  addi sp, sp, -16
+  sw t0, 4(sp)
+  sw ra, 8(sp)
+  /* By convention, q2 holds true IRQ vector, but remains caller-save.
+  We rely on the assumption that compiler-generated code will never touch
+  the QREGs. q3 is truly scratch/caller-save. */
+  picorv32_getq_insn(t0, q2)
+  sw t0, 12(sp)
+
+  jalr t0 // Call the true IRQ vector.
+
+  lw t0, 12(sp)
+  picorv32_setq_insn(q2, t0) // Restore the true IRQ vector.
+  lw ra, 8(sp)
+  lw t0, 4(sp)
+  addi sp, sp, 16
+  picorv32_retirq_insn() // return from interrupt
 
 
 /*
@@ -132,9 +148,7 @@ _irq:
   /* restore x1 - x2 from q registers */
   picorv32_getq_insn(x1, q1)
   picorv32_getq_insn(x2, q2)
-
-  /* return from interrupt */
-  picorv32_retirq_insn()
+  ret
 
 /*
  * Reset handler, branched to from the vector.
@@ -180,9 +194,23 @@ _crt0:
   la t1, _irq_mask
   sw t0, 0(t1)
 
+  /* Clear BSS */
+  la t0, _fbss
+  la t1, _ebss
+2:
+  sw zero, 0(t0)
+  addi t0, t0, 4
+  bltu t0, t1, 2b
+
   /* set main stack */
   la sp, _fstack
 
+  /* Set up address to IRQ handler since vector is hardcoded.
+  By convention, q2 keeps the pointer to the true IRQ handler,
+  to emulate relocatable interrupts. */
+  la t0, _irq
+  picorv32_setq_insn(q2, t0)
+
   /* jump to main */
   jal ra, main
 
@@ -206,7 +234,7 @@ _irq_enable:
   picorv32_maskirq_insn(zero, t0)
   ret
 
-/* 
+/*
  * Disable interrupts by masking all interrupts (the mask should already be
  * up to date)
  */
@@ -239,7 +267,7 @@ _irq_setmask:
   picorv32_maskirq_insn(zero, a0)
 1:
   ret
-  
+
 
 .section .bss
 irq_regs:
@@ -272,4 +300,3 @@ _irq_mask:
 .global _irq_enabled
 _irq_enabled:
   .word 0
-